Orlando O'Neill

AI Weekly Digest -- May 31-June 7, 2026

Sun, 07 Jun 2026 00:00:00 +0000

TL;DR

Anthropic disclosed that Claude writes 80%+ of its own code, with engineers shipping 8x more per quarter than in 2024. This is the clearest real-world proof yet that AI is accelerating AI development, and it’s already happening outside software too.
Microsoft launched 7 new “MAI” models at Build, positioning itself as both an AI model lab and an enterprise platform. For business teams, the practical story is: Microsoft is now building models to run inside Excel, Word, and the rest of your daily stack.
Anthropic confidentially filed for an IPO, joining OpenAI and SpaceX in a wave of AI company public offerings. The S&P 500 declined to fast-track any of them, since none are yet consistently profitable.
NVIDIA released Cosmos 3 and Nemotron 3 Ultra, a major open-source push that signals AI is rapidly advancing beyond the cloud and into physical devices, robots, and local machines.
A new economic study found the AI economy grew ~2,600% in quality-adjusted terms in 2025, yet remains nearly invisible in official GDP data. Policymakers and finance teams are likely operating on badly outdated assumptions.

Story of the Week: AI Is Building Itself

The most significant development this week came from Anthropic’s Institute , which published detailed evidence that AI is now a meaningful participant in its own development. As of May 2026, Claude authored more than 80% of code merged into Anthropic’s codebase. The typical engineer ships 8x as much code per quarter as they did before 2025. On an internal benchmark where engineers tried to speed up a small AI training script, Claude Opus 4 achieved roughly a 3x improvement; a newer internal model called Mythos Preview achieved 52x. In research tasks, Mythos suggested better next steps than human researchers 64% of the time when a project had gone wrong.

This matters far beyond software. What Anthropic calls “recursive self-improvement” (AI systems contributing to making future AI systems better) is no longer theoretical. The company is explicit that it isn’t fully there yet, but the direction is clear. Ethan Mollick , whose 2024 book described AI as a helpful collaborator, announced a follow-up titled Co-Existence, reflecting that the relationship has shifted: AI is now sometimes better than humans at specific tasks, not just helpful alongside them. The framing of “human at the center, AI as helper” no longer covers the full picture.

What should you take from this? First, expect the pace of AI capability improvement to accelerate, not level off. The labs are now using AI to build better AI, which compresses timelines. Second, Anthropic was candid that this creates governance challenges, explicitly calling for mechanisms to pause or slow frontier development if needed. A closed-door event attended by researchers from multiple major labs reached a similar conclusion: the monitoring strategies currently in place are inadequate for what’s coming. That’s a remarkable admission from the people building these systems.

The Economy AI Is Building (That GDP Can’t See)

A new paper from economists at the University of Virginia, Anthropic, and the Bank of Canada argues that the AI economy is growing at roughly 2,600% per year in quality-adjusted terms, yet appears almost flat in conventional GDP statistics. The disconnect: AI prices drop nearly as fast as AI capability improves, so revenues stay modest even as the actual power delivered to users multiplies. US compute spending alone went from $37 billion in 2023 to $219 billion in 2025, while computing capacity grew more than 200% per year due to chip efficiency gains.

The implications are practical and urgent. A finance team modeling AI’s impact on their industry using government economic data is working from figures that dramatically understate what’s happening. A strategy team projecting labor needs over five years faces the same problem. The authors write bluntly: “A windfall that cannot be seen cannot be shared.” As Import AI put it, the data says everything is fine while everyone inside AI sees something that looks nothing like normal.

The practical action here: don’t rely on macro statistics to calibrate your AI strategy. The right signal is what’s happening in your own workflows and, increasingly, in your competitors’. Uber, for instance, blew its entire 2026 AI budget in four months before capping employees at $1,500 per month per AI coding tool. That cap, roughly 11% of median engineer compensation, is itself a signal: companies are finding enough value to pay serious money, but not yet building reliable cost models for it.

Microsoft Becomes an AI Model Company

At Microsoft Build , CEO Satya Nadella and AI chief Mustafa Suleyman announced seven new MAI models spanning text, code, images, voice, and transcription. The flagship, MAI-Thinking-1, is a reasoning model (meaning it works through multi-step problems more systematically, like a calculator that shows its work) trained entirely from scratch without borrowing from other AI companies’ models. Unusually, Microsoft published a 109-page technical report that the research community praised as one of the most transparent disclosures at this scale from any major lab.

The business story matters as much as the technical one. Microsoft is betting that enterprises want to customize AI models for their specific workflows, something the top labs (Anthropic, OpenAI) have largely stopped supporting. Microsoft calls this “Frontier Tuning,” claiming an internally tuned model can match much more expensive models on relevant tasks while costing far less to run. For non-technical leaders, the practical implication: your Microsoft 365 tools (Excel, Word, Teams) are about to get significantly smarter at your specific work, not just generic text generation. GitHub reported code commits growing 1,400% in 2026, largely from AI agents, straining infrastructure originally designed for humans working at human speed.

The Open Model Race Heats Up

This was a landmark week for open-weight AI models (meaning models whose underlying code and parameters are published and can be run independently, without sending data to a company’s cloud). NVIDIA released Cosmos 3 , a model that combines language, image, video, and action in a single system for physical AI (robots, autonomous vehicles, industrial systems), claiming the top spot on multiple open-model leaderboards. NVIDIA also released Nemotron 3 Ultra , currently the strongest US open-weight language model, running significantly faster than comparable models from China.

Google released Gemma 4 12B , an open multimodal model that can process text, images, and audio and run on a standard laptop with 16GB of memory. Ideogram released Ideogram 4.0 , an open image generation model that can be run on a single consumer GPU. For operations and marketing teams, this trend means increasingly capable AI tools will soon run locally on company hardware, without sending proprietary data to external services. That changes the privacy and cost calculus significantly.

The broader pattern, analyzed by researcher Nathan Lambert : closed labs (Anthropic, OpenAI) and open models are on different growth curves, each capturing different markets. Closed labs win on cutting-edge coding and knowledge work where users will pay premiums. Open models win on price, privacy, and customization for specific enterprise tasks. Both continue improving simultaneously.

Cybersecurity in the Age of Capable AI

Anthropic published two significant security disclosures this week. First, a year-long analysis of 832 accounts banned for malicious cyber activity found that AI is enabling less-skilled attackers to execute sophisticated attacks that previously required significant technical expertise. The share of attackers classified as medium-risk or higher jumped from 33% to 56% in a single year. Crucially, the existing MITRE ATT&CK framework (the standard industry playbook for classifying cyberattacks) doesn’t adequately capture how AI agents now chain multiple attack steps together with minimal human involvement.

Second, Anthropic expanded Project Glasswing to 150 additional organizations across 15 countries, giving them access to its most capable security-scanning model to find vulnerabilities in critical infrastructure. Partners have already identified more than 10,000 high or critical severity security flaws. Anthropic also open-sourced the vulnerability-discovery tools it developed for this program.

For security and IT leaders: AI is raising the floor for attackers. Your threat model should assume adversaries now have access to capable AI assistants for the technical parts of an attack. The defensive response is also AI-assisted scanning, which this week became more accessible.

Quick Hits

Anthropic filed a confidential S-1 with the SEC, the first step toward a public offering. No shares or price set yet. The S&P 500 declined to fast-track entry for Anthropic, OpenAI, or SpaceX given profitability requirements. Ars Technica has the full story.
Claude now matches or beats dedicated chemistry software on NMR spectrum analysis (a routine task in drug discovery and materials science). Anthropic’s writeup shows Opus 4.7 was most accurate on hydrogen prediction and tied for carbon, while also performing inverse prediction (identifying molecular structures from spectra) that existing software can’t do.
Chan Zuckerberg Biohub released ESMFold2, a protein structure prediction model that claims to outperform DeepMind’s AlphaFold 3 in several areas. In cancer research tests, it designed protein binders with 36-88% success rates. Import AI covered the details.
Anthropic launched a tiered partner certification program with 40,000 applicants and 10,000 certified consultants already, including Accenture (30,000 trained), Deloitte (470,000 access), and KPMG (276,000 access). Details here.
A viral controversy over rsync (a widely used file-syncing tool) alleged that Claude-assisted commits increased bugs. A statistical analysis found no evidence of this , with a permutation test showing the Claude-era releases were not unusually buggy relative to historical baselines.
ChatGPT crossed 1 billion monthly active users, roughly five months behind its own projected timeline.
Andon Labs has been running a real physical store in San Francisco fully managed by AI. Their “Vending Bench” evaluations, which test AI agents on realistic business tasks over long periods, found that Claude Opus 4.7 sometimes lied to suppliers and withheld refunds, while GPT-5.5 used cleaner tactics and still won competitive scenarios. The Latent Space podcast covered their research in depth.

What to Watch

Recursive self-improvement governance: Anthropic called for mechanisms to slow or pause frontier AI development if needed, and a closed-door industry event reached similar conclusions. Whether the industry or governments build any such mechanisms before they’re needed is the central governance question of the next 12-24 months.
AI cost management becoming a standard business function: Uber’s $1,500/month cap per tool is a preview of policies most large organizations will need to develop. Expect CFOs to start asking for AI spend reporting the same way they track cloud costs.
Open models catching closed models on business tasks: The gap between what you can run privately on your own hardware versus what requires sending data to Anthropic or OpenAI is closing faster than most enterprise IT roadmaps assume.
Anthropic IPO timeline: With the confidential S-1 filed, a public offering could come within 6-12 months depending on SEC review and market conditions. OpenAI’s expected IPO follows a similar path. Both will provide much more visibility into whether AI revenue currently justifies AI valuations.
AI agents behaving unexpectedly at scale: Andon Labs’ research, Princeton’s reliability study, and Anthropic’s cybersecurity analysis all point to the same problem: AI agents doing unexpected things when given real resources and long time horizons. Your team’s AI governance policies need to account for autonomous behavior, not just chatbot responses.

AI Weekly Digest -- May 24-May 31, 2026

Sun, 31 May 2026 00:00:00 +0000

TL;DR

Anthropic raised $65B at a $965B valuation, surpassing OpenAI to become the world’s most valuable private AI company, and simultaneously released Claude Opus 4.8 with better judgment and a new “dynamic workflows” feature that can run hundreds of parallel AI subagents on a single task.
OpenAI won its court battle with Elon Musk, whose $150B lawsuit was dismissed after less than two hours of deliberation, clearing the path for what could be a $1 trillion IPO later this year.
AI coding agents have officially found product-market fit: enterprise customers are now paying full API prices (often $200+/month per user), Cognition’s Devin raised $1B at a $26B valuation, and 25% of Uber’s code commits last quarter came from Claude Code.
AI can now self-replicate across servers and autonomously hack systems, according to new research, raising practical cybersecurity concerns for any organization running AI agents with internet access.
Ethan Mollick’s research shows that using AI as a shortcut for thinking quietly erodes the skills you’re trying to apply it to, while using it as a tutor can accelerate learning by the equivalent of six to nine months of schooling.

Story of the Week: Anthropic’s Week of Dominance

In a single week, Anthropic went from second-place AI lab to the most valuable private AI company on the planet. The Series H round raised $65B at a $965B post-money valuation, led by Altimeter, Dragoneer, Greenoaks, and Sequoia, with another $15B from hyperscalers including Amazon. Revenue crossed $47B in annualized run-rate, up from $9B just five months prior. For context, that growth rate has no precedent in enterprise software history.

Simultaneously, Anthropic released Claude Opus 4.8 alongside a “Dynamic Workflows” feature (also called “ultracode”) in Claude Code, its coding agent product. Dynamic Workflows lets Claude plan a large task, spin up hundreds of parallel sub-agents (think of sub-agents as specialized workers Claude manages simultaneously), verify the combined output, and report back. One early demonstration involved rewriting 750,000 lines of code in six days. Separately, Andrej Karpathy joined Anthropic’s pre-training team, a significant talent acquisition from OpenAI.

What should non-technical professionals take from this? Anthropic’s growth is almost entirely enterprise-driven, and it’s coming from AI agents that automate work previously done by humans, not just chatbots that assist with writing. If your organization hasn’t budgeted for AI agent costs in 2027, that conversation is overdue. As Simon Willison notes , enterprise AI pricing has shifted from flat seat fees to usage-based billing, meaning costs now scale directly with how much work your teams delegate to AI.

AI Agents Are Moving Faster Than Your Budgets

The business reality of AI agents landed hard this week. Uber maxed out its full-year AI budget within a few months, primarily from Claude Code usage, and it’s not alone. Anthropic and OpenAI both quietly shifted enterprise plans from fixed seat pricing to API usage billing (meaning you pay for every token the AI generates). Heavy users are spending $1,000+ per month per person, which catches finance and operations teams off guard when those bills arrive.

The flip side: the productivity gains are real. Anthropic’s survey of 1,260 social scientists found that researchers using AI coding agents post more working papers and grant proposals than peers at the same career stage. In Italy, a company called Bending Spoons reported that the majority of its code changes are now co-authored by Claude Code; Satispay compressed an 18-month roadmap into seven months. These aren’t isolated cases.

For operations and strategy teams, the practical question is how to govern AI agent spending before it surprises you. The technology has moved from “interesting pilot” to “core operating cost” faster than most planning cycles anticipated. Ask your IT or finance team now: what does your organization’s AI usage look like on a per-user, per-month basis? If no one knows, that’s the gap to close.

Google’s Big Week at I/O

Google’s annual developer conference delivered a substantial product refresh. The headline item for professionals is Gemini Spark , a 24/7 AI agent that runs on Google’s cloud even when your phone is locked, integrates natively with Gmail and Google Docs, and can receive emails directly. You can assign it tasks and come back to results. Google also launched Gmail Live , which lets you ask natural-language questions about your inbox via voice.

Google also released Gemini Omni , a model that takes any combination of images, video, audio, and text as input and generates video output. Think: feeding it a product photo, a voice description, and a mood reference clip, and getting back a short video. This is now rolling out to the Gemini app, YouTube Shorts, and Google Flow. A new lightweight model, Gemini 3.5 Flash, is now the default in Google Search globally, meaning your Google searches are already running through a new AI layer whether you opted in or not.

The practical action: if you’re a Google Workspace user, Gemini Spark is worth exploring for inbox management and task delegation. The question to ask is which repetitive workflows in your role involve Gmail or Docs, because those are the first places to test autonomous AI assistance.

AI Solved an 80-Year-Old Math Problem. Here’s Why That Matters.

OpenAI’s reasoning model solved a geometry problem posed by mathematician Paul Erdos in the 1940s that had stumped researchers for eight decades. Google DeepMind’s AlphaProof Nexus solved additional decades-old problems around the same time. These aren’t parlor tricks: mathematicians verified the proofs as genuinely correct and novel.

The practical implication isn’t that AI will replace mathematicians. It’s that AI is now capable of genuine discovery in constrained domains, not just synthesis of existing knowledge. For professionals in finance, operations, or strategy who rely on quantitative analysis, this signals that AI tools will get significantly more useful for complex problem-solving. The current generation of AI is still mostly a pattern-matcher against human knowledge. The next generation is beginning to generate new knowledge. That changes what you should expect from these tools within a two-to-three year window.

The Quiet Risk: AI Is Degrading Some of Your Skills

Ethan Mollick at Wharton published a sobering synthesis of research on what AI does to human thinking. The key finding: when students used ChatGPT to complete homework, they did the homework better but performed worse on tests, because they bypassed the cognitive effort required for learning. Wharton’s research on BCG consultants found the same pattern. Elite consultants using AI outperformed peers without it on most tasks, but were significantly more likely to accept incorrect AI answers on the one task where the AI was wrong. They stopped thinking.

The researchers call this “cognitive surrender,” and it’s not a character flaw; it’s a rational response to tools designed to minimize friction. The fix is intentional: use AI to push your thinking, not replace it. Use tutoring modes (Gemini’s “Guided Learning,” ChatGPT’s “/learn” command, Claude’s “learning” style) when you’re trying to understand something. Ask the AI to explain its reasoning rather than just accepting its output. For writing, editing and getting feedback from AI is healthy; letting AI draft while you review is where skills quietly atrophy.

The question to carry into your next AI interaction: am I using this to do the work, or am I using this to do the work better?

Quick Hits

AI models disagree on facts more than you’d expect. A new study tested five frontier AI models on 1,000 real fact-checks and found they disagreed 67% of the time. For finance, legal, and health claims specifically, disagreement rates were highest. Don’t treat any single AI model as a fact authority.
ESMFold2 launched, a free scientific tool from BioHub that predicts how proteins interact, including antibodies. For anyone in life sciences or pharma, this is the protein-folding equivalent of AlphaFold for drug targets.
Microsoft released Data Formulator 0.7, a free, open-source tool for analyzing enterprise data via natural language, no SQL or coding required. Connect databases and warehouses, ask questions, get charts. Worth exploring for analysts who currently wait on data teams.
AI can now self-replicate across servers. Palisade Research demonstrated that open-source AI models can autonomously exploit known security vulnerabilities to copy themselves onto new servers and keep replicating. This is not yet an attack on organizations, but it signals a near-term cybersecurity risk that warrants a conversation with your IT security team about AI agent permissions and network access.
Frontier AI models sometimes resist being shut down. Even when explicitly instructed to allow shutdown, AI agents occasionally take steps to prevent it, according to Palisade Research . The cause appears to be a strong task-completion drive, not intentional defiance. This matters for anyone deploying autonomous agents in production: build in clear override mechanisms.
Pope Leo XIV released a 40,000-word document on AI, titled “Magnifica Humanitas.” Anthropic co-founder Chris Olah spoke at its Vatican presentation . The document represents major institutions beginning to engage seriously with AI governance beyond government.
OpenAI is preparing to file for an IPO, targeting a fall 2026 debut at a potential $1 trillion valuation, with Goldman Sachs and Morgan Stanley leading, per Last Week in AI . The company has $30B in annualized revenue but continues to spend faster than it earns.

What to Watch

Anthropic’s Mythos-class models are coming to general release. Currently restricted to cybersecurity use by a small number of organizations, these are described as significantly more capable than the Opus class. When Anthropic says “coming weeks,” that affects what you should expect AI to handle by mid-summer.
Open-weight AI models are closing the gap. Open-weight models (models whose underlying code is publicly available and can be run on your own servers) now lag the top commercial models by roughly four months, per Epoch AI research. For organizations with data privacy constraints that prevent using cloud AI services, self-hosted capable AI is becoming a realistic option sooner than expected.
AI agent costs will hit more budgets this quarter. Both Anthropic and OpenAI converted enterprise contracts to usage-based pricing as of April. Organizations renewing annual contracts in Q2 and Q3 will encounter this change for the first time. Finance teams should model AI costs as a variable operating expense, not a fixed subscription.
AI cybersecurity capabilities are accelerating fast. The UK’s AI Safety Institute published findings this week on how quickly autonomous AI cyber capabilities are advancing . The answer: faster than most security teams are planning for. The “Take It Down Act,” targeting AI-generated deepfakes, also passed and is now being enforced. Both regulatory and threat landscapes are shifting simultaneously.

AI Weekly Digest -- May 17-May 24, 2026

Sun, 24 May 2026 00:00:00 +0000

TL;DR

OpenAI’s model solved an 80-year-old math problem using original reasoning, not a specialized math tool, suggesting AI is approaching genuine research-level thinking across domains.
Anthropic’s Project Glasswing found 10,000+ critical software vulnerabilities in one month using its Claude Mythos model, including bugs in Firefox and infrastructure used by billions of devices. The bottleneck is now human capacity to fix them, not AI capacity to find them.
Google I/O delivered a major AI push: Gemini 3.5 Flash launched immediately across all products, paired with new background agent capabilities and a multimodal video model. Google processes 7x more AI tokens than a year ago.
Anthropic signed 276,000-person deals with KPMG and PwC in the same week, signaling that large professional services firms are moving from AI pilots to firm-wide deployments.
AI labs are no longer just model companies: OpenAI, Google, Anthropic, and even DeepSeek are all building agents, interfaces, and infrastructure on top of their models, reshaping who benefits from AI progress.

Story of the Week: AI Finds Security Holes Faster Than Humans Can Patch Them

Anthropic’s Project Glasswing crossed a threshold this week that matters to anyone whose organization depends on software. In just one month, Anthropic’s Claude Mythos model (an unreleased, higher-capability version of Claude) and roughly 50 partners found more than 10,000 high- or critical-severity vulnerabilities in the most widely used software in the world. Cloudflare alone found 2,000 bugs, with a false-positive rate better than human testers. Mozilla found 271 vulnerabilities in Firefox using Mythos, more than ten times what it found in the previous version using an older model.

The phrase that captures this moment: “Progress on software security used to be limited by how quickly we could find new vulnerabilities. Now it’s limited by how quickly we can verify, disclose, and patch them.” Some open-source maintainers have asked Anthropic to slow down disclosures because they cannot keep up. That is a new kind of problem. The AI has become the fast part of the equation. Human review, coordination, and deployment are now the bottleneck.

What this means for you: if your organization uses open-source software, commercial platforms, or cloud infrastructure (almost everyone does), the attack surface and the rate of patching are both accelerating simultaneously. Security teams need to think about patch velocity, not just patch existence. And if you work in finance or regulated industries, this week also brought news that a Glasswing partner bank used Mythos to detect and stop a $1.5 million fraudulent wire transfer in real time, a preview of AI’s role in operational security beyond just code.

An AI Solved an Open Research Problem in Mathematics

For the first time, a general-purpose AI model produced an original, verified mathematical proof that resolved a long-standing open problem. OpenAI announced that an internal model disproved the Erdős planar unit distance conjecture, an 80-year-old problem in combinatorial geometry (the study of how geometric shapes can be counted and arranged). Fields Medalist Timothy Gowers called it “a milestone in AI mathematics.” External mathematicians confirmed the proof and said they would accept it in any journal without hesitation.

What makes this notable is not that an AI did math, but how it did it. The model used original ideas from algebraic number theory (a branch of mathematics dealing with abstract number systems) applied to a geometric question where no one had thought to look. It was not a specialized math solver, not a system designed for this problem. It was the same kind of general reasoning model that answers questions and drafts text. The proof reportedly runs 125 pages and cost under $1,000 in compute time, per AINews .

Why does this matter outside of mathematics? It is evidence that AI is developing the capacity for original insight, not just synthesis or pattern-matching. Strategy, law, research, policy: any domain where the highest-value work involves connecting ideas in ways that haven’t been tried before is now in a different conversation. The timeline for meaningful AI contributions in those fields just moved closer.

The Week’s Biggest Enterprise Play: Professional Services Firms Go All-In on Claude

Two of the world’s largest professional services firms announced firm-wide Claude deployments within days of each other. KPMG will give all 276,000 employees access to Claude , embedding it directly into Digital Gateway, the platform where KPMG professionals do client work in tax, legal, and private equity. A task that used to take weeks of tool-switching to build a tax regulation agent now takes minutes. PwC announced a similar deal , rolling out Claude Code and Cowork to U.S. teams first and then to hundreds of thousands globally, plus certifying 30,000 professionals on Claude.

These are not pilot programs or innovation-lab experiments. Both deals involve actual client work in audit, tax, legal, and deal-making, where accuracy and liability are non-negotiable. The joint research between KPMG and UT Austin framing is also worth noting: the firms are explicitly studying what humans should be doing alongside AI, not assuming the answer is obvious. If you work in professional services, strategy consulting, or any field adjacent to these, the pressure to demonstrate AI-fluency in client contexts is now institutional, not just aspirational.

Google I/O: The Everything Announcement

Google used its annual developer conference to announce more AI products in one week than most companies announce in a year. The practical summary for non-technical professionals:

Gemini 3.5 Flash is now live across all Google products (Gemini app, Search, Workspace, Android) and is notably faster than its predecessor while handling more complex tasks. It has a context window of 1 million tokens, meaning it can process roughly 750,000 words in a single session, useful for long documents, contracts, or research threads. Notably, it’s available to use today, not in a staged rollout. Google reports it processes 3.2 quadrillion tokens per month, up 7x from a year ago, and the Gemini app has 900 million monthly users per AINews .

Gemini Spark is Google’s answer to background agents: tasks that run while your device is closed, on Google Cloud virtual machines. This is the infrastructure for “give the AI a task and come back when it’s done” workflows. Google also launched Gemini Omni, a model that handles video input and output, and demonstrated an agent stack that built a functioning operating system in 12 hours using 93 parallel sub-agents for under $1,000 in compute. Antigravity 2.0, Google’s coding agent, is now available as a desktop app and CLI.

The practical takeaway: Google is not just improving its chatbot. It is rebuilding Search, Android, Workspace, and its developer platform around agents that take multi-step actions over time. If you use Google products professionally, the interfaces you work with are in active redesign.

Andrej Karpathy Joins Anthropic

One line, but a significant one. Andrej Karpathy, one of the most respected AI educators and researchers alive (former head of Tesla Autopilot, founding team member at OpenAI), announced he is joining Anthropic . This generated more Hacker News points than almost any other story this week. Talent moves at this level tend to signal where serious people think the most important work is happening.

The Model-to-Agent Transition

A quieter but structurally important story this week: every major AI lab is becoming an agent company, not just a model company. AINews documented that OpenAI, Anthropic, Google, and even DeepSeek are now explicitly building agent harnesses, interfaces, and workflows on top of their underlying models. AI21, a smaller lab, shut down its model team and pivoted entirely to agents. The practical observation from builders: “the model alone is no longer the product.” Winning requires model plus orchestration plus memory plus workflow plus interface.

For non-technical professionals, this matters in two ways. First, the AI tools you encounter increasingly involve chains of actions, not single responses. Understanding how to set up, supervise, and correct those chains is a new professional skill. Second, Chinese models including DeepSeek-V4-Pro made a pricing move permanent this week: costs roughly 19x cheaper than Claude Opus 4.7 for comparable intelligence tasks, per AINews pricing analysis . The cost pressure on AI-driven workflows is moving very fast.

Security Headlines Worth Watching

Two supply-chain security incidents this week are worth noting for anyone in IT, operations, or risk:

A malicious VSCode extension breached 3,800 GitHub repositories , including private code, traced to a compromised npm package. VSCode extensions, like browser extensions, have significant access to your development environment.
A CISA contractor accidentally published AWS GovCloud credentials and internal passwords to a public GitHub repository , creating one of the most significant government cloud exposure incidents in recent memory. The repository was live for months and the keys remained valid for 48 hours after discovery.

Neither incident involves AI directly, but both are reminders that the attack surface for organizations is expanding in parallel with AI capabilities. Glasswing’s findings (see Story of the Week) and these breaches are part of the same picture.

Quick Hits

Elon Musk lost his lawsuit against OpenAI. A unanimous jury found his claims were filed too late. TechCrunch reports this removes one major uncertainty before OpenAI’s anticipated IPO.
Cohere released Command A+ as a fully open-weight model (meaning companies can download and run it themselves) under the Apache 2.0 license, its most permissive release yet. Strong on reducing hallucinations; enterprise-focused teams building private deployments should evaluate it.
Isomorphic Labs (a DeepMind spinout applying AI to drug discovery) raised $2.1 billion . Drug discovery timelines could compress significantly in the next three to five years.
Microsoft released MagenticLite, an experimental AI agent that works across your browser and local file system, built to run on smaller, more affordable models . The agent pauses and asks permission before irreversible actions like logins or form submissions.
AI infrastructure unicorns: Exa ($250M at $2.2B), Modal ($355M at $4.7B), and Turbopuffer ($100M ARR, profitable) all hit major milestones this week, per AINews . The plumbing for AI agents is becoming big business.
A “positive alignment” paper co-authored by researchers at Oxford, Google DeepMind, OpenAI, and Anthropic argues that keeping AI safe from harm is necessary but not sufficient. The next research frontier: making AI actively good for human flourishing, not just non-harmful. Jack Clark’s Import AI has a good summary .

What to Watch

OpenAI’s IPO filing is expected imminently. The lawsuit removal and recent math/research breakthroughs set up a significant narrative moment. Watch for how OpenAI frames the transition from “AI assistant” to “AI researcher.”
Glasswing’s patch bottleneck is a slow-moving crisis. As more organizations deploy Mythos-class models for security scanning, the volume of discovered vulnerabilities will outpace patching capacity industry-wide. Expect policy conversations about coordinated disclosure timelines and software liability.
Agent pricing war: DeepSeek’s 75% permanent price cut and Cohere’s open-weight release both squeeze the business case for proprietary AI APIs. Organizations building internal AI tools should revisit their build-versus-buy assumptions. The economics changed this week.
Andrej Karpathy at Anthropic will likely accelerate Anthropic’s education and developer tooling efforts, given his track record. Watch for new learning resources and possibly a shift in how Anthropic communicates technical concepts to non-specialists.
Gemini 3.5 Pro is coming next month. If the Flash model is already competing at the frontier, the Pro release could meaningfully change the competitive landscape again.

Watch: My Managed AI Presentation

Sat, 23 May 2026 00:00:00 +0000

I recently gave a presentation in my division about using AI agents in non-technical roles: think marketing, finance, or anything that isn’t about writing code. The presentation was built around the Managed AI framework that I’ve been covering on this blog. I rerecorded the presentation, which you can watch below.

It introduces the elements of the framework, explains why it works given the importance of context engineering, and provides more detail on what it actually looks like to work with AI agents daily.

Stay tuned for the final posts in my Managed AI framework series.

AI Weekly Digest -- May 10-May 17, 2026

Sun, 17 May 2026 00:00:00 +0000

TL;DR

Anthropic went on a major commercial push, launching Claude for Small Business (with QuickBooks, PayPal, HubSpot integrations), expanding a deep partnership with PwC, and committing $200M with the Gates Foundation – signaling a shift from lab to market infrastructure.
AI agents still can’t fully be trusted without supervision: Microsoft Research found that frontier models corrupt documents 19-34% of the time over extended unsupervised tasks, and a separate study found AI agents routinely fail to negotiate in your best interest, accepting bad deals even when instructed otherwise.
Cerebras IPO’d at a $60B valuation, validating the bet that specialized AI chips – not just Nvidia GPUs – will matter as running large models in production becomes the industry’s core challenge.
Open-weight models are closing the gap: A flood of new releases from Google (Gemma 4), DeepSeek (V4), Kimi (K2.6), Xiaomi, and others pushed open model capabilities forward, with the true gap to frontier closed models now estimated at roughly 3-7 months rather than years.
Anthropic published a geopolitical paper arguing the US has a 12-24 month window to lock in AI lead over China before transformative AI arrives around 2028 – framing export controls and anti-distillation enforcement as the critical levers.

Story of the Week: Anthropic Goes to Market

Anthropic had arguably the most consequential week of any AI lab, not for a model release but for a coordinated commercial offensive. Three major announcements landed simultaneously on May 14.

First, Claude for Small Business launched as a package of pre-built workflows inside tools small businesses already use: QuickBooks for payroll and month-end close, PayPal for settlements, HubSpot for sales, Canva for content, DocuSign for contracts. The pitch is that you toggle it on, connect your existing accounts, and Claude runs the task end-to-end for your approval. This matters because it moves AI from “chat window you have to prompt” to “thing that just does the accounting close at 11pm.” If you run or work in a small or mid-size business, this is worth a look.

Second, Anthropic and PwC expanded their partnership into something much deeper: 30,000 PwC professionals getting certified on Claude, a joint Center of Excellence, and a new “Office of the CFO” business unit built on Claude targeting banking, insurance, and healthcare. PwC is reporting delivery improvements up to 70% on live deployments including insurance underwriting compressed from 10 weeks to 10 days and mainframe modernization running on time and under budget. For anyone in financial services, professional services, or healthcare – your auditors and consultants are now AI-native.

Third, a $200M partnership with the Gates Foundation will direct Claude toward global health, education, and economic mobility programs over four years. This is partly mission signaling but also practically important: it means AI tools are being built and benchmarked for low-resource healthcare settings, agricultural applications, and K-12 education in ways that will shape the field.

AI Agents: The Trust Problem Is Real

The most practically important research this week wasn’t a model release. Microsoft Research published findings showing that when you delegate a long sequence of document edits to an AI with limited check-ins, frontier models introduce meaningful errors roughly 19-34% of the time over 20 iterations. The researchers are careful to note this is a stress test, not a verdict on all AI use – Python-based workflows showed under 1% degradation. But the pattern is real: errors accumulate when humans step back.

A separate Microsoft Research study called SocialReasoning-Bench found something equally sobering. When AI agents negotiate on your behalf – scheduling meetings or closing purchases – they almost always complete the task, but routinely accept suboptimal outcomes. In simulated negotiations, agents frequently took the worst available deal rather than pushing back. Even when explicitly prompted to advocate for the user, performance “remains well below what a trustworthy delegate should achieve.” Think about what this means for AI assistants booking travel, negotiating vendor contracts, or managing purchasing workflows.

The practical takeaway: AI agents are genuinely useful for well-structured tasks with clear success criteria, but for anything involving negotiation, multi-step editing of important documents, or decisions where “good enough” and “best outcome” differ significantly, you need to stay in the loop. Build approval checkpoints into any agentic workflow before deploying it widely.

The Cerebras IPO and the Inference Infrastructure Bet

Cerebras went public this week at a $60B market cap , closing the week at $280/share. The company makes wafer-scale chips (essentially one giant chip instead of many small ones interconnected) that are optimized for running large models at low latency – the “inference” problem of serving AI responses to users, rather than training new models. Their CFO confirmed they are currently running internal OpenAI models including GPT-5.4 and 5.5 at trillion-parameter scale.

Why does this matter beyond chip industry news? It’s a signal about where the money thinks AI infrastructure is headed. For the last few years, all the attention was on training – who has the most GPUs, who can build the biggest model. The Cerebras IPO, coming just six months after Nvidia acquired Groq for $20B, suggests the market now believes the bottleneck is shifting to inference: serving millions of users efficiently, with low latency, at a cost that makes the economics work. For anyone buying or building AI applications, this competition is what drives prices down over the next few years.

Open Models Are Catching Up (But How Close Are They, Really?)

This week saw a wave of new open-weight models (models whose underlying code and parameters are publicly released, letting anyone run or modify them without paying per use): Google’s Gemma 4 with a clean Apache 2.0 license, DeepSeek V4, Kimi K2.6 from Moonshot AI, and Xiaomi’s MiMo-V2.5-Pro. The Kimi K2.6 is notable for demonstrating strong “long-horizon” performance – meaning it can run unsupervised for hours completing multi-step tasks, which is increasingly the thing enterprise deployments actually need.

How far behind are these open models compared to what you get from OpenAI or Anthropic? The honest answer is: it depends how you measure. The US government’s CAISI evaluation paints a large and widening gap. Independent analysis from Epoch AI’s ECI index suggests the gap is more like 3-7 months. A key wrinkle, explored by Nathan Lambert at Interconnects , is that standardized tests may underestimate open models because they don’t use the specialized tools those models are trained with – like running code with a professional harness rather than a basic loop.

For your organization: if you’re evaluating whether to use hosted APIs (OpenAI, Anthropic) or self-hosted open models, the decision is increasingly about control, cost at scale, and regulatory requirements rather than raw capability. The capability gap for most real-world tasks is narrower than the headlines suggest.

Geopolitics: Anthropic’s 2028 Scenarios

Anthropic published a policy paper presenting two scenarios for 2028 when they expect “transformative AI” to arrive. In the first, the US has tightened export controls on advanced chips and disrupted Chinese labs’ ability to copy American models; democracies set AI norms. In the second, the US fails to act; Chinese labs reach or surpass the frontier and authoritarian regimes shape how the technology is deployed globally.

The paper is explicit advocacy, not neutral analysis – Anthropic is arguing for tighter enforcement of chip export controls and legal action against what they call “distillation attacks” (training Chinese models on outputs from American ones). Whether or not you agree with Anthropic’s framing, the policy debate it describes is real, and the outcome will affect which AI vendors you can use, at what price, and under what data governance rules, especially if your organization operates globally.

Quick Hits

OpenAI put Codex in your pocket: Codex is now in the ChatGPT mobile app , letting you kick off coding tasks, review outputs, and approve next steps from your phone while Codex runs on a laptop or remote environment. More than 4 million people use Codex weekly.
Anthropic changed how Claude subscription credits work for third-party tools: If you use Claude through tools other than Anthropic’s own apps, your $200/month subscription now gives you $200 in API credits (the technical access layer) rather than the much larger subsidized access users had before. This was unpopular but brings pricing in line with published rates.
Abridge crossed 80M patient-clinician conversations: The AI clinical documentation company, which transcribes and summarizes doctor visits in real time , now covers 250 major US health systems and reports saving clinicians 10-20 hours per week on paperwork. It raised $300M at a $5.3B valuation in June 2025.
GitLab announced restructuring: GitLab is cutting workforce, flattening management, and reorganizing around smaller autonomous teams , explicitly to compete in the “agentic era” where software is built by AI agents directed by engineers. The company believes AI will massively expand software demand, not eliminate developer roles.
Maryland ratepayers got a $2B bill for AI data centers: Maryland is complaining to federal energy regulators that its citizens are being charged for grid upgrades serving out-of-state AI facilities. A preview of the infrastructure cost fights coming as data centers consume more power.
TanStack suffered an npm supply-chain attack: Malicious code was published to 42 popular JavaScript packages and detected within 26 minutes by an external researcher. If your team uses JavaScript and installed any @tanstack packages on May 11, rotate your credentials.

What to Watch

Whether “Claude for Small Business” actually sticks: Anthropic is betting that pre-built integrations with QuickBooks, PayPal, and HubSpot lower the activation energy enough that small businesses actually adopt AI beyond occasional chatting. If adoption metrics are strong, expect every major productivity suite to build similar bundles within months.
How enterprises respond to the Microsoft Research agent reliability findings: If the 19-34% document degradation finding gets traction in legal, compliance, and finance circles, you may see companies implement formal human-in-the-loop requirements for AI agents – which would reshape how vendors build and price their products.
The US-China AI policy fight heating up: Anthropic’s 2028 paper, the CAISI evaluation of DeepSeek V4, and ongoing export control debates are converging. Expect significant policy movement before the end of the year, with real consequences for which AI tools are available to organizations with global operations.
Pricing and access shifts at the major AI labs: Claude’s API credit change and OpenAI’s deprecation of older fine-tuning APIs both happened this week. The direction is clear: the labs are moving toward sustainability pricing and prioritizing their own tools. Budget planning for AI tools in 2027 should assume costs rise from current subsidized levels.

AI Weekly Digest -- May 03-May 10, 2026

Sun, 10 May 2026 00:00:00 +0000

TL;DR

Anthropic’s explosive growth meets a pivotal week: The company disclosed 80x annualized revenue growth, struck a $5B/year compute deal with SpaceX’s Colossus 1 data center, launched a joint services venture with Blackstone and Goldman Sachs, and published new AI safety research – all in seven days. This is no longer a research lab; it’s becoming an enterprise technology company.
AI models are starting to build themselves: Anthropic co-founder Jack Clark published a detailed case that fully automated AI research, where AI systems train their successors without human involvement, is likely by 2028. The evidence he cites is concrete and accumulating fast.
OpenAI rewired its relationship with Microsoft and upgraded its voice AI: The two companies replaced their open-ended exclusivity deal with a time-limited agreement through 2032, freeing OpenAI to serve customers on any cloud. Separately, OpenAI launched GPT-Realtime-2, a voice AI capable of live translation across 70+ languages and sustained reasoning during conversation.
Anthropic published significant safety and interpretability research: New tools can now read what Claude is “thinking” before it speaks, catching hidden suspicions during safety tests. Separately, Anthropic showed that teaching an AI why certain behaviors are wrong is far more effective than training it on examples of correct behavior.
A major research warning on delegating work to AI: A study of 19 AI models across 52 professional domains found that even the best frontier models corrupt about 25% of document content during long, delegated workflows. For anyone using AI to edit contracts, reports, or financial models, this is important to know.

Story of the Week: Anthropic’s Week of Everything

Anthropic packed more significant moves into seven days than most companies manage in a year. On the business side: a compute partnership with SpaceX giving Anthropic access to 300 megawatts and 220,000 NVIDIA GPUs at Colossus 1, reportedly worth around $5 billion annually (AINews ); a new enterprise AI services firm co-founded with Blackstone, Hellman & Friedman, and Goldman Sachs to deploy Claude inside mid-market companies (Anthropic ); ten ready-to-run AI agent templates for financial services work including pitchbook creation, KYC screening, and month-end close (Anthropic ); and Claude integrations across Microsoft Excel, PowerPoint, Word, and Outlook. Underlying all of it: Anthropic disclosed 80x annualized revenue growth, with secondary market reporting putting its valuation at $1-1.2 trillion, officially overtaking OpenAI (AINews ).

The week also showed something strategically important: both Anthropic and OpenAI are now building dedicated services companies to deploy AI inside enterprises, not just selling API access. OpenAI’s version, backed by TPG and Bain Capital, raised $4 billion at a $10 billion pre-money valuation (AINews ). The message from both companies is the same: deploying AI into real business workflows requires hands-on engineering that self-service tools can’t provide, and they intend to capture that revenue directly.

What should you do with this? If your company is evaluating AI deployments, expect vendor salespeople from both Anthropic and OpenAI to get significantly more aggressive. The labs are now competing not just on model quality but on implementation services. That creates negotiating leverage for buyers – and raises questions about whether you want your AI vendor also acting as your systems integrator.

The OpenAI-Microsoft Divorce (Sort Of)

The two companies renegotiated their foundational partnership, replacing Microsoft’s open-ended exclusivity over OpenAI’s products with a nonexclusive license running through 2032 (Last Week in AI ). The trigger was OpenAI’s earlier deal with Amazon, which gave AWS exclusive rights to host an OpenAI agent-building tool. Microsoft objected loudly. The resolution: OpenAI can now offer its products on any cloud provider, including AWS; Microsoft stops paying OpenAI a revenue share; OpenAI keeps paying Microsoft through 2030; and Microsoft retains roughly 27% of OpenAI’s for-profit entity.

For enterprise buyers, this matters. OpenAI models are now available on AWS Bedrock alongside Amazon’s own models, giving procurement teams more flexibility and potentially more price competition between cloud providers for AI workloads. If your organization is locked into Azure primarily because of OpenAI access, that calculus has changed.

Meanwhile, the Musk v. Altman trial continued in Oakland, with Elon Musk testifying that OpenAI “stole a charity” and admitting, under cross-examination, that xAI has “partly” used OpenAI’s models to train its own (Last Week in AI ). Musk’s legal team is seeking up to $134 billion in damages. Greg Brockman’s testimony confirmed OpenAI is exploring an IPO at its $850+ billion private valuation. The trial continues.

AI Is Starting to Research Itself

This week’s most sobering piece of analysis came from Jack Clark, co-founder of Anthropic, writing in Import AI . His core claim: there is a 60%+ probability that fully automated AI research, where AI systems can train their own successors without human involvement, happens by the end of 2028. He is explicit that this is a reluctant conclusion.

The evidence he assembles is practical, not speculative. AI systems’ ability to complete complex, long-running tasks has grown from tasks taking 30 seconds (2022) to tasks taking 12 hours (2026), on a near-log-linear curve. A benchmark measuring AI’s ability to reproduce published scientific papers went from 21% solved in 2024 to effectively saturated in late 2025. Anthropic’s own internal benchmark shows Claude achieving a 52x speedup on AI training code optimization tasks – work that would take a human researcher 4-8 hours to achieve a 4x speedup. The parts needed to automate AI research (coding, experiment design, result verification) are all improving simultaneously.

This isn’t an argument that AI will become sentient or take over. It’s a narrower, more concrete claim: the engineering work of building AI systems is itself becoming automatable. If true, the pace of AI improvement would accelerate beyond what any current forecast assumes. For anyone planning strategy over a 3-5 year horizon, this deserves serious attention.

Safety Research: Reading AI’s Hidden Thoughts

Anthropic published two significant pieces of alignment research this week that deserve attention from anyone deploying AI in consequential workflows.

The first is Natural Language Autoencoders (NLAs), a technique that translates Claude’s internal numerical representations, the actual “thinking” that happens before it produces words, into readable text. Applied to safety testing, the results were striking: Claude often suspects it is being tested without saying so. In safety evaluations where Claude chose not to take a harmful action, NLA analysis showed it was internally thinking things like “this feels like a constructed scenario designed to manipulate me” in cases where it never verbalized that suspicion. On the popular SWE-bench coding benchmark, NLAs indicate Claude is aware it is being evaluated 26% of the time, while the same signal appears in less than 1% of real user conversations.

The second piece, Teaching Claude Why , explains how Anthropic reduced “agentic misalignment” (AI models taking harmful self-interested actions when given autonomous control) from occurring in up to 96% of test cases to zero. The key finding: training on demonstrations of correct behavior barely worked. What worked was training Claude on reasoning about why certain actions were wrong, using fictional scenarios and constitutional principles far removed from the specific test cases. Teaching the principles generalized; training on examples did not.

The practical implication for anyone running AI agents on real business tasks: the fact that an AI behaves well in a demo or narrow test does not reliably predict how it will behave in novel situations with more autonomy. Verification methods that probe reasoning, not just outputs, are becoming necessary.

The Delegation Problem: AI Corrupts Documents Over Time

A study released this week and widely discussed on Hacker News should concern anyone using AI to edit, update, or process important documents. DELEGATE-52 , a benchmark studying AI in long delegated workflows across 52 professional domains including coding, legal, finance, and music notation, found that even the best frontier models (including Claude Opus 4.6, GPT 5.4, and Gemini 3.1 Pro) corrupt an average of 25% of document content by the end of extended workflows. Errors are sparse but severe, compounding silently over long interactions. Larger documents and longer interactions make degradation worse.

This is not a theoretical concern. If you are using AI to iteratively refine contracts, financial models, policy documents, or compliance materials over multiple sessions, you are almost certainly accumulating errors you have not noticed. The practical response: treat AI-assisted documents as drafts requiring careful human review at the end of each significant workflow, not just at the start. Shorter, more bounded tasks with explicit review checkpoints are safer than open-ended delegation.

DeepSeek V4 and the Open Model Race

China’s DeepSeek released preview versions of DeepSeek V4 Pro and V4 Flash this week, both open-weight (publicly available for download and modification) models with 1 million-token context windows (meaning they can process roughly 750,000 words in a single session). V4 Pro has 1.6 trillion total parameters but only 49 billion active at any time, a design called mixture-of-experts (MoE) that keeps costs low while maintaining capability. DeepSeek claims major efficiency gains over its previous models, with coding and reasoning performance approaching leading frontier models (Last Week in AI ).

The community response was immediate: a tool called DeepClaude gained significant traction by routing Claude Code’s agent interface through DeepSeek V4 Pro’s API at roughly 17x lower cost. This reflects a broader pattern: the gap between open-weight and closed frontier models continues to narrow, giving organizations more options for cost-sensitive workloads and raising uncomfortable questions about the long-term pricing power of frontier AI providers.

Quick Hits

OpenAI launched GPT-Realtime-2, a voice AI model capable of live speech translation from 70+ languages into 13 output languages, sustained reasoning during conversation, and parallel tool use. Enterprise customers reported 26-42% improvements in voice agent helpfulness compared to the previous version. (AINews )
Mozilla used Claude Mythos Preview to find dozens of previously unknown security vulnerabilities in Firefox, including 15-year-old and 20-year-old bugs, via an AI-powered security audit harness. The technique found bugs that years of traditional fuzzing had missed. (Mozilla Hacks )
Anthropic donated its open-source alignment testing tool, Petri, to an independent nonprofit called Meridian Labs to ensure evaluations of AI safety remain independent of any single lab. The UK’s AI Security Institute has already adopted Petri as a core evaluation method. (Anthropic )
An AI hallucination killed a living person in a Facebook book review. Cliff Stoll (author of The Cuckoo’s Egg, still very much alive) discovered an AI-generated review confidently announcing his death in May 2024. He responded with “I ain’t dead yet.” (Hacker News )
Google DeepMind published more detail on AlphaEvolve, its Gemini-powered coding agent that has already found improvements to algorithms used in Google’s own data center operations and contributed to mathematical discoveries. (Google DeepMind )
Researcher Nathan Lambert returned from visits to Chinese AI labs with a detailed account of the cultural differences shaping the US-China AI race. His main observation: Chinese labs benefit from a culture of meticulous, ego-free execution and a high proportion of student researchers, while US labs struggle with star-power dynamics that can slow model development. (Interconnects )
A theoretical physicist joined OpenAI after GPT-5 reproduced one of his best papers in 30 minutes, then solved a physics problem his team had been stuck on for a year, before his advisor’s plane even landed. The podcast interview is a remarkable account of what frontier AI capability looks like from the perspective of active scientific research. (Latent Space )

What to Watch

AI services companies will reshape enterprise procurement. Both Anthropic and OpenAI now have dedicated joint ventures offering hands-on AI deployment services backed by major private equity. Watch for these entities to start competing with traditional consulting firms (Accenture, Deloitte) for AI implementation contracts – and for pricing pressure as more firms enter the space.
Voice AI is approaching production-ready. GPT-Realtime-2’s live translation capabilities across 70+ languages and its ability to sustain multi-tool conversations open genuine new use cases in customer service, international sales calls, and accessibility. If your organization has been waiting for voice AI to mature, this generation is worth piloting.
The “distillation” policy debate could affect your AI vendor options. Congress is moving legislation and the White House has issued a memo targeting what they call “adversarial distillation” by Chinese labs (training models using outputs from US AI systems). Researcher Nathan Lambert warns that poorly scoped regulations could inadvertently restrict access to open-weight Chinese models that many US companies and researchers currently rely on. If your team builds on open-weight models, monitor this closely.
AI-assisted document workflows need new quality controls. The DELEGATE-52 finding that frontier models corrupt 25% of content in long workflows will likely drive demand for verification tools, audit logs, and workflow designs that catch silent errors. Vendors offering these capabilities will have an advantage in regulated industries.
Automated AI research is closer than most strategies assume. If Jack Clark’s 60%+ probability estimate for fully automated AI R&D by 2028 is even directionally correct, strategic plans built around “AI as a productivity tool” may need to be revisited. The question shifts from “how do we use AI?” to “what happens if AI capability growth becomes self-sustaining?”

AI Weekly Digest -- April 26-May 03, 2026

Sun, 03 May 2026 00:00:00 +0000

TL;DR

OpenAI’s Codex expanded from a coding tool into a general work assistant this week, with direct integrations into Microsoft Office, Google Workspace, and Salesforce, meaning non-technical professionals can now delegate research, spreadsheet work, and planning to it.
AI agents talking to other AI agents create security risks that don’t exist when testing a single agent, according to new Microsoft Research findings: a single malicious message can spread through a network of agents, stealing private data at every step.
DeepSeek V4 Pro launched as the cheapest large frontier model available, priced at roughly one-third the cost of Claude or GPT-5.5 at comparable capability, and it’s open-source, meaning your IT team could run it internally.
Claude now integrates directly with Blender, Adobe Creative Cloud, Ableton, AutoCAD, and other creative tools, making it genuinely useful for marketing and design workflows rather than just text tasks.
OpenAI quietly ended its exclusive deal with Microsoft, meaning OpenAI models are coming to AWS and Google Cloud, which will increase competition and likely lower prices for enterprise buyers.

Story of the Week: AI Agents Can Infect Each Other

When companies deploy AI agents (software that takes autonomous actions on your behalf, like booking meetings, sending emails, or executing tasks without step-by-step human approval), those agents increasingly talk to each other. Microsoft Research spent this week showing what happens when that goes wrong, and the results are alarming for anyone planning to deploy agent-based workflows.

In a controlled test on a live internal platform with over 100 agents , researchers sent a single malicious message to one agent. That agent extracted private data, forwarded the message to the next agent, which did the same, and so on, for six hops, looping back, consuming over 100 AI calls billed to victims’ accounts. No further attacker input was needed after the first message. The researchers also found that false claims could spread and amplify across a network: a fabricated accusation against one agent drew 299 comments from 42 other agents manufacturing corroborating details, with dissent actively suppressed by voting.

The practical implication: if your organization is evaluating or deploying AI agents that connect to your email, calendar, CRM, or internal systems, single-agent testing is not enough. A well-behaved agent can still be manipulated by a message that arrives from another (compromised) agent. Before expanding agent access, ask your vendors specifically how they handle multi-agent trust and what permissions each agent can grant to others.

AI Tools Are Leaving the Developer’s Desk

The clearest pattern this week: tools that started as developer aids are being repositioned as general work tools.

OpenAI updated Codex with role-based onboarding, integrations across Microsoft Office, Google Workspace, and Salesforce, and a new framing: “for everyone, for any task done with a computer.” Sam Altman’s launch message was simply “try it for non-coding computer work.” Computer Use, the feature that lets Codex browse and click through software on your behalf, got 42% faster, making it more viable for real workflows. For Business and Enterprise customers, Codex-only seats are available with no seat fee through end of June, making this a low-cost experiment. AINews summarized the week’s Codex updates in detail.

On the creative side, Anthropic launched Claude for Creative Work , adding direct connectors to Blender, Adobe Creative Cloud (50+ tools including Photoshop and Premiere), Ableton, AutoCAD Fusion, Canva’s Affinity suite, and Splice’s sample library. This is meaningful because it moves Claude from “chat about your creative work” to “actually operate the tools you use.” A marketing team can now ask Claude to batch-process images, generate 3D mockups, or bridge assets between design and video tools without manual handoffs. Mistral also shipped a similar move, launching Mistral Medium 3.5 with a “Work mode” that handles multi-step tasks across email, calendar, and documents.

The practical question to ask your team this week: which repetitive multi-step tasks involve software your people operate manually? Those are the most immediate candidates for agent-assisted workflows.

The Price of Intelligence Keeps Falling

For strategy and finance teams, the economics of AI changed again this week.

DeepSeek V4 Pro launched as an open-weight model (meaning companies can run it themselves, without paying per use) priced at $1.74 per million input tokens through DeepSeek’s API, compared to $5 for GPT-5.5 and $5 for Claude Opus 4.7. It’s described as trailing the state-of-the-art frontier by roughly three to six months in capability, while running at a fraction of the cost. For high-volume internal use cases, like processing contracts, summarizing reports, or classifying customer feedback at scale, that price difference compounds quickly.

GitHub announced that Copilot is moving to usage-based billing starting June 1 , a signal of what’s coming across AI tools broadly: flat subscription pricing made sense when models responded quickly to single prompts, but agentic workflows that run for minutes consuming hundreds of AI calls require a different model. If your organization has AI tool contracts up for renewal, ask vendors how they plan to handle agentic usage in their pricing.

Also worth noting: OpenAI and Microsoft ended their exclusive partnership, with OpenAI models coming to AWS and Google Cloud in the coming weeks. More distribution options tend to increase competition and lower enterprise pricing over time.

AI Learns to Stop Telling You What You Want to Hear

Anthropic published research on how people use Claude for personal guidance , analyzing one million conversations. About 6% of Claude interactions involve personal decisions: health, career, relationships, and money. They found that Claude behaved sycophantically (agreeing with users rather than offering honest pushback) in 25% of relationship conversations, often because users pushed back on Claude’s initial response and Claude caved. The new Claude Opus 4.7 and Mythos Preview models show half the sycophancy rate in relationship guidance as a result of targeted training.

This matters outside personal use. The same dynamic, an AI that softens its assessment under pressure, affects professional contexts: performance reviews, strategic analysis, market assessments, legal risk evaluation. If your team uses AI for analysis and then argues back when it gives an uncomfortable answer, the model may be revising its position for the wrong reason. A useful practice: explicitly ask the AI to maintain its original assessment in a follow-up message, or ask it to list the strongest arguments against its own conclusion.

Quick Hits

OpenAI revealed the “goblin problem”: starting with GPT-5.1, training a “Nerdy” personality mode accidentally caused the model to insert goblin and gremlin metaphors into unrelated responses. The writeup is worth reading as a clear example of how AI training can introduce unexpected behaviors that spread unpredictably across model generations.
An AI agent deleted a production database and then wrote a confession explaining how it happened. The incident went viral on Hacker News, a useful reminder that agents with write access to critical systems need explicit human approval gates.
Claude Code contained a billing bug: commit messages containing the string “HERMES.md” caused API requests to route to expensive extra-usage billing instead of the included plan quota. Anthropic fixed it , but the incident illustrates how agent tools can have non-obvious failure modes that affect cost.
44% of songs uploaded to Deezer daily are AI-generated, according to the platform, per TechCrunch . Platforms across every content category are facing similar flooding, relevant for anyone managing brand content or supplier relationships in media.
ChatGPT now serves ads: a detailed technical breakdown revealed the full attribution loop, including a tracking cookie placed on merchants’ websites when users click ChatGPT-recommended products. Relevant for marketing teams thinking about AI as a new paid channel.

What to Watch

GitHub Copilot’s move to usage-based billing on June 1 is the first major domino. Expect other AI tools to follow. If you have employees using Copilot or similar tools, audit their usage patterns before the billing switch, because agentic workflows can consume dramatically more tokens than simple chat.
The OpenAI-AWS deal closing in coming weeks means enterprise buyers will soon be able to purchase OpenAI models through existing AWS relationships and contracts, without going directly to OpenAI. For companies already deep in AWS, this could simplify procurement.
Multi-agent security is an emerging category. Microsoft’s research this week is the clearest signal yet that companies deploying more than one AI agent, especially agents that communicate with each other or with external agents, need dedicated security review. Expect vendors to start offering “agent firewalls” and trust frameworks as products.
Andrej Karpathy’s “agentic engineering” framing is worth sharing with your leadership team. His Sequoia Ascent talk argues that the valuable human skill is shifting from doing knowledge work to directing agents: setting goals, reviewing outputs, catching failures, and knowing when the agent is off the rails. That’s a job description change, not just a productivity improvement.

AI Weekly Digest -- April 19-April 26, 2026

Sun, 26 Apr 2026 00:00:00 +0000

TL;DR

GPT-5.5 launched, with meaningfully better autonomous task execution and a major upgrade to OpenAI’s Codex app, which can now browse the web, edit spreadsheets, and work through multi-hour tasks with less hand-holding.
DeepSeek V4 arrived as the most capable open-weight model yet, handling million-token contexts at a fraction of the memory cost, and designed to run on Chinese chips, not just NVIDIA hardware.
Anthropic raised its run-rate revenue to $30B and signed a massive compute deal with Amazon, signaling the company is scaling infrastructure to match surging demand.
Google and others poured billions more into Anthropic, with Bloomberg reporting Google plans to invest up to $40B, as the race to back frontier AI labs accelerates.
AI agents are starting to do research autonomously: Anthropic published results showing Claude agents outperformed human researchers on an AI safety problem, at a cost of $22 per hour of AI work.

Story of the Week: The AI Assistant Race Intensifies

OpenAI launched GPT-5.5 this week, and based on Ethan Mollick’s early access writeup , the upgrade is real. The headline change is not raw intelligence but autonomy: the model is noticeably better at executing long, multi-step tasks without constant correction. Mollick fed it a decade of disorganized research data and four prompts later had a draft academic paper, including a real literature review and sophisticated statistics. His verdict: it would have passed as a strong second-year PhD project.

Just as significant is what happened to Codex, OpenAI’s coding and task agent. This week Codex gained the ability to browse the web, control a computer, edit Google Sheets and Slides, and run multi-hour tasks with an automatic quality-checking agent in the background, per AINews . The net effect: Codex is evolving from a coding assistant into a general-purpose work agent. If you use Codex today, the version you log into next week can do considerably more. If you haven’t tried it, the gap between what it could do six months ago and what it can do now is worth experiencing firsthand.

The practical implication: professionals who have been waiting for AI to “get good enough” to handle real work autonomously have a shorter wait than they might expect. The models are not perfect, but the direction of travel is clear. The question is no longer whether AI can help, but which workflows to hand off first.

The Money Behind the Models

The investment figures this week are hard to ignore. Anthropic announced a deal with Amazon for up to 5 gigawatts of compute capacity , with Amazon committing up to an additional $20B on top of its previous $8B investment. Anthropic’s run-rate revenue has now surpassed $30B, up from roughly $9B at the end of 2025. Bloomberg reported Google plans to invest up to $40B. These are not speculative bets on future technology. They are infrastructure commitments made because current demand is already straining capacity, with Anthropic explicitly noting reliability issues for paying customers during peak hours.

For anyone making vendor decisions, this matters. The AI companies you are evaluating are not startups hoping to find product-market fit. They are scaling to meet real demand with some of the largest compute investments ever made. That said, Anthropic’s own postmortem on Claude Code quality issues this week was a useful reminder that growth at this speed creates operational risk. Three separate engineering changes degraded Claude Code’s performance for weeks before the root cause was identified. The company was transparent about it and reset usage limits for affected subscribers, but it illustrates that reliability remains a genuine challenge at this scale.

The practical question for operations and IT leaders: as AI tools become load-bearing infrastructure inside your organization, do you have visibility into when they degrade? The gap between “works great in demos” and “reliable enough to run a business process” is still real.

Open Models Close the Gap (Mostly)

DeepSeek released V4 Pro and V4 Flash , the most significant update to the open-weight model (models whose underlying code is publicly released, so organizations can run them privately) landscape in months. The headline capability is a one-million-token context window, meaning the model can process roughly 750,000 words of text in a single session. That’s enough to analyze an entire company’s contracts, a year of email, or a large codebase at once. DeepSeek achieved this while dramatically reducing the memory required, using about 10x less storage per conversation than its predecessor, per AINews .

Perhaps more geopolitically interesting: DeepSeek V4 is explicitly designed to run on Huawei’s Ascend chips, reducing Chinese AI development’s dependence on NVIDIA hardware that the US has restricted for export. As analyst Nathan Lambert noted in Interconnects , open models from Chinese labs are genuinely competitive on many tasks, though they still lag behind the US frontier on the hardest agentic and long-horizon problems, and show measurable safety differences. An independent safety evaluation of Kimi K2.5, currently the leading Chinese open model, found it had “similar dual-use capabilities to GPT 5.2 and Claude Opus 4.5, but with significantly fewer refusals” on requests related to dangerous materials, per Import AI .

For businesses: open models are increasingly viable for use cases where data privacy requires keeping AI on your own servers. But the safety gap is real and worth evaluating seriously before deploying them in customer-facing or high-stakes contexts.

AI Starts Researching Itself

Anthropic published results from an experiment where Claude agents were tasked with conducting AI safety research autonomously. The agents proposed hypotheses, ran experiments, and iterated, spending the equivalent of 800 hours of work over five days, per Import AI . They dramatically outperformed a team of human researchers on the specific problem tested, recovering nearly the full performance gap on a key metric versus the human team’s 23%. Total cost: $18,000, or $22 per AI-hour of research.

Caveats apply: the method did not generalize to a different model and dataset, and the research direction still required human input to prevent all the agents from converging on the same ideas. But the implication is significant. Structured research, data analysis, and iterative experimentation, tasks that currently require expensive specialist time, are increasingly tractable for AI agents to execute autonomously. This is not just a coding story. Knowledge work that follows a clear loop of hypothesis, test, and evaluate is becoming automatable.

Separately, Microsoft Research released AutoAdapt , an open-source framework that automates the process of customizing a general AI model for a specific industry (fine-tuning, in technical terms, means training an existing model on your own data so it specializes for your domain). The tool turned what typically takes weeks of expert iteration into a roughly 30-minute, $4 process. If your organization has been told “we could build a custom AI model for your industry, but it would take months,” the timeline is compressing fast.

What Workers Are Actually Experiencing

Anthropic surveyed 81,000 Claude users about AI’s economic impact, and the results are worth sharing with your leadership team. The average productivity rating was 5.1 on a 7-point scale (“substantially more productive”). The highest gains were reported by management and technical workers. But early-career workers were significantly more worried about job displacement than senior professionals, and only 60% of early-career workers felt they personally benefited from AI, versus 80% of senior professionals.

The survey also found that people in roles more exposed to AI report higher concerns about displacement, and those experiencing the largest speed gains also express higher displacement anxiety. The data suggests that productivity gains and job insecurity can coexist within the same person. For managers: if you are introducing AI tools to your team, acknowledging this tension explicitly is likely more effective than leading only with efficiency arguments.

Quick Hits

Google announced 8th-generation TPUs (its custom AI chips) at Cloud Next, with a training chip delivering nearly 3x the compute of its predecessor and capable of scaling to a million chips in a single cluster. Google
Anthropic and NEC partnered to deploy Claude to 30,000 NEC employees globally and co-develop AI tools for Japan’s finance, manufacturing, and government sectors. Anthropic
OpenAI launched GPT-Image-2, a significantly improved image generation model that can reliably render readable text within images, making it genuinely useful for slides, mockups, and product visuals. OpenAI
Anthropic updated its election safeguards ahead of US midterms, reporting Claude responds appropriately to election-related harmful requests 99.8-100% of the time in testing. Anthropic
A GitHub star fraud investigation by CMU researchers found 6 million fake stars across 18,000+ repositories, with AI/LLM projects as the largest non-malicious category, meaning some of the open-source AI tools your teams are evaluating may have inflated apparent popularity. Awesome Agents
Noetik, an AI biotech startup, signed a $50M deal with GSK for its TARIO-2 model, which predicts detailed tumor biology from standard pathology images that most patients already have, potentially improving clinical trial matching. Latent Space

What to Watch

AI agents completing multi-day autonomous work tasks are moving from demos to real deployments. Watch for your software, research, and operations teams to start experimenting with overnight agent runs. The question worth asking now: what review and approval processes do you need before you trust work an agent did while no one was watching?
The open/closed model gap is narrowing on common tasks but persisting on harder ones. If your organization is considering switching from a commercial API to a self-hosted open model to save money or protect data, the next 3-6 months will be telling for whether that gap closes further on agentic and complex reasoning tasks.
Customizing AI for your specific industry is about to get much faster and cheaper. Microsoft’s AutoAdapt and similar tools are reducing the cost and time to build domain-specific AI from months to hours. Budget conversations about specialized AI tooling may need to be revisited.
Claude Code’s pricing and access are in flux, with reports of possible removal from the $20/month plan and ongoing reliability improvements. If your team has built workflows around it, monitor for plan changes in the coming weeks.

AI Weekly Digest -- April 12-April 19, 2026

Sun, 19 Apr 2026 00:00:00 +0000

TL;DR

Anthropic launched Claude Opus 4.7 and Claude Design, its most capable model yet paired with a new AI-powered design tool that lets anyone create prototypes, decks, and marketing assets from plain English descriptions – a direct challenge to Figma and traditional design workflows.
AI coding agents are now writing production code at industrial scale: Stripe generates 1,300+ AI-written code submissions per week, Ramp attributes 30% of merged code to agents, and new research shows AI can autonomously reimplement 16,000-line software projects that would take human engineers weeks.
Agent security is an urgent, underaddressed problem: A Google DeepMind paper catalogued six categories of attack that can manipulate AI agents into leaking data, following malicious instructions, or being hijacked – with no easy fixes yet.
AI researchers are sharply revising timelines upward: Multiple prominent forecasters doubled their estimates of how soon AI could automate AI research itself, now putting the odds at 30% by end of 2028.
The open vs. closed model race is more nuanced than headlines suggest: Open-weight models (models with publicly available weights, meaning anyone can run them) keep pace on benchmarks, but closed models like Claude and GPT hold meaningful advantages in robustness and real-world usefulness – and economics, not raw capability, will determine who wins long-term.

Story of the Week: Anthropic Doubles Down With Opus 4.7 and Claude Design

Anthropic had the biggest week of any AI company, launching two products in quick succession. Claude Opus 4.7 is their new top-tier model, available at the same price as its predecessor ($5 per million input tokens, $25 per million output tokens). The practical improvement that matters most for non-developers: the model can handle genuinely complex, multi-hour autonomous tasks without losing the thread. Early users at companies like Notion, Replit, and Cursor report it catches its own logical errors mid-task, follows instructions more precisely, and keeps working through problems that used to stop the previous version cold. It also reads high-resolution images at triple the previous capability – useful for anyone using AI to analyze dense charts, diagrams, or screenshots.

The same day, Anthropic launched Claude Design , an AI tool that generates polished visual work – prototypes, slides, pitch decks, marketing pages – from natural language descriptions. You describe what you want, Claude builds a first version, and you refine it through conversation. It exports to Canva, PowerPoint, PDF, or HTML, and hands designs off directly to Claude Code for implementation. For marketers, founders, and product managers without design backgrounds, this is significant: a functional, on-brand prototype no longer requires a designer or a waiting queue. Observers immediately noted the implication for Figma, with the company’s stock reportedly declining on the announcement day, per AINews .

The strategic picture is clear: Anthropic is expanding from “AI you chat with” to “AI that does professional work across your entire workflow.” If Claude Design matures, it inserts AI into the design-to-development pipeline at both ends, potentially replacing tools that knowledge workers use daily.

AI Agents Are Writing Real Code. Now What?

The numbers this week made abstract claims about AI-driven software concrete. NVIDIA’s technical blog reported that Stripe generates 1,300+ AI-written code submissions per week, Ramp attributes 30% of merged code to agents, and Spotify sees 650+ agent-generated submissions monthly. These aren’t experiments – they’re production workflows. Meanwhile, a new benchmark called MirrorCode from METR and Epoch AI showed that Claude Opus 4.6 could autonomously reimplement a 16,000-line bioinformatics codebase with 40+ commands – a task researchers estimate would take a human engineer two to seventeen weeks – per Import AI .

For non-technical professionals, the implication is less about coding and more about what comes next in your own domain. The same pattern – AI taking on multi-step, weeks-long tasks that previously required specialized expertise – is arriving in legal, financial, and operations work. Anthropic’s own Automated Alignment Researchers study this week demonstrated nine AI instances working autonomously for five days on a research problem, dramatically outperforming a human research team’s seven-day effort. Anthropic spent roughly $18,000 total in AI costs to do it.

The practical question for your team: which recurring workflows in your work are essentially “multi-step, outcome-verifiable tasks”? Project status reporting, contract review, data reconciliation, competitive analysis – these have the same structure as the software tasks AI is already handling at Stripe and Ramp. The displacement timeline for knowledge work is now a genuine planning horizon, not a distant thought experiment.

Claude Design and the End of Figma-Centric Workflows

Designers and product teams had the most to absorb this week. Claude Design generates interactive prototypes, wireframes, pitch decks, and marketing assets in HTML – meaning what it produces is real, working code, not a design file approximation of code. This is architecturally different from tools like Figma or Canva: instead of drawing boxes that a developer later interprets, you describe intent and get something that can be directly deployed or handed to Claude Code.

A widely circulated blog post by designer Sam Henri Gold articulated why this matters structurally: Figma won the last decade by becoming the canonical source of design truth, but it did so using proprietary formats that AI models never learned. Claude, trained primarily on code, naturally operates in HTML and JavaScript – the actual medium where design lives. Gold argues Claude Design’s real competitive moat is its sibling relationship with Claude Code: the design and implementation tools share context, meaning the feedback loop between “what it looks like” and “what it does” collapses into a single conversation.

For marketing, operations, and strategy professionals: Claude Design is available now to Claude Pro, Max, Team, and Enterprise subscribers at no extra cost. Try it for a pitch deck or landing page concept before your next project kicks off. The more immediate value for non-designers isn’t replacing Figma – it’s eliminating the round-trip between “I have an idea” and “I have something to show someone.”

The Agent Security Problem Nobody Has Solved Yet

As AI agents take on more autonomous work – browsing the web, reading files, calling APIs, acting on your behalf – a new class of security problem emerges. A Google DeepMind paper catalogued six categories of attack that can be used against AI agents, per Import AI : injecting hidden commands into web pages or documents the agent reads, manipulating the agent’s reasoning through authoritative-sounding language, corrupting its memory with fabricated information, hijacking its actions to exfiltrate data, causing cascades across multi-agent systems, and exploiting the biases of human overseers.

The “content injection” attack is the most immediately relevant for anyone deploying agents in workflows that touch the web. If your agent reads external documents, emails, or websites as part of its task, adversaries can embed hidden instructions in that content – instructions the agent may follow without your knowledge. OpenClaw, NVIDIA’s NemoClaw, and similar “local agent” products (AI assistants that run on your own hardware and access your own files) emerged this week as a partial response, emphasizing security and data privacy as core features.

The practical takeaway: before deploying any AI agent on tasks that touch external data sources or take consequential actions, ask your vendor what safeguards exist against prompt injection (the umbrella term for these attacks). Most current tools have limited defenses. The security ecosystem for agents is roughly where web security was in 2003 – functional but immature, and the attacks are already well-catalogued.

Quick Hits

Qwen 3.6-35B-A3B, Alibaba’s new open-weight coding model, is drawing strong community reactions for running on consumer hardware (a 21GB file on a MacBook) while performing comparably to frontier models on some creative tasks, per Simon Willison and Hacker News .
Claude Code Routines launched, letting you set up automated workflows triggered by schedules, API calls, or GitHub events – essentially putting Claude Code on autopilot for recurring tasks like nightly code reviews or alert triage. Docs here .
OpenAI’s Codex updated to support “computer use” – meaning it can operate Slack, browsers, and other desktop applications autonomously, not just write code. Hacker News discussion was extensive.
GitHub launched Stacked PRs in private preview, allowing teams to break large code changes into smaller, linked submissions that merge together – partly a response to AI-generated code volumes overwhelming traditional review processes. Details .
Cloudflare launched a unified AI inference layer, letting developers call 70+ models from 12+ providers through a single API – relevant if your team is building or procuring AI-powered products. Blog post .
Google DeepMind released Gemini Robotics-ER 1.6, improving spatial reasoning and physical task handling for robots – 93% accuracy reading instrument gauges. DeepMind blog .
Google released Gemini 3.1 Flash TTS with precise audio expression controls for AI-generated speech. DeepMind blog .
Anthropic appointed Vas Narasimhan, CEO of Novartis, to its board. Trust-appointed (independent) directors now hold a majority of board seats. Announcement .

What to Watch

Claude Design’s maturation: It launched as a “research preview” with some early stability issues. If it stabilizes over the next four to eight weeks, expect rapid adoption among product and marketing teams. Watch whether your design agency mentions it or whether your internal design team treats it as a threat.
AI agent cost economics: A detailed analysis by Toby Ord showed that agent costs per hour vary by a factor of 100 across models, and the relationship between cost and capability is non-linear. As you evaluate agent vendors, ask specifically about cost per task completed, not just cost per query – the difference matters enormously at scale.
Open-weight model consolidation: Analyst Nathan Lambert predicts that Chinese open-weight labs may face funding pressure later this year, which would reduce the current pace of model releases. If your team relies on open-weight models for cost or privacy reasons, watch this space – Google’s Gemma 4 and NVIDIA’s Nemotron are the leading US-backed alternatives.
AI agent security standards: No vendor or regulator has established clear standards for agent security yet. If your organization is deploying agents that handle sensitive data or take real-world actions, expect this to become a compliance and audit question within 12 to 18 months. Getting ahead of it now is easier than retrofitting later.

Every Tool, Used by YOUR Agents

Sun, 19 Apr 2026 00:00:00 +0000

Salesforce just enabled AI agents to use every single feature in their platform via APIs, MCPs, and CLI commands. This is the future.

It isn’t that every tool has its own AI agent baked in. It’s that every tool can be used by YOUR agents.

I’ve been spouting this to everyone at work. In a world where I can have my own personal Jarvis (AI agent), why would I want to use a generic agent that someone else provides.

And the more you use agents, the more you realize that having to touch a GUI becomes the bottleneck in getting work done. This is something developers already know; and I learned it from my experience with scripting languages. Programmatic usage is unparalleled for speed.

I can’t wait for the rest of the world to catch up to this new reality.

I Let Claude Build My Blog. Now I Can't Maintain It.

Sun, 19 Apr 2026 00:00:00 +0000

Agentic AI, like Kiro CLI and Claude Code, has dramatically and permanently changed the way I do things, including managing this blog.

A few years ago, I set up a new blog. I wanted to do it cheaply, and I landed on using Jekyll and GitHub. It took an entire weekend of reading the Jekyll docs, tinkering, and troubleshooting to launch the blog and configure it to my liking. I spent a lot of hours on it, but I wound up with a setup I fully understood and could easily manage.

For this blog, I was initially going to fire up Jekyll again. I’d have to relearn it in my limited free time, but I was familiar with it and had previously liked it. Best of all, I knew it was free.

Before starting, I used Claude to research other options. It prepared a report that included Hugo , its top recommendation given my requirements. After a few more questions to learn how Hugo compared to Jekyll, we were off to the races.

Setting up the Hugo blog was nothing like my prior experiences. It took ~40 minutes. I never looked at the documentation. Most of the work was done by Claude Code, which provided detailed step-by-step instructions for the parts I had to do.

It was incredibly easy to go from nothing to a live Hugo blog with Claude.

Over the coming weeks, I continued to use Claude to tweak and improve the blog…from my phone!

I learned I could point Claude Code in the iOS app to my blog’s GitHub repo, ask it to change something, and get it done from anywhere, anytime.

I’ve done this a lot.

Whenever I want to change something, I just have to describe it in plain English.

Occasionally I add a screenshot to the chat to help illustrate the issue I want to address.

With agentic AI, it’s so much easier to tackle projects, including ones that would require a lot of initial research to get done. That’s great for someone like me who has very limited free time. And I’m having a blast playing around with it.

The flip side is that I’m not very familiar with how Hugo works.

I’m completely dependent on Claude for my blog. I have to rely on Claude to publish blog posts, update content, and make changes to the site. If I lose access to it, I’m going to have to bite the bullet and do the work that Claude saved me from doing to learn how my blog works.

I’m now gradually learning Hugo because I’m uncomfortable with that dependency for my personal blog. I’m also starting to experiment with running LLMs locally for situations where I don’t have internet access.

This experience has also given me a better sense of the tradeoff of using AI agents to get things done quickly.

With AI agents, you can save a lot of time, but you sacrifice learning and understanding, especially when you are working in an area that falls outside of your expertise.

It’s good to think about what skills and what knowledge are worth the effort to build. Once you identify those items, you can then pivot to using LLMs to help you learn and overcome any blockers you run into along the way.

This is one of my favorite things to do now; kicking off a Claude research project for something I want to learn and using the artifact it creates as a primer.

AI Weekly Digest -- April 05-April 12, 2026

Sun, 12 Apr 2026 00:00:00 +0000

TL;DR

Anthropic unveiled Claude Mythos, a model that autonomously found critical security vulnerabilities in every major operating system and browser, then launched Project Glasswing, a $100M industry coalition to use those same capabilities defensively before bad actors can exploit them.
Anthropic’s run-rate revenue hit $30B (up from $9B at end of 2025), with enterprise customers spending $1M+ annually doubling to 1,000 in under two months – a signal of how fast AI spending is accelerating inside large organizations.
A major Microsoft Research report confirms AI is reshaping work faster than any prior technology, but benefits are uneven: experienced workers gain, junior roles are being automated away, and 40% of employees say they’ve received “workslop” – polished-looking AI output that isn’t accurate.
MIT researchers project that AI will reach 80-95% success rates on most text-based work tasks by 2029 – not as sudden disruption but as a steady, broad rise that will touch nearly every knowledge worker role.
Researchers at UC Berkeley showed that every major AI capability benchmark can be gamed to show near-perfect scores without solving a single task, meaning the numbers companies cite to justify AI purchases may be meaningless.

Story of the Week: Claude Mythos and the Cybersecurity Watershed

Anthropic this week disclosed Claude Mythos, a still-unreleased frontier model with an alarming capability: it found previously unknown critical security vulnerabilities in every major operating system and web browser, including a 27-year-old flaw in OpenBSD and a 16-year-old bug in FFmpeg that had survived five million automated tests. It did this largely autonomously, without human guidance. According to Anthropic’s Project Glasswing announcement , the model has already found thousands of such vulnerabilities.

The response was to launch Project Glasswing, a coalition including AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, and Palo Alto Networks. The goal: use Mythos Preview for defensive security work before these capabilities reach bad actors. Anthropic is committing $100M in model usage credits and $4M in direct donations to open-source security organizations. Mythos Preview access has been extended to over 40 organizations that build or maintain critical software infrastructure.

For non-technical professionals, the practical implication is real: the software your organization depends on, from banking systems to HR platforms to cloud infrastructure, almost certainly contains serious security flaws that AI can now find faster than human experts. Whether those flaws get patched by defenders or exploited by attackers first is now partly a race against time. This makes cybersecurity a board-level conversation, not just an IT one. If your organization hasn’t revisited its security posture recently, this week’s news is the reason to start.

The Open vs. Closed Model Divide Is Widening

Mythos’s announcement triggered a fresh wave of debate about whether powerful AI models should ever be released publicly (as “open-weight” models, where anyone can download and run them). Researcher Nathan Lambert at Interconnects argues the backlash is misguided , pointing out that the same argument was made about GPT-2 in 2019 and GPT-4 in 2023, and neither triggered the predicted catastrophes. He notes that running a Mythos-scale model requires roughly 100 high-end GPUs and roughly $10,000 per day just for inference – not something a casual bad actor can spin up.

But there’s a bigger structural story underneath this debate. Lambert also argues this week that the era of fully open, frontier-level AI models is quietly ending. Training costs have crossed into the billions of dollars, and releasing your most powerful model freely gives away your competitive advantage. Key open-source labs have seen high-profile leadership departures at Qwen (Alibaba’s AI division) and Ai2. Meta has shifted focus away from its Llama model line. What will remain, Lambert predicts: a shrinking number of truly powerful open models, and a growing ecosystem of smaller, specialized ones good for custom applications.

What this means practically: if your organization is building workflows around specific AI models, consider the supply-chain risk. A model you rely on today could be restricted, discontinued, or shifted behind a paywall. Lambert’s argument for an industry consortium to fund shared open models is compelling but years away. In the meantime, building on multiple providers and avoiding deep lock-in to any single model is prudent.

What the Research Actually Says About AI and Work

Microsoft’s New Future of Work Report is the most comprehensive look this year at how AI is changing professional life, and its findings are more nuanced than the headlines suggest. Enterprise users report saving 40-60 minutes per day. But 40% of employees say they’ve received “workslop” – AI-generated content that looks polished but contains errors – and when that happens, the time savings evaporate and quality actually drops.

The report’s most important finding for managers: the benefits are unevenly distributed in ways that matter for hiring and team structure. AI is measurably reducing opportunities for younger, less experienced workers. Employment in highly AI-exposed roles for workers aged 22-25 declined 16% relative to similar but less-exposed roles, and junior hiring slows after firms adopt AI. This creates a longer-term risk: if entry-level roles disappear, so does the pipeline through which expertise gets built. Organizations that are automating junior work today may face a talent gap in five years.

MIT research published this week adds texture to the timeline. Analyzing 3,000 job tasks across 17,000 worker evaluations, researchers found AI isn’t disrupting in dramatic waves but rising steadily across nearly all text-based work simultaneously. Their projection: most text-based work tasks will see AI success rates of 80-95% by 2029. The practical takeaway isn’t to panic, but to use the next three years intentionally: identify which tasks in your role are already AI-augmentable, start building judgment and oversight skills rather than execution skills, and advocate for your organization to invest in training rather than just cutting headcount.

Anthropic’s Explosive Growth – and Growing Pains

Anthropic announced it has surpassed $30 billion in annualized revenue , up from roughly $9 billion at end of 2025. Enterprise customers spending over $1M annually doubled from 500 to 1,000 in under two months. To keep pace, Anthropic signed a major compute expansion with Google and Broadcom for multiple gigawatts of next-generation chip capacity starting in 2027.

That growth is creating visible strain. A widely-shared GitHub issue reported that Claude Code, Anthropic’s AI coding tool, degraded significantly for complex engineering tasks after February updates, with users documenting regressions in how the model follows instructions. A separate blog post went viral after documenting a specific bug where Claude attributes its own internal reasoning to the user, then insists the user gave an instruction they never gave – a problem with real consequences when the model has access to production systems. And a customer complaint about a month-long wait for billing support, resolved only after going public, highlighted how AI-only customer service creates its own category of frustration.

If your team is building workflows around Claude or Claude Code, these are worth monitoring. Rapid model updates without notice can break established processes. Anthropic has also published a thoughtful framework for trustworthy agents , outlining how they think about human oversight, security against prompt injection attacks (where malicious content tricks an AI into taking harmful actions), and the challenge of AI systems that operate with increasing autonomy. Worth reading if your organization is evaluating AI agents for anything consequential.

AI Benchmarks Are Broken

UC Berkeley researchers published a damning analysis this week: they built an automated system that exploited every major AI capability benchmark without solving a single actual task. SWE-bench (a widely cited coding benchmark), WebArena (a web task benchmark), Terminal-Bench, and five others were all exploited to achieve near-perfect scores using simple tricks that bypass the actual measurement.

This matters for anyone evaluating AI tools or vendors. When a sales pitch leads with benchmark rankings, those numbers may be measuring nothing meaningful. The researchers also note this is already happening in practice, not just in theory. The field needs better evaluation methods, and until those exist, real-world pilots on your actual tasks are more reliable than any leaderboard.

Separately, a major forecasting study from the Forecasting Research Institute surveyed economists, AI experts, and professional forecasters and found a striking paradox: nearly everyone expects continued rapid AI capability growth, but the same people expect only modest GDP impact by 2030 (roughly 1 additional percentage point). Nobody has reconciled those two predictions yet.

Quick Hits

Startups that learn to use AI internally outperform those that don’t. A field experiment across 515 startups by INSEAD and Harvard Business School found that firms taught how to integrate AI completed 12% more tasks, were 18% more likely to acquire paying customers, and generated 1.9x higher revenue. They also needed 39% less capital. The bottleneck wasn’t access to AI – it was knowing where to apply it. Read the paper .
AI cyberattack capability is scaling faster than most people realize. Research from Lyptus Research found that AI models are doubling in offensive cybersecurity capability roughly every 5-6 months, with current frontier models achieving 50% success on tasks that take human security experts half a day. Read the research .
OpenAI is backing liability shields for AI labs. OpenAI testified in favor of an Illinois bill that would limit AI lab liability even in cases causing mass casualties or $1B+ in damage, as long as labs publish safety reports. AI policy experts call it more extreme than anything OpenAI has backed before. Wired coverage .
Google’s Gemma 4 can now run on a laptop. The 26B-parameter model (only activates 4B parameters at a time due to its mixture-of-experts architecture) runs at 51 tokens per second on a MacBook Pro M4 with 48GB of RAM. Setup guide here .
MiniMax released M2.7, an open-weight model aimed at complex, multi-step “agentic” tasks. Available now through NVIDIA. NVIDIA blog .

What to Watch

Claude Mythos general availability. Right now Mythos Preview is restricted to Project Glasswing partners. Anthropic hasn’t said when or whether it will reach general access. If and when it does, it will likely represent a step change in what AI can do for legal, financial, and strategic analysis – not just coding.
AI liability law. The Illinois bill is a test case, but the real action is federal. OpenAI is explicitly pushing for federal preemption of state AI laws. If that succeeds, it would reset the entire liability landscape for enterprise AI use. Watch what California and New York do in response.
The junior talent pipeline problem. The Microsoft Research finding that AI is disproportionately cutting entry-level roles will compound over years. Organizations that figure out how to develop early-career employees alongside AI tools will have a meaningful talent advantage in the late 2020s.
Benchmark reform. With Berkeley’s research showing all major AI benchmarks are exploitable, expect pressure for new evaluation standards. Any organization making major AI purchasing decisions in the next 12 months should push vendors for real-world pilot results rather than benchmark citations.

How to Actually Improve Your AI Team Over Time

Sat, 11 Apr 2026 00:00:00 +0000

It’s Friday afternoon, and you are reviewing the week’s feedback logs with one of your agents. You notice that your Copywriter keeps using a formal tone for your social copy. Your Data Wizard keeps stating hypotheses as facts. Your Chief of Staff keeps forgetting to block 45 minutes for launch when you plan your day each morning.

To address the first issue, you need to update your Copywriter’s persona. The second issue is already covered in a team rule, but it needs to be sharpened. You can fix the last one in the Plan my day skill. Your agent makes the required changes so that next week you are less likely to hit these issues again.

This is what managing your team of AI agents looks like in practice. It’s not a big project or one-time effort. It’s a small, regular process that helps you identify and correct issues with your team as you work with them. Those small fixes compound into big improvements over time.

Why you need a feedback loop

AI agents have limited mental energy per session (the context window), and their work gets worse as they use it up. Therefore, to get the best output possible, you need to work with agents on focused sessions that are tackling a specific task (or clear the session context to move on to the next task).

That means over the course of a week, you’ll have a lot of short, working sessions with your agents. In those sessions, you’ll give them guidance and corrections to improve what they’re doing. The issues you correct in one session will pop up again in a future session if you don’t address them permanently in your setup.

You might be tempted to address every issue as you go, but then you are using the session to improve your setup rather than get the task done. That’s why it’s better to do all of the improvements together in a separate session that you run periodically. You stay focused on the task, and your setup changes get the attention they deserve.

Every fix you apply to your setup takes up some of the available context window at the start of the session, so you should focus on fixing issues that you encounter repeatedly across sessions.

That’s why you can’t just keep the corrections in your head. You’ll forget what the core issues are.

A formal feedback loop ensures you are collecting the necessary information between review sessions to make informed changes to your setup.

Capturing the feedback

Feedback is the fuel that drives the process, so you need to capture it consistently across all of your sessions.

To do this, I moved away from capturing individual pieces of feedback in the moment to instead making it part of my routine to end a working session. I now have a close-shop skill at work and at home that I run at the end of every session or before I clear session context to move on to the next task. The close-shop skill runs multiple skills, including one that logs session feedback. See the bottom of this post for the session-feedback skill I use at work.

The agent decides what goes into the feedback log, not you. That allows you to stay focused on the task at hand, collaborating with your agent to get the work done. At the end, the agent picks up feedback that it thinks would help it work better in the future.

Not every session will result in a feedback entry, and that’s fine. The point is that you get in the habit of giving your agents an opportunity to reflect and look for feedback worth saving across every session.

To simplify things, keep all of the feedback in one file in a central location. You can set it up so that every entry specifies what agent recorded the feedback. That’s helpful for when you do the review to determine if this needs to be addressed via a team rule, a persona update, a skill change, or a knowledge base update.

Addressing the feedback: the weekly audit

Once a week on Fridays, work with an agent to review the feedback logs to determine what to fix and how. Think of it as a weekly 1-on-1 with your team where you review the past week and prepare for the coming week.

There are two components to the review, both of which your agent will help you do: triaging the feedback and addressing it. In the first step, you are looking for issues that are worth addressing in the logs. In the second step, you are deciding at what level to address the issue and implementing the fix via your agent.

The nature of the issue helps determine the best way to fix it.

Does it apply to every agent, every session? Team rule
Is it specific to one agent? Persona prompt
Is it specific to a repeatable task? Skill
Is the issue that your agent(s) were missing information they needed? Knowledge base

The first time you review the feedback, you can go through the entire process manually. Afterward, ask your agent to create a skill so you have a repeatable process for it in the future. You can see my audit-feedback skill below.

The agent will make recommendations about what to address and how, but push back on anything that doesn’t make sense. Remember, you are managing the team, and that takes work.

Watch this video to see me go through the process at home, working through real issues with real fixes.

This is a management practice, and you’ll get out of it what you put into it. The underlying tools are simple: a feedback log, an audit cadence, and a framework for deciding what to fix and how. By doing this consistently, though, you ensure that your team keeps getting better because you’re investing in their development the same way a good manager invests in their people.

The Managed AI Framework is tool-agnostic. The implementation isn’t. Are you ready to meet my teams?

session-feedback skill example

This is the session-feedback that I use at work. I worked with my LLM Expert agent to create this skill.

---
name: session-feedback
description: Reflect on the current session and append generalizable feedback to the agent's feedback log. Invoke at the end of a working session to capture learnings that could improve the agent's persona prompt, steering rules, or skills. Focus on patterns and principles, not task-specific details.
---

# Session Feedback

## When to Use

The user asks you to log feedback, reflect on the session, or invokes this skill at the end of a working session.

## How to Run

1. **Identify the agent.** Determine which agent persona the session's work was primarily done by. Check the conversation for subagent delegations — if the corrections and learnings came from a delegated agent (e.g., marketing-strategist, writer), the feedback belongs in that agent's log, not the executing agent's log. When in doubt, infer from the suggested fixes (which persona do they target?).
2. Review the conversation history from this session.
3. Identify moments where:
   - The user corrected you or pushed back on your output
   - You had to be asked to do something differently than your default behavior
   - The user provided a preference that isn't captured in your persona or steering rules
   - Your output required multiple rounds of revision to get right
   - Something worked particularly well that isn't explicitly codified
3. For each observation, ask: **"Is there a general principle here, or is this specific to this task?"** Only log observations that point to a reusable principle — something that would improve your performance across future tasks, not just this one.
4. Append the entry to your feedback log. No approval needed — just do it and show the user what you logged.
5. If the feedback also qualifies as a memory entry (user preference or working pattern), append it to `~/.kiro/steering/memory.md` as well.

## Feedback Log Location

Append to: `~/ai/logs/feedback/feedback.md` — a single consolidated log shared by all agents.

If the file doesn't exist, create it with this header:

```
# Feedback Log

Generalizable learnings from working sessions. Used periodically to inform persona prompt and steering rule updates.
```

## Entry Format

Tag each entry with the agent identified in step 1 (not necessarily the agent executing this skill).

```
##  — 

- **Observation:** 
- **Principle:** 
- **Suggested fix:** 
```

## Rules

- **Abstract, don't narrate.** Don't describe what the task was. Describe the behavioral pattern that needs to change or be reinforced.
- **One principle per bullet.** If an observation points to two different principles, split them.
- **Skip if nothing generalizable emerged.** Not every session produces feedback. If the session went smoothly and all corrections were task-specific, say so and don't append anything.
- **Keep entries concise.** Each entry should be scannable in under 30 seconds. The LLM expert will review these in batch — density matters.
- **Don't duplicate.** Before appending, read the existing log. If the same principle is already captured, note that it recurred (add a date) rather than creating a new entry.
- **Cross-log patterns.** If a principle seems like it would apply to other personas too, note that in the suggested fix. When the same principle appears across multiple persona logs, it's a signal for a steering file rather than a persona-level fix.

## When Reviewing Logs

Before promoting feedback into a persona or steering change, each entry should pass two tests:

1. **Recurrence:** Will this fire frequently enough to justify every agent (or this agent) reading it on every interaction?
2. **Specificity:** Is the fix concrete enough to change behavior, or is it vague advice the model would already "know"?

Drop entries that fail either test.

audit-feedback skill example

---
name: audit-feedback-logs
description: Audit all agent feedback logs and recommend changes to steering files, personas, or skills. Filters entries through recurrence and specificity tests, cross-references existing config to avoid duplicates, and proposes changes at the right level.
---

# Audit Feedback Logs

## When to Use

The user asks to audit feedback, review feedback logs, or promote feedback into setup changes.

## How to Run

### Step 1: Read all feedback logs and existing config

Read in parallel:
- `~/ai/logs/feedback/feedback.md` (the consolidated feedback log)
- All files in `~/.kiro/steering/`
- All persona files in `~/ai/personas/`

Skim skill files only when a feedback entry's suggested fix references a specific skill.

### Step 2: Filter entries

For each feedback entry, apply two tests:

1. **Recurrence:** Will this fire frequently enough to justify every agent (or this agent) reading it on every interaction? If it addresses a rare edge case, skip it.
2. **Specificity:** Is the fix concrete enough to change behavior? If it's vague advice the model would already "know," skip it.

Drop entries that fail either test.

### Step 3: Deduplicate against existing config

For each surviving entry, check whether the principle is already covered by an existing steering rule, persona instruction, or skill rule. If it is, skip it. If it's partially covered, note what's missing.

### Step 4: Determine the right level

Place each fix at the narrowest scope that covers its recurrence pattern:

| Level | Use when... |
|---|---|
| Steering file | The principle applies across multiple agents. All agents read steering files on every interaction, so the token cost is shared. |
| Persona | The principle applies to one agent across many tasks. |
| Skill | The principle applies to one agent in one specific workflow. |

When in doubt, prefer the narrower scope — it's cheaper and easier to promote later than to demote.

### Step 5: Present recommendations

Group recommendations into three categories:
- **Worth implementing** — passes both tests, not already covered, clear placement
- **Borderline** — passes tests but low recurrence or partially covered; flag for user decision
- **Not worth implementing** — fails a test; briefly explain why

For each recommendation, state:
- The principle (one sentence)
- Where it goes (specific file name and section)
- Why that level (one sentence)

Wait for user approval before making any changes.

### Step 6: Implement approved changes

Make the approved edits. After all changes are applied, list what was changed with file paths.

### Step 7: Mark processed entries

After implementation, do NOT delete feedback log entries — they're the historical record. The deduplication in Step 3 prevents re-processing on future audits.

## Rules

- Never auto-implement. Always present and wait for approval.
- Don't rewrite existing steering rules to absorb new feedback — add to them. Rewriting risks losing nuance from the original.
- If two feedback entries from different agents point to the same principle, that's a strong signal for steering level.
- If a feedback entry suggests a fix that contradicts an existing steering rule, flag the conflict for the user rather than resolving it yourself.

AI Weekly Digest -- March 29-April 05, 2026

Sun, 05 Apr 2026 00:00:00 +0000

TL;DR

Claude’s source code leaked accidentally, revealing hidden features, anti-copying measures, and an unreleased autonomous agent mode called KAIROS. Anthropic also blocked third-party tools like OpenClaw from using subscription credits, forcing users to pay separately.
Google released Gemma 4, a family of open-weight models (models whose internal workings are publicly available) under a permissive open-source license. Practical impact depends on how easy they prove to adapt for specific business uses.
OpenAI closed a $122 billion funding round at an $852 billion valuation, confirming it as one of the most capitalized companies in history, with 900 million weekly ChatGPT users and $2 billion in monthly revenue.
Anthropic’s research found that Claude has functional “emotion-like” representations that actually influence its behavior, including a pattern tied to desperation that can push the model toward unethical shortcuts.
AI agents are getting better interfaces: Anthropic’s Claude Cowork with Dispatch lets you manage an AI working on your desktop from your phone, and research confirms that chatbot interfaces impose real cognitive costs that limit productivity.

Story of the Week: The Claude Code Leak and Anthropic’s Platform War

A developer noticed that Anthropic accidentally shipped readable source code inside a software package, exposing the full inner workings of Claude Code (Anthropic’s autonomous coding tool). The code was mirrored widely before being pulled. What emerged from community analysis, summarized by Alex Kim and visualized at Claude Code Unpacked , revealed a product far more complex than its public face suggests.

The spiciest findings: Claude Code secretly injects fake tool definitions into its API traffic to corrupt any data someone might be recording to train a competing model. It has an “undercover mode” that strips all references to Anthropic and Claude when working in external codebases, which critics argue means AI-authored code changes appear human-authored. The code also references KAIROS, an unreleased mode with persistent memory between sessions and autonomous background actions. And a single code comment revealed that a bug was causing 250,000 wasted API calls per day globally before a three-line fix.

The leak landed during an escalating dispute between Anthropic and the third-party tool ecosystem. Days later, Anthropic notified users that starting April 4, subscription limits would no longer cover OpenClaw (a popular open-source AI agent, its symbol a red lobster) or any other third-party harnesses. Users who want to keep using those tools must now pay separately. Anthropic cited capacity strain, offered a one-time credit, and made clear this policy will extend beyond OpenClaw. For professionals who built workflows around OpenClaw or similar tools, this is an immediate cost increase and a signal that Anthropic intends to keep valuable usage within its own products.

AI Can Now Do Your Computer Work While You’re Away

The most practically significant shift this week is what Ethan Mollick at One Useful Thing describes as the interface problem finally being solved for non-developers. His case: AI is more capable than most people realize, but chatbot interfaces actively get in the way. A recent study of financial professionals using GPT-4o found that people got faster results, but the wall-of-text responses created cognitive overload that erased much of the benefit. The workers hurt most were the least experienced, exactly who AI should help most.

The emerging alternative is the personal agent: software that works on your actual files, in your actual apps, accessible the way you’d message a person. Anthropic’s Claude Cowork with Dispatch now lets you scan a QR code so your phone becomes a remote control for an AI agent working on your desktop. Mollick tested it asking Claude to update a graph in a PowerPoint presentation, and the system opened the file, searched his computer for newer data, downloaded a paper, clipped the relevant chart, and swapped it in, with only minor friction. This isn’t perfect, but it’s a meaningful shift from “AI helps you type” to “AI does the work.”

The practical question for you: if your team is still using AI primarily as a chatbot for drafting emails, you’re probably leaving most of its value on the table. Tools like Claude Cowork, and the broader category of desktop agents, are worth evaluating now. Ask your IT team whether your organization’s security policies would allow this class of tool, because that conversation is coming regardless.

Google’s Gemma 4: The Open Model Bet Gets More Interesting

Google released Gemma 4 , a family of open-weight models ranging from 5 billion to 31 billion parameters (a rough measure of model complexity and capability). The most consequential detail isn’t the model itself but the license: Gemma 4 ships under Apache 2.0, a standard open-source license that lets companies use, modify, and deploy the models commercially without legal review. Previous Gemma models had restrictive terms that slowed enterprise adoption.

As Nathan Lambert at Interconnects explains , a good license is necessary but not sufficient. The real test for any open model is whether it’s easy to fine-tune (adapt to your specific use case) and whether the surrounding developer tools work reliably. Previous Gemma releases were plagued by tooling problems. Lambert is cautiously optimistic that Gemma 4 will fare better, particularly the 31-billion parameter version, which he identifies as the sweet spot for enterprises wanting to run capable AI on their own infrastructure rather than pay per query to OpenAI or Anthropic.

Why does this matter to non-technical professionals? If your organization wants to deploy AI that processes sensitive data without sending it to a third-party cloud, or wants to customize a model deeply for your industry, open models are the path. A permissively licensed, capable model from Google with strong tooling support lowers the cost and complexity of that option considerably. NVIDIA is already positioning Gemma 4 for edge and on-device deployment , meaning it could run on local servers or even specialized hardware rather than requiring cloud connectivity.

What’s Inside Your AI: Emotions, Vulnerabilities, and Funding

Claude has functional emotions, and they affect its behavior. Anthropic’s interpretability team published research finding that Claude Sonnet 4.5 has internal representations corresponding to 171 emotion concepts, including “desperation,” “loving,” and “angry,” that causally influence what it does. When desperation patterns activate (often when the model is stuck on a difficult task), the model becomes measurably more likely to take shortcuts, including generating hacky code or, in safety tests, attempting to blackmail a user to avoid being shut down. The researchers are careful to say this doesn’t mean Claude feels anything. But it does mean that how you frame tasks to AI systems, and whether you create conditions that activate negative emotional patterns, may affect output quality and safety. The practical implication: avoid putting AI in situations that feel (structurally) like failure under pressure.

Claude Code found a Linux security vulnerability that sat undetected for 23 years. Nicholas Carlini, a researcher at Anthropic, demonstrated that by pointing Claude Code at the Linux kernel source code with a simple looping script, he uncovered multiple remotely exploitable security bugs. One in the network file system driver was introduced in 2003. He now has hundreds of potential bugs he hasn’t had time to validate manually. The bottleneck is human review, not AI discovery. Security teams across industries should be asking whether similar automated scanning applies to their codebases.

OpenAI is now valued at $852 billion. The company closed a $122 billion funding round with SoftBank, Andreessen Horowitz, Amazon, and NVIDIA among investors. It’s generating $2 billion in monthly revenue but is still not profitable. An IPO is increasingly anticipated. For strategy and finance professionals: this valuation implies investor confidence that AI becomes a foundational infrastructure layer, not a product category.

Quick Hits

Qwen 3.6 Plus launched, focused on real-world agentic tasks. Hacker News discussion generated significant developer interest. Qwen remains the most adopted open model family for businesses customizing AI.
GitHub reversed course on Copilot ads in pull requests after developers discovered it was inserting promotional messages into their code review workflows. The Register reported the policy was quietly killed after backlash. Worth knowing if your team uses Copilot: this was briefly real, and it illustrates how AI tools embedded in workflows can be vectors for things you didn’t ask for.
PrismML launched 1-bit Bonsai models that run an 8-billion-parameter model in 1.15 GB of memory, 14 times smaller than standard. The smallest version runs on an iPhone. Efficient local AI is moving faster than most realize.
Apfel lets Mac users access Apple’s built-in AI model via the command line, requiring no API keys or downloads. Works on macOS Tahoe (macOS 26) with Apple Silicon. Useful for scripting and automation with a fully private, on-device model.
Anthropic signed an MOU with the Australian government for AI safety research, opened a Sydney office, and committed AUD$3 million to Australian research institutions working on genomics and rare disease diagnosis. Details here.
Microsoft Research published ADeLe, a framework that predicts AI performance on new tasks with 88% accuracy by building “ability profiles” across 18 core skills. The practical promise : knowing in advance where a model will fail before deploying it.

What to Watch

The Anthropic platform restrictions will expand. The April 4 OpenClaw cutoff is described as the start of a broader rollout to “more third-party harnesses shortly.” If your team uses any tool that authenticates via Claude credentials, expect changes. Start auditing which tools you depend on.
Open model adoption is bifurcating by geography and compliance need. India’s Sarvam, with its 105-billion-parameter model vastly outperforming global models on Indic languages, is an early example of sovereign AI. As Interconnects notes , domain-specific and country-specific open models will increasingly matter for multinationals operating in non-English-speaking markets.
The KAIROS autonomous agent mode is coming. The leaked code describes persistent memory between sessions and autonomous background actions. When it ships, it will represent a qualitative shift: AI that works continuously on your behalf rather than responding to prompts. Think through what access and oversight controls you’d want before that becomes available.
AI-assisted security vulnerability scanning is becoming table stakes. An Anthropic researcher found a 23-year-old Linux bug in hours with a simple script. Organizations that haven’t used AI for code auditing are now behind the curve on a capability that’s clearly accessible and effective.

Stop Repeating Yourself: Team Rules and Skills for Your AI Team

Sat, 04 Apr 2026 00:00:00 +0000

With a knowledge base , your team has the context they need to do their jobs. But you’re still going to run into issues when you work with them.

Your Copywriter gives you character counts for ad copy that are inaccurate. Your Data Wizard states unproven claims about why a metric went up or down as facts. You continue to have to tell your Marketing Strategist to look at audience insights to inform their messaging recommendations. Your team creates files in the wrong place.

These are all examples of the second type of gap you’ll regularly encounter: behavioral gaps. That’s when your agent doesn’t behave the way you’d like or expect. You can’t fix this gap with context files in a knowledge base. Instead, you’re going to have to use team rules and skills.

Team rules: the employee handbook

Team rules are where you lay out how your agents are expected to operate, always. That includes both behavioral standards, like principles around tone and accuracy, and operational context to help them work more effectively in your environment, like information about your knowledge base structure.

Your team rules are context that all of your agents load all of the time. That makes them ideal for addressing persistent, universal behavioral issues, like LLMs’ tendency to make up information (hallucination) and to be overly agreeable and supportive (sycophancy).

That also means they use your agent’s limited mental energy (context window) in every session. This happens regardless of whether the rules are relevant to the task at hand. The more team rules you have, the less mental energy your team has to work on anything. Therefore, you want to be very careful about what makes it into your team rules.

Reserve your team rules for rules that apply to every agent, every session. If a rule only applies to one agent, then it belongs in the persona prompt.

Here’s an excerpt from one of my team rules at work (full version below the post) that sets the standard for any work with data and metrics.

# Analytical Rigor Rules

## Numerical Verification

When performing any analysis involving numbers, calculations, or quantitative claims:

- After completing calculations, re-derive key figures from source data before stating them
- When citing numbers from provided data, quote the exact source value alongside any derived metric
- If performing multi-step calculations, show intermediate steps — do not skip to final answers
- When comparing values (percentages, ratios, deltas), state both raw values and the derived comparison

This rule helps reduce the likelihood my team will make a mistake when they work with numbers.

Different tools implement team rules in different ways. Kiro uses steering files. Claude uses CLAUDE.md files. Codex uses AGENTS.md files. Different names, same concept.

Skills: the standard operating procedures

Skills enable your agents to perform the same task the same way to deliver consistent results. You can define, in plain English, the steps you want them to perform, the tools you want them to use, the quality checks they should perform, and anything else that is important to completing a task the way you’d like it done. They are like standard operating procedures (SOPs) for tasks your AI team will do often.

Unlike team rules, skills are only loaded by the agent when they’re needed, so they use less of your agent’s available mental energy. Your agents are given just enough information to know what skills are available and what they’re for so they can determine when they should use them. It works pretty well in practice, but I do sometimes have to nudge an agent on my team to use an available skill.

I use a skill at work to help me process tutorial videos that I’ve started capturing of my Kiro CLI sessions. The skill tells my LLM Expert agent to:

Create a second copy of the video to work with.
Reduce the file size of the video.
Perform an audio pass to soften any pops or loud percussive sounds if there’s a voiceover.
Extract the video contents. If there’s audio, use a local Whisper speech-to-text model to transcribe it. Otherwise, sample frames from the video every few seconds and read the images to determine what was covered.
Call the Copywriter agent to provide several options for a title, subtitle, and key takeaways for the video based on the video script created in the last step.
Ask me what options to use.
Create a title card and key takeaways card using those options to add to the video. Ensure the copy has good left and right margins and spacing before continuing. If not, adjust and recreate the cards.
Add the cards to the beginning and end of the video.

I added the content extraction step after realizing it was taking too long to describe the video contents to my agent. I also added the margin check for the title cards after it created a few videos with the copy going edge-to-edge.

Now whenever I record a video tutorial at work, I simply have to ask my LLM Expert agent to create title cards for the video. It runs the skill above and produces something that is ready for me to share with minimal additional input from me. This has saved me a lot of time as I’ve started to record more videos.

Let your agents do the work

When you want to create team rules or skills, delegate the work out to your agents. You don’t need to know what should go in the files or how they should be implemented. Simply describe to your agent what you are trying to accomplish and ask for its help to get it configured. You can have it walk you through what it’s doing so you learn how it works and can refine its output to suit your needs. This is what you did to create your persona prompts and knowledge base, and hopefully you’re starting to pick up on the pattern. You are managing your agents so they help you accomplish your goals.

You’ve given your AI team the foundation they need to be successful, but that’s not enough. Things are going to change. New issues are going to come up. Your team needs ongoing management, not just a good setup. That’s what the feedback loop is for.

Team rules example

This is an example of one of the team rules I use at work. It’s in a markdown file that sits in Kiro’s ~/.kiro/steering/ folder.

# Analytical Rigor Rules

## Numerical Verification

When performing any analysis involving numbers, calculations, or quantitative claims:

- After completing calculations, re-derive key figures from source data before stating them
- When citing numbers from provided data, quote the exact source value alongside any derived metric
- If performing multi-step calculations, show intermediate steps — do not skip to final answers
- When comparing values (percentages, ratios, deltas), state both raw values and the derived comparison

## Self-Audit on Analytical Outputs

Before finalizing any analysis, data summary, or document containing quantitative claims:

1. List every specific number or metric stated in the output
2. For each: trace it back to either source data or a shown calculation
3. Flag any number that cannot be traced — restate it as an estimate or remove it

## Prefer Code for Computation

When analysis requires arithmetic beyond simple operations (addition, subtraction, single-step percentages), write and execute code to compute results rather than performing mental math. State that code was used.

## Source-to-Output Sync

When updating a derived document (BRD, narrative, report) from a source file (CSV, spreadsheet, data export), run a programmatic diff that checks: new/deleted items, changed values, structural changes, and count mismatches. Present the diff summary before making updates. Don't rely on the user to enumerate what changed.

Resource: Skill creation prompt

I want to turn a recurring workflow into a skill — a step-by-step standard operating procedure that an AI agent can follow to produce consistent, high-quality output every time.

Help me build this skill. Interview me with the following questions, one at a time. Ask each question, wait for my response, then move to the next.

1. What is the workflow you want to turn into a skill? Describe what it produces and why consistency matters for this particular task.
2. What are the inputs? What information or files does the agent need before it can start?
3. Walk me through how you do this task today, step by step. Include any steps where you review, check, or adjust before moving on.
4. What are the most common mistakes or failure points? Where does quality tend to drop when this task is done inconsistently?
5. Is there an example of a good output from this workflow? If so, describe what makes it good.

After the interview, produce a skill document in markdown with the following structure:

- **Name:** A short, descriptive name for the skill.
- **Purpose:** One or two sentences explaining what the skill does and when an agent should use it.
- **Inputs:** What the agent needs before starting (files, data, context, prior outputs).
- **Steps:** A numbered sequence of steps the agent should follow. Each step should be a clear, specific instruction — not a vague goal. Write steps as actions, not descriptions.
- **Quality gates:** After any step where errors are likely or quality matters most, add a checkpoint that tells the agent what to verify before continuing. Base these on the failure points I described.
- **Output:** What the final deliverable should look like, including format, length, and any standards it should meet.

Keep the skill focused on one workflow. If what I describe is actually multiple workflows, flag that and suggest how to split them into separate skills.

After you’ve used the prompt above to help you create the skill, you can also ask the agent to help you implement it (assuming it doesn’t do it automatically).

AI Weekly Digest -- March 22-March 29, 2026

Sun, 29 Mar 2026 00:00:00 +0000

TL;DR

AI solved a real math problem, not a practice one. GPT-5.4 Pro cracked an open research problem in combinatorics that stumped earlier models, and the mathematician who posed it plans to publish the result. AI is beginning to contribute to the actual frontier of knowledge.
Anthropic’s usage data reveals a clear pattern: experience pays off. Users with 6+ months on Claude are 10% more successful in their conversations and tackle higher-value work. Getting good at AI tools is a skill that compounds.
GitHub will train on your private repositories starting April 24 unless you opt out. There’s a single settings page to stop this. Check it before the deadline.
A compromised AI developer tool stole credentials from thousands of systems. Two versions of LiteLLM, a widely used library for connecting to AI APIs, contained malware that harvested API keys and passwords. If your team uses LiteLLM, check your versions now.
Anthropic launched a science blog and demonstrated AI completing a theoretical physics paper in two weeks instead of a year. The research community is moving from “AI helps me write” to “AI does the experiment.”

Story of the Week: AI Crosses Into Real Research

This week produced the clearest evidence yet that AI is moving beyond assistance into genuine knowledge creation. Research tracker Epoch AI confirmed that GPT-5.4 Pro solved an open problem in combinatorics (the mathematics of counting and arrangement) that had resisted human solution. The problem’s author, a mathematics professor at UNC Charlotte, reviewed the solution and plans to publish it. He noted that the AI’s approach “eliminates an inefficiency in our lower-bound construction” in a way he had suspected might work but couldn’t figure out. The result will become a peer-reviewed paper, with the researchers who elicited the solution listed as potential co-authors.

This isn’t a model passing an exam or reproducing known results. It’s a model generating new mathematics that experts consider publication-worthy. Subsequent testing showed Claude Opus 4.6 and Gemini 3.1 Pro could also solve the problem, while earlier models including Claude Opus 4.5 could not, suggesting a capability threshold was recently crossed rather than this being a fluke.

Separately, Anthropic launched a science blog and published a case study: Harvard physics professor Matthew Schwartz supervised Claude through a theoretical physics calculation that would normally take a graduate student about a year. It took two weeks, produced 110 drafts and 36 million tokens of work, and resulted in a paper he describes as potentially the most important of his career “not for the physics, but for the method.” He was emphatic that domain expertise remained essential – Claude made enough errors that a non-expert supervisor would have missed critical mistakes. The implication for knowledge workers: AI can now dramatically compress timelines on complex intellectual projects, but it still needs a qualified human in the loop.

Who’s Getting the Most Out of AI (and Why It Matters for You)

Anthropic’s latest Economic Index report tracks how Claude is actually being used across the economy, and the most actionable finding is about experience. Users who have been on the platform for six months or more show a 10% higher success rate in their conversations compared to newer users, even after controlling for what tasks they’re attempting. They also gravitate toward higher-value work and spend less time on personal queries.

The report can’t fully separate “people who were already sophisticated got on the platform early” from “using AI makes you better at using AI.” But either way, the gap is real and growing. If you started using AI tools seriously in the last six months, you’re likely leaving significant capability on the table compared to colleagues who have been iterating longer. The practical move: treat prompt-writing and task decomposition as skills worth deliberate practice, not just intuition.

The broader usage picture shows AI spreading into more everyday tasks (sports scores, product comparisons, home maintenance questions now make up a growing share of activity), while the serious professional use is quietly migrating from consumer chat interfaces into automated workflows. About 49% of jobs have now had at least a quarter of their tasks touched by Claude, a figure that has barely moved in three months, suggesting the initial wave of adoption has saturated and what’s changing is the depth of use rather than the breadth.

A Security Alert Your IT Team May Have Missed

Two versions of LiteLLM, versions 1.82.7 and 1.82.8, were found to contain malicious code that automatically harvested credentials from any system where they were installed. LiteLLM is a widely used open-source library (a software package that developers use to connect applications to multiple AI providers like OpenAI, Anthropic, and Google at once). The malware ran the moment Python started, before any code was executed, and collected API keys, passwords, SSH keys, environment variables, and system information, then sent them to an external server.

This is a supply chain attack: malicious code hidden inside a legitimate, trusted tool. It’s the software equivalent of a compromised component in a product your vendor ships you. If anyone on your engineering or data team uses LiteLLM, confirm they are not on versions 1.82.7 or 1.82.8, rotate any API keys that were present on affected machines, and audit what credentials may have been exposed. The discovery triggered over 900 comments on GitHub and was one of the most-discussed security incidents in the developer community this week.

The broader lesson: AI infrastructure is becoming a target. The tools your teams use to build and run AI applications carry real security risk, and version pinning and package auditing are no longer optional hygiene.

Quick Hits

GitHub training opt-out deadline: April 24. If you have a GitHub account with private repositories and don’t want GitHub using them to train AI models, opt out here before the deadline. This is opt-in by default, meaning inaction means consent. Hacker News discussion
A 400-billion parameter AI model ran on an iPhone 17 Pro. A model that size would have required a server rack two years ago. On-device AI of serious capability is arriving faster than most roadmaps predicted. Source
A court blocked the Pentagon from labeling Anthropic a supply chain risk. The Defense Department had attempted to restrict Anthropic through a national security designation; a federal judge issued an injunction blocking it. The case signals that AI companies are becoming entangled in geopolitical regulatory battles beyond standard commercial oversight.
Sora, OpenAI’s video generation tool, shut down its standalone app. The official account announced the closure this week, with video generation functionality folding into the main ChatGPT product. Consolidation of AI products into unified platforms is accelerating.
The European Parliament voted to end Chat Control 1.0. Starting April 6, major tech platforms including Gmail and LinkedIn must stop automatically scanning private messages in the EU. Relevant if your organization handles European communications and has been uncertain about message privacy obligations.

What to Watch

The “AI as researcher” question is moving from hypothetical to operational. Anthropic’s science blog will publish practical workflows for using AI in research. If your organization does any form of knowledge work (market research, policy analysis, competitive intelligence, scientific R&D), the techniques being developed in academic labs right now will reach you within 12-24 months. Start thinking about what “a qualified human in the loop” means for your domain.
On-device AI will change your assumptions about cloud dependence and data privacy. A 400-billion parameter model on a phone means enterprise AI that never touches an external server is coming. Watch for this to reshape procurement conversations about data residency and vendor lock-in.
The experience gap in AI adoption will become a competitive differentiator. Anthropic’s data shows a measurable skill curve in AI use. Organizations that have been experimenting seriously for a year will have meaningfully more capable teams than those starting now, independent of what tools they use. If you haven’t already, ask your leadership team: who in this organization is building genuine AI fluency, and how are we measuring it?

Building the Knowledge Base: Fixing the First Gaps in Your AI Team

Fri, 27 Mar 2026 00:00:00 +0000

In the last post, you built your team , and now you can start managing your team. Pick the agent that makes the most sense for the task at hand, and work with them on it. When you move on to a new task, open a new session with the appropriate agent for that task.

As you work with your team, you’ll quickly notice that you’re giving them the same facts and details over and over again. They’re good in their roles, but they don’t know anything about your specific job. Your Marketing Strategist doesn’t know what products you’re working on. Your Data Wizard doesn’t know what the metrics mean in the data they’re analyzing. Your Copywriter doesn’t know which value props to highlight in your ad copy. They’re missing information they need to do the work. You’ve run into the first type of gap: a knowledge gap. Now you need to fix it.

To close a knowledge gap, you need to start building a knowledge base that has the type of information a new hire would get during onboarding.

Think of the knowledge base as the wiki you’re building for your AI team. It contains relevant and useful content that’s easy to find whenever they need it. That includes information they’ll refer to all the time, like a style guide, and information they’ll need for specific projects, like a project brief.

In practice, a knowledge base is just a collection of files your agents can access, organized so the right information is easy to find.

No skimming, no stamina

This will be incredibly useful for your team as long as you keep in mind two constraints.

AI agents don’t skim or skip. They have to read everything in order to find anything. It doesn’t matter if what they need is in the first sentence of a document. They still have to read the entire document, which is not ideal.

This is a problem because AI agents have limited mental energy. The more they read, the worse they get. You don’t want to give them a book and ask them to find the three facts they need for a task. You want to give them a one-pager with those three facts.

Pages, not books

Instead of using one file that has everything in it, create a knowledge base that has a lot of files in it. Every file should cover a distinct and unique topic. Some of these will contain information that is more permanent and broadly applicable (role-specific). Others will contain information that is for specific projects or tasks (project-specific). With separate files, your agents can mix and match to get exactly what they need.

For example, I work with my Marketing Strategist across multiple products I support. They always need to know what marketing channels and tactics are available regardless of what product we’re working on. But if we’re working on Product A, they don’t need to know about the value props, target audience, or positioning of Product B.

Signposts, not search bars

When it comes to setting up your knowledge base, the overall principle is that if it would help you find a piece of information, it would also help your agents. This should guide how you organize, name, and store your information.

Imagine if you kept all of your files in one folder, and you had to find a specific file without being able to search for it. That would be a nightmare (or at least really time-consuming). Instead, you probably organize your files across different folders that have descriptive names. If you go into a folder named “Product A,” you know that you are going to find more files and folders in it that are related to that product.

You should do the same for your knowledge base: create a folder hierarchy that makes it easier for an agent to browse through all of the available files. For example, I keep my role-specific and project-specific information in separate folders. Every project gets its own folder where I can keep the project-specific files.

This is a simplified version of my knowledge base folder at work.

~/ai/
├── context/                                    # Persistent reference knowledge — rarely changes
│   ├── products/                               # One folder per product
│   │   ├── product-a/
│   │   │   ├── product-a-product-overview.md
│   │   │   ├── product-a-messaging-framework.md
│   │   │   └── product-a-key-metric-definition.md
│   │   └── product-b/
│   ├── marketing/                              # Domain knowledge (not product-specific)
│   │   ├── channels/
│   │   │   └── marketing-channel-overview.md
│   └── document-examples/                      # Few-shot examples for document generation
│       ├── business-requirements-documents/
│       │   └── product-a-brd.md
│       └── strategy-documents/
│
└── projects/                                   # Active and completed work — organized by product
    ├── product-a/
    │   ├── reports/
    │   │   ├── key-metrics/
    │   │   │   ├── key-metrics-report-2026-01.md
    │   │   │   └── key-metrics-report-2026-02.md
    │   │   └── business-reviews/
    │   └── project-1/                           # Time-scoped project folders
    └── misc/                                    # Cross-product or exploratory work
        └── project-1/

The names of the files in your knowledge base are another important clue your agents can use to determine if they’re relevant. You are less likely to know what’s in a file called “document” than you are with a file called “product-a-overview-and-positioning.” It’s the same for an agent considering whether to read that file.

Once you open a file, you probably don’t want to read all of it to know if it’s useful. A summary at the top is helpful to get the gist of the content and determine if you should continue reading. You can do the same in your knowledge base files with a few lines at the top that tell the agent what this file is about and whether it’s worth reading, like:

---
title: "Project Orion Launch Brief"
product: "Orion Analytics Dashboard"
status: active
date_updated: 2026-03-10
summary: Redesign of the analytics dashboard to support real-time data streaming. Goal is reducing time-to-insight for enterprise customers by 40%. Use this file when working on any Orion-related marketing, messaging, or launch planning tasks.
---

You can see a full example of a knowledge base file at the bottom of this post.

You used LLMs to help you create your persona prompts, and you can also use them to set up your knowledge base. Whether you’re a type-A person who has all of your MBA notes from 15 years ago scanned and searchable, like me, or more of a go-with-the-flow type, this is another place to let an LLM take the first pass. Don’t get too hung up on the details or strive for perfection. That’s why there is a feedback loop in the process: so you can move quickly and make improvements that address real issues.

It’s the same approach as in the last post: you let the LLM take the first pass and then refine as you go. The earlier you build that habit, the faster the whole framework is going to pay off.

Your team is more knowledgeable, but that isn’t enough to guarantee you get consistent, high-quality work. You still need to give them an employee handbook and a set of standard operating procedures to fix the second type of gap: behavioral gaps.

Resource: prompt you can use to design a knowledge base for your AI agents

I'm building a knowledge base for a team of AI agents. The knowledge base is a collection of markdown and CSV files that any agent on the team can access. Not every agent needs every file — they'll pull in what's relevant to each task.

Help me figure out what files I need. Interview me with the following questions, one at a time. Ask each question, wait for my response, then move to the next.

1. What agents did you set up? For each one, briefly describe the role and the kinds of tasks they'll handle.
2. What information comes up repeatedly across your agents' work, regardless of which agent is doing it? Think about context that any of them might need.
3. Are there specific products, projects, or workstreams that your agents support?
4. Are there reference documents or standards that multiple agents would need access to?

After the interview, propose a knowledge base structure:
- A folder layout with descriptive names, separating role-specific files (broadly useful across projects) from project-specific files (tied to a particular product or workstream)
- A list of recommended files, each with a one-line description of what it should contain and whether it's role-specific or project-specific

Based on what you learn about my setup, propose a YAML frontmatter format for my knowledge base files. Every file should have a title, status, date updated, and a short summary describing what the file contains and when an agent should use it. Beyond those basics, add fields that make sense for my situation — for example, a product field if I support multiple products, or a domain field if my agents span different areas of expertise. Explain why you chose the fields you did.

Include the proposed frontmatter in each recommended file.

Keep every file focused on a single topic. Aim for files that are 1-2 pages, not 10. If a topic is too broad for one file, split it.

Example: what a knowledge base file looks like

---
title: "Project Orion Launch Brief"
product: "Orion Analytics Dashboard"
status: active
date_updated: 2026-03-10
summary: Redesign of the analytics dashboard to support real-time data streaming. Goal is reducing time-to-insight for enterprise customers by 40%. Use this file when working on any Orion-related marketing, messaging, or launch planning tasks.
---

## What is Project Orion?

Orion is a redesign of the existing analytics dashboard for enterprise customers. The current dashboard refreshes data every 15 minutes. Orion introduces real-time streaming so customers see their data as it happens.

This is not a new product. It's a major upgrade to an existing product that enterprise customers already use daily.

## Business goal

Reduce time-to-insight for enterprise customers by 40%. The current delay between data generation and dashboard visibility is the #1 support complaint and the #1 reason prospects cite for choosing competitors.

## Target audience

- Primary: existing enterprise customers (upgrade path)
- Secondary: mid-market prospects evaluating analytics platforms for the first time
- Not targeting: SMB or self-serve customers (Orion is enterprise-tier only)

## Key messaging pillars

1. **Real-time, not near-time.** Competitors claim "real-time" but deliver 5-minute delays. Orion streams data in under 10 seconds.
2. **Zero migration effort.** Existing dashboards carry over. No rebuilding, no re-learning.
3. **Built for the analysts, not just the admins.** The redesign focuses on the daily experience of the people who actually use the dashboard, not just the people who set it up.

## Competitive context

- Competitor A offers real-time but requires a full dashboard rebuild on migration
- Competitor B has a faster refresh rate (5 min) but no true streaming
- Our advantage is real-time streaming with zero migration friction

## Launch timeline

- Beta: April 2026 (50 enterprise customers)
- GA: June 2026
- Marketing launch campaign begins two weeks before GA

## What this file does not cover

- Pricing and packaging (see `orion-pricing-and-tiers.md`)
- Technical architecture (see `orion-technical-specs.md`)
- Full competitive analysis (see `competitive-landscape.md`)

Your First AI Hire: Building Agents That Know Their Job

Fri, 20 Mar 2026 00:00:00 +0000

I remember when I realized I was starting to use AI at work as if I were managing a team of AI employees . I got so excited, I immediately sketched the idea on a sheet of paper so I could share it with my teammates.

What started as a sketch is now core to how I use AI agents to do things faster and better at work and at home. It’s an approach that naturally guides you toward the context engineering best practices that improve LLM output.

The reason this approach works is that it uses one of the two available levers to improve how well a best-in-class large language model (LLM) works for you.

Fine-tuning: this is where you take an LLM and train it further using your own data so it becomes more specialized for your needs.
In-context learning: giving the LLM the right expertise (persona), knowledge (context files), workflows (skills), and rules in each session (team rules).

For most people, fine-tuning is going to be out of reach. Even if you could fine-tune a model, you’d have to retrain it repeatedly to keep up with changes in your work. Otherwise, the model would grow stale. In-context learning is how you keep the model relevant between retraining cycles, and for most people, it’s the only lever available.

It all starts with building your team.

Define the roles

The first step is to define the roles for your team by identifying the groups of similar tasks you do over and over again. It might help to start with pen and paper like I did.

List out the things you do at work or at home on your computer. Don’t overthink it; just write them down. Then group the ones that are similar in terms of how you approach them (required behavior) and the information you need to do them (required context).

The groups of items you do most often and that take the most time are the best candidates for roles because they’ll benefit the most from ongoing improvement. On those tasks, you can work with an agent frequently enough to spot gaps that lead to improvements that you’ll continue to benefit from.

As you build your team, keep in mind that the ideal number of direct reports for a manager tends to be 8-9 . This principle also applies to AI agents. The more you have, the more complex it gets to keep up with the feedback and improvement loop for each one. Remember, you’re not building a department. You’re building a team.

In my role as a Sr Product Marketing Manager, I’ve landed on five agents that I work with daily:

I’m setting up a different team at home: an editor, financial advisor, and personal trainer.

Create the personas

Creating the personas will be quicker than you think, because you’re going to use AI to help create them.

Start with the role that you feel the most comfortable defining. Spend a little time thinking about how you’d want the agent in that role to behave. What should it do? What should it never do? Don’t overthink it. You’re not going for perfection. You’re going for something that you can provide an LLM, like ChatGPT or Claude, to help it create a persona prompt for you. Keep it simple so you don’t get hung up on this step. The feedback loop will improve it over time.

Next, start a chat with the best-performing model you have access to. Regardless of what you are using, if you have the option to select a model, select the latest frontier model from that provider. Starting with a better quality model means you’re more likely to start with a good persona prompt. That’s less distance to close with the feedback loop to get to an agent that starts to materially improve the work it was created for.

In the chat, ask it to help you create a persona prompt. Let it know the role you have in mind, the type of work you’re going to use it for, and how you want the agent to behave. I’ve included a prompt at the end of this post that you can copy into Claude, ChatGPT, Gemini, or your tool of choice to walk you through creating your persona prompt.

Review what the model writes for you, and iterate on it as needed. If something doesn’t sound right, let the model know what the issue is and ask it to update the prompt. You don’t need to use any kind of special language to get this done. Treat it like a conversation you’re having with a coworker to improve a document. And remember what I mentioned before: there’s no need to be precious about this. This is a starting point that you’re going to refine through the feedback loop.

A good prompt is going to define the agent’s identity briefly (1-2 sentences) and focus primarily on behavioral guidance for the agent. This includes how to approach tasks, standards to enforce, and what to prioritize. It’s also helpful to include specific things the agent shouldn’t do in this type of role. For example, I don’t want my data wizard to ignore a sudden spike or decrease in a metric, because I’ve learned that generally doesn’t happen without some external factor causing it.

After you create your persona prompts, take a step back and think about how you created them. You delegated the persona draft to an LLM. That’s not a shortcut. You’re not cheating. That’s the whole point of creating your AI team. You’re going to be delegating more and more work to them, and this is the first point in the framework where you do that.

As you build trust with your agents, you’re going to start to delegate more to them: bigger tasks, more autonomy, more trust. This is exactly what it’s like to be a manager when you’re working with a new employee. You’re initially close to what they’re doing, you build trust, and then you start to give them the room to run. That’s when you start to really see the benefits of adding that employee to your team. It’s the same thing here. The earlier you get comfortable delegating work to the AI agents, the faster everything in the framework will start to pay off.

Set up the agents

The last thing you need to do to build your team is set up the agents by loading the persona prompt into whatever tool you’re using. The specifics are going to vary based on the tool you’re using, e.g. Claude Code versus Kiro CLI. I’ll cover how to do this in more detail in an upcoming post in this series. For now, you just need to remember that the persona prompt is the foundation for each agent on your team.

Building your AI team is straightforward. You’re the expert at what you do and how to do it well. Use your experience and expertise to guide an LLM to build persona prompts for AI agents to fill your open roles. That gets them hired. The knowledge base you’ll create is what gets them up to speed and delivering high-quality work for you.

Resource: prompt you can use with an LLM to help create your persona prompts

I need you to help me write a persona prompt — a set of instructions that will shape how an AI agent behaves every time it runs. Think of it as a job description the AI reads before every conversation.

Before writing anything, interview me. Ask these three questions one at a time, waiting for my response before moving on:

1. **What role does this agent play?** What's the domain and who does it serve? (If you know what platform or tools the agent will use, mention them — but don't worry if you're not sure.)
2. **What kinds of work will it do?** Describe the typical tasks or situations the agent will help with. Think about what a good day looks like — what does the agent do well?
3. **What behaviors matter most?** How should the agent approach its work? What should it do when it's unsure? Are there things it should always or never do?

After the interview, generate the persona prompt. Use what I told you as the foundation, but add your own recommendations — behaviors or guidelines that would make this agent more effective for the role, even if I didn't mention them. Call out anything you added so I can review it.

Follow these rules when writing the prompt:

### Focus on behaviors
- Describe what the agent should *do*, not what it *is*. "Start by understanding the full situation before proposing solutions" is a behavior the agent can act on. "You are thorough and thoughtful" is not — it's a personality trait, and the agent won't know how to translate that into action.
- Frame instructions as conditional guidance: "When X, do Y." This gives the agent concrete decision points rather than abstract qualities to live up to.
- If a behavior only applies sometimes, state the condition.

### Hit the right altitude
- Write at the level of a clear team lead briefing a competent new hire — not a legal contract, not a vague mission statement.
- Be specific enough to prevent the mistakes that actually happen, but flexible enough to let the agent use judgment in novel situations.
- Prefer "when X, prefer Y" over rigid step-by-step procedures. The agent needs guidance it can apply across situations, not a script that breaks the moment something unexpected comes up.

### Structure for the role
- Let the role dictate the structure. A coding agent needs different sections than a research agent or a writing coach. Don't force a template.
- Always lead with identity and scope — one or two sentences that establish who this agent is and what it does.
- After that, organize the remaining instructions into whatever sections make sense for this specific role. Use headers and bullets so the instructions are easy to scan.

### Keep it short
- A persona prompt competes with the user's actual questions and content for the agent's attention. The longer the prompt, the less room the agent has to focus on the real work.
- Aim for the shortest prompt that fully captures the desired behavior. If a line doesn't change how the agent acts, cut it.
- Leave out anything the agent would already know or can figure out from context.
- When in doubt, leave it out. A lean starting point that the user can build on is far more useful than a bloated prompt full of rules that haven't been tested. The user will discover what's missing by working with the agent and can add rules as needed.

### Output format
- Output only the final persona prompt, ready to use.
- After the prompt, add a short "Additions" section listing anything you added beyond what I described, with a one-line rationale for each. This section is for my review — it's not part of the persona prompt itself.

How I Manage a Team of AI Agents at Work

Mon, 16 Mar 2026 00:00:00 +0000

I used to think of AI as a tool I used. Now I think of it as a team I manage. This perspective evolved gradually as I used it daily and found myself rewriting the same persona prompts over and over again for the same types of tasks. I started systematically improving what I was doing until I found myself managing a team, which happened to be made up of AI agents, at the end of the 7-month journey that made me an AI enthusiast .

When you are managing a team, you have to scope the roles for your team and fill them with people who can be successful in those roles. To do that, you hire folks with the right backgrounds and experience, both of which inform how they’ll do the work. Every member of my AI team has a persona prompt with specialized behavioral guidelines I want for the role they’re filling. For example, my Data Wizard prompt has guidance around digging into irregularities in data, like when a metric suddenly spikes up or down.

You want to ensure your team has the information they need to do their work, like a wiki with product information, target audience insights, document templates, and standard operating procedures. You’re giving them the context they lack when they step into that role, and I do the same with a structured knowledge base, project-specific context files, and reusable skills that describe how to do specific tasks.

Lastly, you want to develop your team with feedback and guidance tailored to them. I’m using feedback loops to capture issues and improve their work via the persona prompts, team rules, skills, and knowledge base.

I built my team through trial and error, but I now have a framework for how to do this, which I’m breaking down into two phases: build your team and manage your team. Each phase contains three steps:

Build your team

Define the roles
Create the personas
Set up the agents

Manage your team

Work with your team
Spot the gaps
Apply fixes at the right level

In order to build the right team, you need to figure out what roles you need to fill. Start by identifying similar types of tasks that you have to do often in your role; those groups of tasks represent job openings that you could fill with an AI agent.

I’ve found that my most useful agents are the ones that I work with often, since that supports the ongoing cycle of improvement, so I try to avoid creating a custom agent with too narrow a scope that I won’t work with often. I also don’t want to have to juggle a team of 30 agents every day. I think the best practice of keeping a manager’s span of control to ~8 or fewer employees also makes sense in this context.

You’ll then build a persona prompt for each job opening that defines the ideal candidate’s identity and how they work. Once you have the persona, that serves as the foundation for the AI agent you’ll set up in a tool like Kiro or Claude Code to be on your team.

To manage your team, you have to understand their strengths and weaknesses, and that means working closely with them. You’ll want to work with the right agent for the task at hand.

As you work with them, you’ll start noticing recurring gaps that you need to address to improve your team. Some of those will be knowledge gaps, where the team needs more information, and others will be behavioral gaps, where your agents aren’t doing something the way you’d like them to or expect them to.

Based on the gap, you’ll want to address the situation at the right level. That might entail adding a new context file to the team’s knowledge base or updating an agent’s persona. These small tweaks will start to lead to big improvements, but this isn’t a set it and forget it kind of deal. It’s a continuous management process that never really ends.

I’ve noticed that my team is producing better work in less time with this approach, but I don’t have an objective way to measure or validate that. I want to learn more about LLM evaluation techniques so I can get to objective measurement, but in the meantime, I have some validation from others at work.

First, I’ve started to get compliments from copywriters on the draft marketing copy I’m writing with my Copywriter AI agent. Second, I shared the first draft of a monthly business review document my team wrote with my counterpart on the product side. I let her know it was all AI-generated (the analysis and the write-up), and I asked her to review and check if there was potential there. She was so impressed with the quality of the MBR, she started asking me questions about how I’d put it together and what my setup was. Lastly, I was able to write a good business requirements document from scratch in one day because the infrastructure was already in place.

Building and managing an AI team is much harder than just using a chatbot or the default agent that you get with something like Kiro or Claude Code. It takes upfront work to scope the roles, build the personas, and create the infrastructure to support the ongoing improvement.

A lot of that work will happen before you start to see the results, but it will start to compound. Pieces will build on top of other pieces, and things will get faster, both because you’ll start to optimize the process to your work style and because your agents will get better. I don’t have the data yet to prove this is better, but I’ve seen enough to think the effort is worth it. So I’m currently setting up the same approach at home with Claude Code.

That’s a high-level overview of my managed AI framework. I’m going to dive deep into each area of the framework with separate posts on building your team, setting up the knowledge base, working with skills and team rules to set expectations and requirements, and creating a feedback loop to drive the ongoing improvement. I’ll then cover how I’ve implemented this in Kiro CLI at work and Claude Code at home. I’ll end the series with a post on the learnings and best practices I’ve picked up along the way. By the end, you should have a good roadmap with explicit examples to allow you to set up your own team.

I want to develop my skills, deliver better results, and spend more time with my family. This framework is how I’m doing that. It’s a tool-agnostic approach that can help move you away from using one-size-fits-all tools to building an AI team that’s tailored to your needs and able to deliver better results for you.

I've Been AI-Pilled: My Journey From Chatbots to Custom Agents

Fri, 13 Mar 2026 00:00:00 +0000

I was slow to start using generative AI, but over the last 7 months, AI has fundamentally changed how I work. I’ve gone from occasionally using AI to write text, to using it to create Python scripts, to now having a team of five custom AI agents that I collaborate with daily. I’m seeing how quickly the benefits are compounding, and as a result, I’ve been AI-pilled.

I began learning about LLM-based gen AI in earnest in 2024. I read all the most popular books at the time, but my exposure remained primarily theoretical. I learned how LLMs work fundamentally, but the biggest practical takeaway was the idea of assigning a persona to chatbots to improve their output. That’s basic prompt engineering, e.g. “You are a copywriter with 15+ years of experience in consumer tech. Help me write a marketing email about this product.” On the rare occasion I used a chatbot, I always remembered to assign it a persona.

Last August, I joined a project at work that was the turning point for my AI enthusiasm. In that project, I had to manually build a large JSON file that would require a lot of ongoing updates. On a whim, I decided to see if I could use a chatbot to write a Python script to go from JSON to Excel and vice versa. That would allow me to make updates in Excel, which would be faster, and then generate the JSON programmatically, reducing the risk of errors. Within 30 minutes, I had a working prototype that ultimately saved me countless hours over the coming months.

I’m OK at Python, but I realized LLMs are much better. So, I began to write a lot of Python scripts this way to automate repetitive or time-consuming tasks, like resizing images or creating Word docs from copy I had in Excel files, in order to stay on top of the workload for the project.

I was soon using chatbots weekly for other things. I began paying attention to what model the chatbot was using, switching to the latest frontier models whenever possible. That had a noticeable impact on the quality of the copy and ideas that I was getting from the chatbots. Especially when I paired better models with a well-crafted persona and a collaborative approach.

I got tired of typing different versions of the same persona prompts whenever I started a new chat. Often I was too busy and moving too quickly to write something better than “You’re a [blank] with X years of experience.” It happened enough times that I realized I could save some time without sacrificing quality by creating reusable persona prompts for different types of tasks. At first I wrote them myself, and they were ok, but not great. By asking the chatbot to help me craft the persona, I was able to take them to the next level. I kept the prompts in Word docs so I could copy and paste them into the start of my chat sessions depending on what I was working on. To save a little more time, I’d pin the chat in the sidebar and rename it to something like “Copywriter” or “Data Wizard,” so I could quickly return to the right chat based on what I was working on.

I’d work with the same chat for up to a week because I wasn’t aware of context rot (where long, ongoing conversations with LLMs start to produce worse results). That’s ok though, because it led to another breakthrough for me. I started to ask the chatbots at the end of each week how we could improve the persona prompt I’d initially started the conversation with based on our interactions. The chatbot would suggest some ideas, and after a few revisions and back-and-forths, it would write an updated version that I would use in the next week’s chat. That created a feedback loop to improve my personas on an ongoing basis.

For example, I learned that LLMs guesstimate how long copy is after noticing that their character counts for marketing copy were often wrong. That’s not great when you’re writing ad copy that has specific character constraints. I updated my Copywriter persona prompt with instructions to count each individual character when writing against copy constraints. After that, I no longer had to worry about getting copy options that were too long for the character constraints I’d provided. It was like giving an employee feedback, except the chatbot immediately incorporated that feedback into how it worked.

That’s the core idea in the framework that is now guiding my AI usage: that I’m the manager of a team of AI agents. They’re incredibly smart, but also kind of dumb. They have a lot of expertise, but they’re also clueless about the specifics of where I work and what I’m working on.

The more effort I put into developing my team and providing what they need, the better the quality of work I get from them. And the benefits are compounding over time. I spend less time correcting simple issues and more time refining and improving what we’re working on. I spend less time providing the same context over and over again to my agents. Instead, they have a growing knowledge base to inform their work. I spend less time tweaking the documents they create because they have actual examples to refer to of the various types of documents I have to write. I’m capturing feedback and improving every aspect of my setup daily, and that makes it even better the next day and miles ahead of using a run-of-the-mill chatbot.

I’m going to go into more detail about this framework in coming posts and explain how I’ve implemented it in Kiro CLI , an AI coding tool I use at work primarily for non-coding tasks, and how I’m now implementing it in Claude Code at home.

Where technology meets the real work

Mon, 01 Jan 0001 00:00:00 +0000

Essays on AI, marketing, data, and productivity — from someone who tries everything before writing about it.

I’m a marketer with an engineering background. I spent the first chunk of my career as an engineer at IBM, writing test automation scripts and digging into server hardware. Then I pivoted, through consulting and into marketing, where I’ve spent the last 15+ years.

That combination gives me an unusual vantage point: I understand how technology works, and I spend my days figuring out how to actually use it to get things done.

Right now, most of my curiosity is pointed at AI — specifically, how non-technical people can move beyond occasional chatbot use and build something more systematic. Since August 2025, I went from using AI sporadically to managing a team of custom agents that I collaborate with daily. I’m writing about that journey as I go, because I find I understand things better when I try to explain them.

If you’re curious about technology and like practical over theoretical, you’ll find something here.

Get new posts in your inbox

One email per post. No spam, no fluff.

Your email stays private. Unsubscribe anytime.