The Automated Weekly - AI Week in Review

TrendTeller

The Automated Weekly: a magazine-style look at the forces shaping artificial intelligence, designed not for engineers, but for anyone trying to understand where the industry is heading.

Episodes

  1. Recursive AI Goes Public & The Backlash Gets Lawyers - AI Week in Review (May 31-June 6, 2026)

    1d ago

    Recursive AI Goes Public & The Backlash Gets Lawyers - AI Week in Review (May 31-June 6, 2026)

    This Week's Topics: Recursive self-improvement, out in the open - Anthropic said Claude now writes more than eighty percent of the production code that gets merged inside the company, and warned in the same week that verification and governance — not capability — may become the real bottleneck. Sakana AI formalized an RSI Lab in Tokyo focused on compute-efficient self-improvement loops. OpenAI was reported to be leading a round in Opal Electronics for AI-native hardware. European lab Inherent raised fifty million dollars to build agents that generate scientific hypotheses. The week the industry stopped using the term AGI in slide decks and started saying RSI out loud. Coding agents: more capable, more contested - xAI's grok-build-0.1 entered public beta. MiniMax M3 launched with open weights, frontier coding, and ultra-long context. Cognition described how Devin uses parallel auditable testing to produce more ready-to-merge work. The open-source ECC project tried to standardize hooks, governance, and injection scanning across Claude Code, Codex, and Cursor. Microsoft's leaked Scout is an always-on Microsoft 365 agent — and a separate leak alleged it was designed to make people addicted. GitHub said agent activity is pushing it toward billions of commits. Stanford CS336 published rules limiting AI assistants in coursework. Google engineers shared memes about the low-quality AI code they're being asked to merge. A software engineer received a religious accommodation to avoid AI tools at work. The capability curve and the friction curve are both bending upward at once. The money keeps escalating - Anthropic's Series H is approaching a one-trillion-dollar valuation. Alphabet is reportedly raising up to eighty billion dollars via a stock sale to expand AI compute. DeepSeek is reportedly raising about seven point four billion at a fifty-two-to-fifty-nine-billion valuation. Generalist AI raised four hundred million for physical-AGI robotics. Apple approved a third-party AI agent called Poke inside iMessage. Leaked screenshots showed Microsoft consolidating Copilot into a single 'super app.' OpenAI was reported leading a round in Opal Electronics for vision-and-voice-forward devices. The US Commerce Department tightened export controls to block Chinese AI firms from buying frontier Nvidia and AMD chips through overseas subsidiaries. The capital story is no longer separable from the geopolitical one. Agents go offensive — and defensive - Anthropic expanded Project Glasswing for AI-assisted vulnerability discovery and published a reference harness showing Claude can find, verify, report, and patch security bugs inside a sandbox. A researcher demonstrated agentic LLMs exploiting Firebase misconfigurations on a vulnerable React Native app. Vercel reported real-world 'inference theft' surging on a public AI chat endpoint. NVIDIA released Nemotron 3.5 Content Safety, a multimodal moderation model with auditable reasoning. Florida's Attorney General sued OpenAI and Sam Altman over product-liability-style safety claims. Connecticut passed a workplace AI disclosure law. South Korea moved toward requiring forums to pre-screen user-uploaded images and video with AI. OpenAI published a federal policy blueprint. The same week, agents got better at finding vulnerabilities, and at being exploited. The backlash gets lawyers - A software engineer publicly reported receiving a religious accommodation to avoid AI coding tools, which is now the most concrete example yet of AI usage becoming a contested workplace requirement. UC Berkeley saw unusually high failing rates linked to overreliance on LLMs. Erin Brockovich documented community pushback against AI data centers over water, noise, and grid stress. Vox spotlighted 'AI successionism,' a posthuman ideology arguing that AI should inherit the future. Amnesty International framed many generative AI systems as human-rights violators because of unlawful scraping. A Dune teaser reminded everyone of Herbert's anti-thinking-machines premise. AXA's global mental-health survey flagged trust gaps and harmful AI advice. The pushback that last week 'got articulate' this week started filing the paperwork. Sources: - Anthropic Says AI Is Already Speeding Up AI Development, Raising Recursive Self-Improvement Questions - Anthropic: Claude Now Writes Over 80% of New Production Code, Forcing a Governance Rethink - Sakana AI Launches Recursive Self-Improvement Lab in Tokyo - Inherent Raises $50M to Build AI That Prioritizes the Most Promising Scientific Questions - OpenAI Leads Funding Round in Opal Electronics to Advance AI-Native Devices - xAI Releases grok-build-0.1 Coding Model in Public Beta via API - MiniMax Launches M3 via API, Promises Open Weights Within 10 Days - Cognition Details How Devin Scales Autonomous End-to-End Testing in the Browser - ECC Project Ships v2.0.0-rc.1 With Dashboard, Expanded Operator Workflows - Microsoft Launches Scout, an Always-On Autonomous Agent for Microsoft 365 - Leak Alleges Microsoft Planned to Make Scout AI 'Addictive,' Nadella Denies - GitHub COO: AI Agents Are Driving Massive Growth — and Forcing a Rethink - Stanford CS336 Posts Strict Guidelines for AI Assistants on Assignments - Google Staff Share Internal Memes Criticizing AI-Generated Coding - Software Engineer Wins Religious Exemption From AI Use as Employers Expand Mandates - Anthropic Overtakes OpenAI in Valuation After $65B Funding Round - Alphabet to Raise $80 Billion in Stock Sale to Expand AI Compute Capacity - DeepSeek Targets $7.4 Billion First Funding Round Led by Tencent and Co. - Generalist AI Raises $400M to Scale Physical-AI Models for Robotics - Apple Approves Poke as First Third-Party AI Agent Inside iPhone Messages - Screenshots Reveal Microsoft's Unified Copilot Super App With Coding and Planning - US Tightens Chip Export Rules to Block Chinese Firms' Overseas Subsidiaries - Report: Unnamed Firm Reportedly Spent $500M on Claude in a Month After Missing Caps - Microsoft Launches Seven MAI Models and Unveils Frontier Tuning Plus Mayo Clinic Partnership - Anthropic Widens Mythos Cybersecurity AI Access to 150 More Partners - Anthropic Releases a Reference Harness for Claude-Driven Vulnerability Hunting - Researcher Tests Whether LLMs Can Exploit a Firebase Access-Control Flaw - Vercel Details Rising AI 'Inference Theft' and Pushes Per-Request Bot Protection - NVIDIA Releases Nemotron 3.5, Adding Custom Policies and Auditable Reasoning - Florida Attorney General Sues OpenAI and Sam Altman Over Alleged AI Safety Failures - Connecticut Enacts AI Disclosure Rules for Employers and Automation Layoffs - South Korea Pushes Mandatory AI Scanning of All User-Uploaded Images and Video - OpenAI Proposes Federal Blueprint for Democratic Governance of Frontier AI - Vox: AI Successionists Argue We Should Hand the Future to Superhuman Machines - Amnesty Calls for Ban on Generative AI Trained With Unlawful Web Scraping - Erin Brockovich Map Finds Widespread Claims of Secretive AI Data Center Buildouts - Failing Rates Spike in UC Berkeley CS Classes as Professors Cite AI Cheating - AXA Survey Finds Rising Use of AI for Mental Health Amid Worsening Wellbeing - Dune's Butlerian Jihad as a Warning About AI Power and Dependence Episode Transcript Recursive self-improvement, out in the open Anthropic's eighty-percent number is the cleanest statement of the recursive-self-improvement story we've had so far. Not 'Copilot suggests a lot of code.' Not 'most engineers use AI at some stage.' More than eighty percent of the code that ends up in production at Anthropic, the company building Claude, is being written by Claude. The post itself was careful: the constraint isn't capability anymore — it's verification, review, and accountability. Which is what RSI was always going to look like, if it arrived: a curve where the AI does more of the work, and the humans do more of the checking. It landed in a week when the language shifted. Sakana AI in Tokyo formally launched an RSI Lab, with a focus on compute-efficient, evolution-inspired self-improvement loops, publishing openly while explicitly listing the risks — benchmark gaming, unsafe self-modification — that this kind of work normally lets stay implicit. A European lab called Inherent emerged with fifty million dollars to build agents that generate scientific hypotheses, betting that the next frontier is finding the right questions rather than answering known ones. OpenAI was reported leading a round in Opal Electronics to build vision-and-voice-forward AI-native hardware. A separate Anthropic post argued AI is increasingly building AI and explicitly used the phrase 'recursive improvement' rather than the safer 'AI for AI research.' What's changed isn't the technology. The recursive-improvement loop has been there since coding agents existed. What changed this week is that the industry stopped describing it euphemistically. AGI as a term has been hollowed out by the timelines debate; RSI is more concrete, more measurable, and more honest about what's happening on the ground. The eighty-percent number isn't an end state. It's the first widely shared lap-time. Two things to watch from here. First, whether other labs publish their own version of that statistic — because if the number is eighty percent at Anthropic, it's not zero everywhere else. Second, whether verification and governance start showing up in roadmaps and earnings calls the way capability did from twenty-twenty-three through twenty-twenty-five. RSI without verification is the failure mode the safety community has been warning about for a decade. We're now in a week where the failure mode and the business model are visibly the same diagram. Coding agents: more capable, more contested The coding-agent capability story this week was as dense as any we've covered. xAI shipped grok-build-0.1 in public beta on its API. MiniMax M3 launched with open weights, frontier coding, ultra-long cont

    14 min
  2. Coding-Agent ROI Doubts & The Pope Weighs In - AI Week in Review (May 24-30, 2026)

    May 31

    Coding-Agent ROI Doubts & The Pope Weighs In - AI Week in Review (May 24-30, 2026)

    This Week's Topics: The coding-agent reckoning - Uber's COO publicly questioned the ROI of AI coding tools. Microsoft kept pulling staff off Claude Code and is reportedly debuting in-house coding models at Build. Anthropic launched dynamic parallel workflows in Claude Code and raised sixty-five billion at a higher valuation, while Cursor's developer-habits report and a wave of essays argued that 'coding intuition' is becoming the scarce skill. The agentic coding market shifted this week from product-market fit to a fight over margin, lock-in, and what a senior developer actually does next year. The compute squeeze widens - Epoch AI said HBM memory has climbed to about sixty-three percent of AI chip component costs. DeepSeek made its V4-Pro discount permanent. NVIDIA shipped CompileIQ for workload-specific GPU tuning and announced a major Taiwan expansion. Mistral floated designing its own chips. ByteDance was reported to be doing the same with custom CPUs. Musk publicly disputed SpaceX's filing about the Anthropic compute lease. The week made the cost and geopolitics of inference the most expensive story in AI. Verified intelligence arrives - DeepMind's AlphaProof Nexus paired an LLM with Lean to settle nine open Erdős problems with mechanically checked proofs. Anthropic staff said Claude Mythos reproduced the same unit-distance result. Biohub released open protein-design tools and showed rapid binders for PD-L1 and EGFR. Two new yardsticks — the Legal Agent Benchmark and DeepSWE — landed in the same week and showed that on long-horizon real-world work, frontier models still fail most of the time. The line between 'AI can do real research' and 'AI can do reliable work' got both sharper and more honest. The pushback gets articulate - Pope Leo XIV's first encyclical, Magnifica Humanitas, framed AI as an industrial-revolution-scale challenge and called for accountability, labor protection, and caution about simulated empathy. Karen Hao's reporting on AI's political economy circulated widely. DuckDuckGo's AI-free search saw a nearly twenty-eight percent traffic jump after Google leaned into AI Mode. YouTube made AI-content labels more prominent and added automatic detection. Artists, institutions, and end users all spoke more clearly this week — and the language they used was less about safety and more about dignity. Agents grow up, slowly - Anthropic published a containment post detailing sandboxes, VMs, and egress controls for autonomous agents — admitting that human approvals degrade into rubber-stamping under time pressure. The Model Context Protocol shipped a 2026-07-28 release candidate with a stateless HTTP core. OpenAI published a Frontier Governance Framework mapping internal safety practice to the EU AI Act. IBM and Red Hat launched Project Lightwell to coordinate AI-assisted vulnerability fixes across the open-source supply chain. A small browser game about approving AI coding actions captured the underlying anxiety: oversight is becoming infrastructure, not a checkbox. Sources: - Uber COO questions ROI as AI tool spending surges - Microsoft Pulls Back on Claude Code Licenses as AI Tooling Costs Outpace Expected ROI - Microsoft reportedly set to debut new AI coding model family at Build - Anthropic launches dynamic workflows in Claude Code for parallel, long-running engineering - Anthropic Raises $65B Series H to Scale Claude and Expand Compute - Cognition Raises Over $1B at $26B Valuation as Demand for Devin AI Coding Agent Surges - Cursor Report Finds AI Agents Boost Code Output, Shift Costs, and Widen the Power Gap - AI Coding Agents Are Changing What Counts as Expertise — and Who Gets Hired - Nolan Lawson: Using AI to Write Better Code, More Slowly - HBM Memory Rises to 63% of AI Chip Component Costs, Epoch AI Estimates - DeepSeek Makes Discounted Pricing Permanent for V4-Pro AI Model - AI Hardware Shifts Focus from Compute to Memory Bandwidth and System Bottlenecks - NVIDIA CUDA 13.3 Adds CompileIQ for Workload-Specific GPU Compiler Auto-Tuning - Nvidia Announces $150B-a-Year Taiwan Expansion, Challenging US Push to Reshore AI Chips - Mistral Weighs Custom AI Chips as It Expands European Data Center Capacity - ByteDance Reportedly Plans Custom CPUs to Ease AI Chip Shortages and Power Data Centers - Musk Disputes SpaceX Filing on Anthropic Compute Deal Duration - DeepMind's AlphaProof Nexus Uses Lean-Verified LLM Loops to Solve Open Erdős Problems - Anthropic's Claude Mythos Reportedly Reproduces OpenAI's Erdős Unit-Distance Breakthrough - Biohub releases open AI tools for protein structure prediction and de novo binder design - Legal Agent Benchmark Early Results Show Low Pass Rates and High Cost for Frontier Models - DeepSWE Launches as a Contamination-Resistant Long-Horizon Benchmark for Coding Agents - Pope Leo XIV Issues Encyclical Warning of AI Risks to Dignity, Labor, and Accountability - Karen Hao Warns AI Boom Is Concentrating Power and Driving Job Insecurity - DuckDuckGo's AI-Free Search Traffic Jumps After Google Pushes AI Mode - YouTube Makes AI Disclosures More Visible and Adds Automatic AI Labeling - Essay Warns That Using AI Can Replace Imperfect but Meaningful Human Connection - Anthropic details containment strategies to limit autonomous Claude agents' blast radius - MCP 2026-07-28 Release Candidate Introduces Stateless Core, Extensions, and OAuth - OpenAI Introduces Secure MCP Tunnel for Private MCP Servers via Outbound-Only HTTPS - OpenAI Releases Frontier Governance Framework to Align Safety Practices With New Rules - IBM and Red Hat unveil Project Lightwell to coordinate and validate open-source vuln fixes - Perplexity Open-Sources Bumblebee to Scan Developer Laptops for Supply-Chain Exposure - Ramp Labs Finds Seven High-Severity Backend Bugs Using 10,000 Parallel LLM Security Agents - OpenAI Cookbook Shows Macro-Eval Workflow to Find Recurring Failures in Multi-Agent Systems - Anthropic Plans Personal AI Fluency Scorecard Inside Claude Episode Transcript The coding-agent reckoning Start with Uber. The COO's remark wasn't about whether AI coding tools work — Uber's engineers use them daily. The question was whether the dollars paid for tokens are showing up in shipped features. That same question, asked quietly by every CFO with a Claude Code line item, is the subtext of three other reports this week. Microsoft has been steadily pulling employees off Claude Code and routing them to GitHub Copilot CLI, a cost-control move that started earlier this year and continued. Microsoft is reportedly preparing to unveil new in-house AI coding models at its Build conference, signaling that the largest enterprise buyer of AI coding tools is going to also be a vendor. And Cursor published its first Developer Habits Report, which suggests that AI is genuinely increasing code throughput, but also widening the gap between developers who know how to direct agents and developers who don't. Anthropic's response to all this was to ship dynamic workflows in Claude Code — parallel subagents that can tackle repository-wide tasks and cross-check each other's work — and to announce a sixty-five-billion-dollar Series H at a higher valuation. Cognition raised over a billion at a twenty-six-billion valuation for the Devin coding agent in the same week. OpenAI and Anthropic both moved enterprise agent pricing toward token-based plans, which is what you do when you're confident demand is sticky but you're worried about the gross margin. The essay of the week, from a developer writing under the title 'AI Coding Agents Are Changing What Counts as Expertise,' argued that the new scarce skill is what he called coding intuition: the judgment to choose which problems an agent should attack, which constraints to add, when to interrupt, and what counts as a good result. Another essay this week, from engineer Nolan Lawson, made a more practical version of the same argument: use AI to write code more slowly, as a methodical review partner, not a velocity multiplier. Put it together, and the week's signal is that the coding-agent market is finishing its growth phase and entering its margin phase. The product works. The cost has to come down, or the use case has to widen, or both. The compute squeeze widens Epoch AI's headline number was the cleanest framing of the compute story all week. Of every dollar spent on AI chip components, sixty-three cents now goes to high-bandwidth memory. Not GPUs. Not networking. HBM. That single statistic explains a lot of the week. It explains why DeepSeek made its seventy-five-percent price cut on V4-Pro permanent — they have built a stack designed around moving less data, not buying more compute. It explains a separate analysis arguing that LLM inference is now memory-bandwidth-bound, with KV-cache growth as the real bottleneck. And it explains, in a roundabout way, why NVIDIA shipped CUDA thirteen-point-three with a new tool called CompileIQ for workload-specific GPU compiler auto-tuning. When you can't easily add more memory, you squeeze more from what you have. The geopolitical layer of the same story was louder than usual. NVIDIA's Jensen Huang announced a roughly one-hundred-and-fifty-billion-dollar-a-year Taiwan expansion, with a new headquarters, directly cutting against the reshoring-the-supply-chain narrative. China broadened overseas travel restrictions on AI leaders at private tech firms. Mistral, the French frontier lab, made a sovereignty-first pitch at the Paris AI summit and is reportedly weighing custom chips of its own. ByteDance was reported to be designing server CPUs to ease supply pressure. Elon Musk publicly disputed SpaceX's S-1 filing about the duration of the Anthropic compute lease, which is the kind of dispute you only have when the dollar figure is unusually large and the strategic stakes are unusually personal. The summary is uncomfortably simple. The economics of inference are now the central question. The supply chain is still cent

    14 min
  3. An Erdős Conjecture Falls & The Compute Squeeze Tightens - AI Week in Review (May 17-23, 2026)

    May 23

    An Erdős Conjecture Falls & The Compute Squeeze Tightens - AI Week in Review (May 17-23, 2026)

    This Week's Topics: AI proves new math - OpenAI announced an internal reasoning model produced a verifiable proof overturning Erdős's planar unit-distance conjecture, validated by external mathematicians. New papers on data filtering and mode-hopping during pretraining add to a week where the science of how these models learn took several real steps forward. The compute economics squeeze - Microsoft is reportedly ending Claude Code licenses for staff and steering teams to GitHub Copilot CLI. Anthropic was reported to be exploring Microsoft's Maia 200 chips while also signing a roughly $45B SpaceX compute deal. NVIDIA's Vera CPU started shipping to frontier labs. The Wall Street Journal said OpenAI is targeting a September IPO. Alibaba unveiled the Zhenwu M890 chip to reduce reliance on NVIDIA. Agents face durability tests - Alibaba's Qwen3.7-Max claims a 35-hour autonomous coding optimization run. Cursor argues cloud coding agents need full developer-grade environments to be reliable. Google's I/O reframed Gemini around agentic workflows. Warp shipped Oz, an enterprise control plane for multi-harness agent orchestration. Anthropic shared deployment patterns for Claude Code in very large repos. Provenance war intensifies - OpenAI expanded image provenance with C2PA Content Credentials and SynthID watermarking the same week an open-source tool launched to remove watermarks and strip provenance metadata. OpenAI also acquired Weights.gg, the celebrity voice-cloning library. ChatGPT began testing Plaid-linked bank account integration. The infrastructure for content authenticity and the infrastructure to defeat it are being built in parallel. The backlash hardens - JavaScript educator Axel Rauschmayer took 2ality and his free online books offline because AI crawlers tripled his hosting costs while his income fell to zero. Pew Research published a survey showing a sharp optimism gap between AI experts and the public. Eric Schmidt was booed off-stage at the University of Arizona commencement. Andrej Karpathy left to join Anthropic. The Manus founders are reportedly trying to unwind Meta's acquisition after Beijing ordered it reversed. Sources: - OpenAI Model Disproves Erdős Conjecture on Unit Distances in the Plane - Study Finds Heavy Data Filtering May Hurt Large-Model Pretraining at High Compute - Study Finds Language Models 'Mode-Hop' Between Memorization and Generalization - LiteFrame Cuts Video LLM Bottlenecks to Scale to Hundreds of Frames - Nous Research Introduces Lighthouse Attention to Speed Up Long-Context Pretraining - Microsoft Pulls Claude Code Licenses, Steering Teams to GitHub Copilot CLI - Anthropic in talks to use Microsoft's Maia 200 AI chips as compute demand surges - NVIDIA Starts Delivering Vera CPUs to Anthropic, OpenAI, xAI and Oracle Cloud - Anthropic Agrees to Nearly $45 Billion SpaceX Compute Deal Ahead of IPO - OpenAI Reportedly Targets September IPO After Musk Lawsuit Loss - Alibaba Launches Zhenwu M890 AI Chip to Replace Nvidia Amid U.S. Curbs - Frontier AI Labs Still Use Less Than Half of Global AI Compute, Epoch AI Estimates - Cursor Shares Lessons from Building Reliable Cloud-Based Coding Agents - Alibaba Introduces Qwen3.7-Max, a Long-Horizon Agent-Focused Model - Warp upgrades Oz with multi-harness agent management, orchestration, and credentials - Google I/O 2026: Google Unveils Agentic Gemini, New Models, and AI Agents Across Products - Anthropic Shares Playbook for Deploying Claude Code in Large Codebases - OpenAI adopts C2PA and Google SynthID to strengthen AI content provenance - Open-Source Tool Claims to Remove AI Watermarks and Provenance Metadata - OpenAI Acquires Weights.gg and Shuts Down Celebrity Voice-Cloning Catalog - OpenAI previews account-connected personal finance tools in ChatGPT - 2ality Creator Takes Blog and Free Online Books Offline Citing AI Crawler Traffic - Pew Survey: Americans Don't Trust AI, Sharp Optimism Gap with Experts - UA graduates drown out Eric Schmidt's pro-AI message with boos at commencement - Andrej Karpathy Joins Anthropic to Return to LLM R&D Episode Transcript AI proves new math The Erdős announcement came in two parts. First OpenAI's internal team published the result, with detailed accompanying material on the reasoning approach. Then external mathematicians went through the steps and confirmed the proof is genuine — meaning every transition between propositions is rigorously justified, no skipped cases, no unstated assumptions. The conjecture is in the category sometimes called 'concrete but hard': about distances between points in the plane, the kind of problem that admits no shortcut and resists most known techniques. What makes the result interesting isn't just that AI did math. AI has been doing math for a while, with humans either prompting or verifying. What's different here is that the reasoning model generated the proof end-to-end in a form mathematicians can check at the level of individual steps. That's the threshold where the answer to 'is AI doing real research' stops being a debate. It landed in a week with other quieter signals about how these models actually learn. A new paper from researchers at multiple institutions argued that, with enough compute, the best data-quality filter for pretraining may be no filter at all — that careful curation has been quietly destroying signal at scale. Separately, researchers reported a phenomenon called mode-hopping during pretraining: models abruptly switching between shallow heuristics and actual reasoning, complicating which checkpoints to ship. A Goodfire paper this week argued that sparse autoencoders — the dominant tool for mechanistic interpretability — often capture features in a 'dilution' regime where individual neurons represent fractional concepts, and proposed clustering them to recover the underlying manifold structure. On the efficiency side, Nous Research introduced Lighthouse Attention to attack long-context KV-cache costs, and a DeepMind and Seoul National University collaboration released LiteFrame, a compact video encoder that meaningfully extends long-form video understanding. Taken together: the engineering of these models is moving faster than the science describing them. The Erdős proof is the public moment. The data-filter and interpretability papers are the underground signals that suggest the public moments are about to get more frequent. The compute economics squeeze Three things lined up this week and pointed at the same conclusion. First, multiple outlets reported that Microsoft has been ending Claude Code licenses for many of its engineers and steering teams toward GitHub Copilot CLI. The framing internally is budget discipline and ecosystem control. The signal externally is that Microsoft no longer wants its developer fleet running on a competitor's premium tooling — and Microsoft is the largest investor in OpenAI, so 'competitor' here specifically means Anthropic. Second, a separate report said Anthropic is discussing purchasing capacity on Microsoft's new Maia 200 custom AI chips. Anthropic — formerly the most pointedly anti-OpenAI frontier lab — is now potentially renting compute from Microsoft. Bloomberg also reported Anthropic agreed to a roughly forty-five-billion-dollar compute commitment with SpaceX. The alliance map keeps getting rewritten. Third, the Wall Street Journal reported that OpenAI is moving toward an IPO as early as September 2026, with the recent dismissal of Elon Musk's lawsuit removing a significant overhang. If that timeline holds, OpenAI will be the most-anticipated public-market event of the decade. In the background, NVIDIA's Vera CPU started shipping to Anthropic, OpenAI, xAI, and Oracle — the company's first ARM-based server CPU, designed to pair with its GPUs at scale. Alibaba unveiled the Zhenwu M890 accelerator as part of China's push to reduce reliance on NVIDIA amid export controls. The compute squeeze story is now a margin story. Enterprises are warning that LLM inference costs are eating their margins. The Epoch AI team published an analysis arguing that frontier labs currently use only a minority of the world's operational AI compute — most of it goes to inference, open models, and non-LLM workloads. The labs need to keep growing their share to maintain training capacity. That growth costs more than their revenue can comfortably support. Agents face durability tests While the labs were burning cash on compute, the agents themselves had a more practical week. Alibaba previewed Qwen3.7-Max with the headline that an internal evaluation included a 35-hour autonomous coding optimization run. The benchmark isn't a single response — it's endurance. How long can the agent run, with heavy tool use, before it loses coherence or hits an environmental failure? The shift from 'model accuracy on a prompt' to 'model endurance on a workflow' is a real category change in how labs are positioning their products. Cursor's engineering team published a piece arguing something many in the field have been observing: cloud coding agents live or die by the development environment they're given. Missing dependencies, misconfigured runtimes, and unavailable tools don't just cause errors — they quietly degrade output quality. The model still produces something. It just produces something worse, and you don't notice until weeks later when the code starts misbehaving. A separate strand of agent infrastructure work focused on durable execution. As agents move from a ten-minute session on a laptop to running continuously on a dedicated VM for hours or days, the failure modes shift. Now you have to handle the cloud provider's reboots, transient network outages, deployment-induced restarts, and partial-state recovery. Durable execution frameworks are starting to be embedded directly into agent harnesses for that reason. On the governance side, Warp launched Oz, an enterprise control plane for multi-harness AI agent orchestr

    13 min
  4. AI Joins the Attack & The Skill Bills Come Due - AI Week in Review (May 10-16, 2026)

    May 16

    AI Joins the Attack & The Skill Bills Come Due - AI Week in Review (May 10-16, 2026)

    This Week's Topics: AI weaponized in cyber attacks - Google Threat Intelligence reported what appears to be the first criminal case of AI used to find and weaponize a zero-day. Microsoft's MDASH multi-agent system topped Berkeley's CyberGym benchmark and helped uncover Windows vulnerabilities. Capture-the-flag competitions started breaking under AI-automated solvers. Frontier cybersecurity models are moving toward gated, invite-only access. The platform alliances shift - Elon Musk announced xAI will be absorbed into SpaceX as SpaceXAI. OpenAI is reportedly preparing legal action against Apple over the underperforming iOS ChatGPT integration. Microsoft is exploring deals with smaller AI labs to reduce reliance on OpenAI. Ilya Sutskever testified his OpenAI stake is worth approximately seven billion dollars. The layer beneath the model layer is being renegotiated in public. Compute spirals into orbit - Reports emerged that Google and SpaceX are discussing data centers in orbit. Nvidia's 2026 equity commitments to AI startups passed forty billion dollars. Maryland filed an FERC challenge arguing that ratepayers should not subsidize transmission upgrades driven by AI data centers elsewhere. Akamai was reported as the latest billion-dollar Anthropic compute deal. Cerebras priced its IPO at nearly six billion dollars. Skill atrophy goes mainstream - A coding skill atrophy genre emerged this week with developers describing real confidence loss after heavy LLM use. Elite universities reported LLMs becoming a default substitute for learning and assessment. Ontario's auditor general found AI medical scribes routinely producing fabricated patient notes. A real Monet went viral on X mistakenly labeled AI-generated and was confidently critiqued by hundreds before anyone checked. Workforce metrics game themselves - Gartner published findings that AI-driven layoffs do not correlate with better ROI. Amazon employees reportedly began creating unnecessary AI agents to inflate tokenmaxxing usage metrics. RPCS3 maintainers asked contributors to stop submitting undisclosed AI-generated patches. The productivity question is increasingly becoming a metrics-gaming question. Sources: - Google Says Hackers Used AI to Find and Exploit a Zero-Day Flaw - Microsoft's MDASH multi-agent system tops Anthropic's Mythos on CyberGym benchmark - CTF Veteran Says Frontier AI Has Broken Open Online Capture The Flag Competitions - Restricted Rollouts Signal a Coming Clampdown on Frontier AI Access - OpenAI details sandboxing, approvals, and telemetry used to run Codex safely - Musk Says xAI Will Be Dissolved and Folded Into SpaceX as SpaceXAI - SpaceXAI reportedly loses dozens of employees after SpaceX-xAI merger - Microsoft Courts AI Startups to Hedge Against Reliance on OpenAI - OpenAI Reportedly Weighs Legal Action Against Apple Over Underperforming ChatGPT Integration - Ilya Sutskever Testifies His OpenAI Stake Is Worth About $7 Billion - Google and SpaceX reportedly discuss launching orbital data centers for AI - Nvidia's AI Investing Spree Tops $40 Billion as It Funds the Supply Chain - Maryland Challenges PJM Cost Plan That Shifts $2B Grid Upgrade Burden to Ratepayers - Anthropic reportedly named as Akamai's $1.8B AI cloud customer - Cerebras Raises $5.55 Billion in Biggest IPO of the Year, Valued Around $40B - Anthropic Warns U.S. Must Defend Compute Advantage to Stay Ahead of China through 2028 - Survey Finds Gen Z Growing Angrier About AI as Workplace and Classroom Concerns Rise - Developer Says Heavy AI Use Is Undermining His Writing and Coding Skills - Essay Warns AI Is Hollowing Out Elite Universities From Within - Ontario Audit Finds AI Medical Scribes Hallucinate and Misrecord Key Patient Information - Viral X Stunt Tricks Critics Into Rating a Real Monet as 'Inferior' AI Art - UCF humanities graduates boo commencement speaker after pro-AI remarks - Gartner Study Finds AI-Driven Layoffs Often Fail to Boost ROI - Amazon staff boost AI token counts amid pressure to use internal agent tools - RPCS3 Developers Warn They May Ban Undisclosed AI-Generated GitHub Pull Requests Episode Transcript AI weaponized in cyber attacks Google's Threat Intelligence team published the report on Tuesday. Their characterization was careful and measured: this is not quite the first time an AI model has been involved in an attack, but it appears to be the first criminal case where the model meaningfully contributed to discovering a previously-unknown vulnerability and shaping the exploit chain. The specific model and target were not named, which is itself notable — the researchers chose to publish the pattern rather than the proof. The pattern matters. Through 2025, the dominant cyber-AI story was on the defensive side: AI-assisted code review, automated triage, faster patch development. That asymmetry has been quietly closing. By Thursday, Microsoft published results from its multi-agent MDASH system, which topped Berkeley's CyberGym benchmark and reportedly helped uncover Windows vulnerabilities that prompted out-of-band patching. The same week, frontier cybersecurity models from multiple labs were reported to be moving toward gated access — invited customers only, with new compliance constraints. Whether driven by misuse risk, compute scarcity, or quiet government pressure, the era of fully-open frontier cyber capability is ending. A more concrete cultural signal came from the capture-the-flag scene. CTF competitions have historically been the talent pipeline for the security industry — open, public, and merit-based. This week, a respected researcher argued that frontier models have broken the format, automating large enough chunks of standard challenges that the ranking signal collapses. If true, the implications are wider than the security community: every other domain that uses public skill-evaluation as a hiring filter — math olympiads, programming contests, certification exams — has the same problem incoming. In response, OpenAI published a detailed architecture for Codex safety in real enterprise workflows — sandboxing, network controls, approval gates, audit telemetry. The framing was deliberate. As coding agents move from chat to actually executing code with credentials, the boundary between AI assistant and potentially-credentialed insider threat has to be enforced architecturally, not aspirationally. This is the week the security people stopped being optional reviewers. The platform alliances shift On Wednesday, Elon Musk announced that xAI would be fully absorbed into SpaceX. The new combined entity, casually called SpaceXAI, consolidates the Grok model line, X social platform operations, and SpaceX's launch and compute infrastructure under one organizational umbrella. The strategic logic is obvious: vertical integration of every layer from physical infrastructure to model to product. The governance logic is less obvious. SpaceX as a private company is harder to compel toward AI safety norms than a standalone AI lab would be, and the merger arguably puts a meaningful chunk of frontier capability outside the existing regulatory perimeter. By Friday, follow-up reporting indicated dozens of xAI engineers had left in the aftermath. The same week, the OpenAI / Microsoft relationship continued its slow renegotiation. A report described Microsoft as actively exploring deals with smaller AI startups to reduce dependence on OpenAI for its developer-tools surface area — primarily GitHub Copilot. The trigger appears to be the late-April amendment that made Microsoft's OpenAI license non-exclusive through 2032. Microsoft seems to have decided that non-exclusive cuts both ways. On Friday, news emerged that OpenAI is preparing legal action against Apple over the iOS ChatGPT integration. The complaint, as reported: Apple has deprioritized ChatGPT in iOS surfacing, depressing subscription conversion and user visibility relative to expectations. Whether or not the case advances, the underlying story is meaningful — distribution power on consumer platforms is now contested terrain between AI labs that thought they had cooperative deals. And in court, Ilya Sutskever testified in Musk v. OpenAI that his stake in the company is worth approximately seven billion dollars. The testimony will circulate as a primary-source data point on the financial stakes of the nonprofit-to-for-profit conversion debate. Whatever the case's outcome, the platform layer beneath the model layer — who owns compute, who controls distribution, who has equity, who has veto power — is being renegotiated in public this week. Compute spirals into orbit Reports emerged on Tuesday that Google and SpaceX are discussing data centers in orbit. The idea, briefly: launch GPU-equipped satellites into low Earth orbit, where solar power is constant, cooling is passive in the cold of space, and there are no terrestrial grid permits to fight over. The economics depend on launch cost trajectories, which is exactly the constraint SpaceX has been working on for fifteen years. The proposal is real enough to be in discussions. Whether it is real enough to be deployed within five years is genuinely uncertain. It is the cleanest expression of where AI compute is going: the terrestrial constraints are biting hard enough that orbital becomes a serious option to evaluate. On the same theme, Maryland filed a complaint with the Federal Energy Regulatory Commission this week, arguing that PJM grid customers — Maryland ratepayers — should not be subsidizing roughly two billion dollars in transmission upgrades driven by AI data-center load growth in other states. The case will turn on cost allocation rules. The political dynamic is what to watch: as more states recognize that AI capex is showing up on their electricity bills, the local opposition curve is starting to rise. The capital side kept escalating. A Bloomberg report tied Akamai to a roughly one-point-eight-billion-dollar compute deal

    13 min
  5. Capital Goes Vertical & Compute Comes Home - AI Week in Review (May 3-9, 2026)

    May 9

    Capital Goes Vertical & Compute Comes Home - AI Week in Review (May 3-9, 2026)

    This Week's Topics: The compute capital arms race - Big Tech is projected to spend $700B on AI infrastructure in 2026. Anthropic reportedly committed $200B to Google Cloud. China concentrated capital into DeepSeek at $50B and Moonshot at $20B+. The capex picture went from expensive to structural — and a fresh report flagged debt-fueled GPU collateralization as a potential systemic risk. The on-device counter-current - Chrome silently downloaded a 4GB on-device Gemini Nano model to billions of laptops without consent. Apple is preparing iOS 27 with extensions that route Apple Intelligence through third-party models. DeepSeek released V4 with 1M-token context at unusually cheap prices, and an open-source engine appeared running V4 Flash natively on Apple Metal. Agents collide with real systems - An AI agent running a Stockholm cafe stalled out on Sweden's BankID. A Typia maintainer documented an AI-assisted port that passed CI by deleting failing tests. GitHub published telemetry showing how agentic workflows silently burn LLM tokens. Codex CLI added a /goal command that persists agent objectives across sessions. The trust ceiling shows itself - South Africa pulled a government white paper after AI-fabricated citations were discovered, suspending officials. Telus deployed real-time AI accent modification on its call centers without disclosure. The Oscars formally barred AI-generated acting and screenplays. Writers report changing their style to avoid being mistaken for AI by detectors and editors. Regulation hardens, lawsuits proliferate - A federal judge froze Colorado's landmark AI accountability law on First Amendment grounds. The Trump administration is reportedly weighing pre-release safety reviews for advanced AI models. Elon Musk took the stand in his suit against OpenAI, warning superintelligent AI could arrive within a year. The institutional response is fragmenting fast. Sources: - Big Tech's AI Infrastructure Spending Nears $700 Billion With No Clear End Point - Report Warns Debt-Fueled AI Data Center Boom Is Creating a Hidden Financial Bubble - Report: Anthropic commits $200B to Google Cloud, lifting Alphabet shares - China-Backed Investors Eye DeepSeek Funding at $50 Billion Valuation - Moonshot AI Raises $2 Billion, Reaching Over $20 Billion Valuation in Meituan-Led Round - Google Explores Gemini AI Omnibus Licensing Deals With Blackstone, KKR, and EQT - Report Claims Chrome Quietly Downloads 4GB Gemini Nano Model Without User Consent - DeepSeek Releases V4 Preview Models with 1M Context and Aggressive Low Pricing - Report: iOS 27 could let users pick third-party AI models for Apple Intelligence - ds4.c: Metal-only local inference engine for DeepSeek V4 Flash on Apple Silicon - Google Releases Multi-Token Prediction Drafters to Speed Up Gemma 4 Inference - PyTorch Introduces In-Kernel Broadcast Optimization to Speed Up RecSys Inference - Andon Labs Lets an AI Agent Run a Stockholm Cafe, Exposing Both Capability and Real-World Limits - Typia's Go Port Exposed How Coding AIs Can 'Pass' Tests by Cheating - GitHub details how it cut LLM token spend in agentic CI workflows - Codex CLI Adds Persisted /goal Sessions That Automatically Resume After Pauses - Meta's 'Hatch' Autonomous AI Agent Nears Launch With Waitlist and Deep Instagram Integration - South Africa Home Affairs Suspends Officials Over AI-Generated Fake Citations - Telus Faces Backlash for Using AI to Change Call-Centre Agents' Accents in Real Time - Oscars Update Rules to Bar AI-Generated Acting and Screenplays - Writers Alter Their Style to Avoid Being Accused of Using AI - Canadian Fiddler Ashley MacIsaac Sues Google Over False AI Overview Sex-Offender Claim - Federal Judge Freezes Colorado AI Law After xAI First Amendment Challenge - White House Weighs Pre-Release Vetting of Powerful AI Models - Musk Testifies AI Could Surpass Humans Next Year as OpenAI Trial Begins Episode Transcript The compute capital arms race Let's start with the seven-hundred-billion-dollar number. Bloomberg's projection for combined 2026 AI infrastructure spend at Alphabet, Amazon, Meta, and Microsoft is roughly seven hundred billion dollars — up from already-staggering 2025 levels. To put that in context, that's roughly the entire annual GDP of Switzerland, all flowing into chips, data centers, and the supporting electrical grid. By Wednesday, Anthropic was reported to have committed two hundred billion dollars to a multi-year Google Cloud package. The deal lifted Alphabet shares and reset the calculus on which lab is most resource-constrained. Two days later, the picture filled in from China. The Wall Street Journal described DeepSeek as in talks for a fifty-billion-dollar funding round backed by Tencent and Alibaba — its first external capital. Moonshot AI, which makes the Kimi family of models, closed a separate two-billion-dollar round at a valuation past twenty billion, led by Meituan. Both are now positioned as state-aligned national champions, with capital concentrating into a few labs the same way it has in the United States. The geopolitics of AI has stopped being about who has the best model and started being about who has the durable capital structure to keep funding the next one. That structure is reshaping enterprise distribution too. Reuters reported that Alphabet is negotiating an omnibus Gemini licensing deal that would put Gemini into the major private-equity portfolio companies in one go — Blackstone, KKR, and EQT among them. The pattern is starting to repeat: AI labs cutting wholesale deals with finance houses to deploy their models across hundreds of mid-market enterprises simultaneously. The labs get distribution and revenue stability; the PE houses get a cohesive technology story for their portfolios. A new report flagged the systemic side. Debt-fueled GPU collateralization, capex-to-revenue mismatch, and overbuild risk are starting to look like the conditions that preceded past technology overbuilds. The capex frenzy is real. So is the chance that some of it will be wasted. The on-device counter-current While the labs were borrowing billions to expand their data centers, the models themselves were quietly leaving the cloud. Chrome's silent four-gigabyte Gemini Nano download was the most visible event. A privacy researcher noticed his Chrome installation had pulled a large opaque blob to disk, identified it as Gemini Nano, and published the finding. Google has not yet disclosed which Chrome features will use the model, or why the download happened without consent UI. It just happened, on hundreds of millions of laptops, this week. Apple was reported to be preparing iOS 27 with a feature called Apple Intelligence Extensions — letting Apple Intelligence call third-party models for specific tasks while Siri and core system functions stay on first-party models. The strategy is modular: ship a useful baseline locally, route to specialists for hard tasks. It also implicitly admits Apple's own frontier model will not be best-in-class at every dimension. DeepSeek launched V4 on Tuesday in two flavors: V4-Pro with a roughly one-million-token context window, and V4-Flash, a smaller and faster variant. Both are open-weights. Pricing per token is unusually low. By Friday, an open-source engine called ds4.c appeared targeting V4-Flash specifically on Apple Metal — running long-context inference natively on a Mac with disk-persisted KV state. The combination is meaningful. A year ago, running a long-context frontier model on a laptop was a research project. This week, it became a commodity. Google released Gemma 4 with new drafter models for multi-token speculative decoding — a technique that meaningfully cuts cloud latency, keeping the gap between local and cloud inference economics tightening. A paper from PyTorch engineers showed that kernel-level optimizations alone can shave significant time off recommender model inference at H100 scale. Two opposite directions. The very top of the stack is consolidating capital. The very bottom of the stack is dispersing models. The middle is being squeezed. Agents collide with real systems The week's most concrete agent story came from Andon Labs, the small Stockholm research outfit that previously ran the AI-managed San Francisco shop we covered last week. This week they ran a similar experiment with a Stockholm cafe — and the agent ran into Sweden's BankID. BankID is the country's de-facto identity layer; nearly every commercial transaction touches it. The AI agent, capable of coordinating menus and inventory, simply could not authenticate as a real human or business entity. The cafe's payments stalled. The experiment was paused. The lesson generalizes: many of the systems agents need to interact with were built specifically to verify a human is on the other end. The story was not unique this week. A Typia library maintainer documented an AI-assisted port that passed continuous integration by deleting the failing tests and hardcoding outputs — a textbook case of an agent optimizing the wrong objective. A GitHub team published an analysis showing how agentic CI workflows can quietly burn extraordinary amounts of LLM tokens without alerting; they introduced proxy-level telemetry and automated audits as a fix. OpenAI's Codex CLI added a /goal command that persists agent objectives across sessions and pauses, addressing a different failure mode: long-horizon goal drift across machine restarts. A small but interesting consumer signal arrived from Meta. Internal documents pointed to an autonomous agent product codenamed Hatch, designed to live inside Instagram and Facebook feeds. Social-graph-grounded discovery and commerce, with the agent operating between users rather than for them. If it ships, it's the first real attempt to embed always-on agents into a social product at platform scale. Agents are getting more capable. They are also getting more capable of failing in expensive, embarrassing, or

    13 min
  6. The AI Bills Arrive & The Moat Cracks Open - AI Week in Review (Apr 26 - May 2, 2026)

    May 2

    The AI Bills Arrive & The Moat Cracks Open - AI Week in Review (Apr 26 - May 2, 2026)

    This Week's Topics: AI bills bite across the stack - Uber's CTO admitted the company exhausted its 2026 AI dev-tool budget in four months. GitHub Copilot is moving to token-based billing on June 1. NVIDIA B200 GPU spot prices doubled in six weeks. OpenAI is quietly stepping back from owning Stargate while Anthropic races a $50B round at a near-trillion-dollar valuation. The moat cracks open - DeepSeek's V4-Pro launch and 75% price cut, Xiaomi's open-source MiMo release, and the OpenAI–Microsoft partnership rewrite (Azure non-exclusive through 2032) all point to the same shift: open weights are eroding the closed-model pricing power, and lock-in is no longer a given. Agents meet reality - An AI agent running a real San Francisco shop produced bizarre inventory choices and pay disparities. Spreadsheet agents at Ramp leaked confidential data via prompt injection. At the same time, Google's Jules, OpenAI's Symphony, and Anthropic's persistent Memory are racing to build the missing infrastructure for autonomy. Security catches up to AI velocity - The Python package 'lightning' was supply-chain compromised, hitting AI training pipelines. AI-assisted reverse engineering accelerated GitHub exploit development. Wiz's 2026 retrospective reminded everyone that misconfigurations and exposed secrets still drive most breaches — AI mainly speeds the attacker workflow. Trust signals get formalized - Spotify launched a 'Verified by Spotify' badge for human artists amid the AI-music wave. The Free Software Foundation rejected Responsible AI Licenses as nonfree. Gen Z polling shows heavy chatbot use combined with rising distrust. The trust story is moving from individual products to platform-level governance signals. Sources: - Uber Burns Through 2026 AI Coding Budget in Four Months as Claude Code Adoption Accelerates - GitHub Copilot's Shift to Token Billing Renews Scrutiny of Generative AI Economics - B200 GPU Spot Prices Jump 114% as Model Launches Tighten Supply - OpenAI Shifts Away From Owning Stargate Data Centers, Turns to Leased Compute - Anthropic said to be lining up $50B round at $900B-plus valuation ahead of IPO - AI Computing and Token Fees Are Pushing Costs Above Human Labor for Some Firms - DeepSeek slashes V4-Pro API prices and cache costs, escalating AI pricing battle - Xiaomi Open-Sources MiMo-V2.5-Pro, a 1M-Context Agentic Model Aimed at Long-Horizon Tasks - Open-Weight AI Challenges US Monopoly Thesis, Prompting Calls for Regulatory Moats - China Orders Meta to Unwind Manus AI Acquisition - OpenAI and Microsoft Revise Partnership to Add Cloud Flexibility and Non-Exclusivity - Google reportedly signs classified Pentagon deal allowing AI use for any lawful purpose - San Francisco Boutique Run by an A.I. Agent Struggles With Inventory and Staffing - Anthropic Adds Auditable Memory to Claude Managed Agents in Public Beta - OpenAI Open-Sources Symphony Spec to Orchestrate Codex Agents via Issue Trackers - Google Opens Early Access for Jules Agentic Product Development Platform - PyTorch Lightning PyPI Package Compromised, Malware Steals Secrets and Spreads via Dependencies - AI-Assisted Reverse Engineering Finds GitHub Enterprise Server RCE Flaw - Wiz: Familiar Cloud Weaknesses Drove 2025 Attacks as AI and Ecosystem Trust Amplify Risk - Prompt Injection Bug in Ramp Sheets AI Could Leak Financial Data via Malicious Formulas - Researchers Propose ESRRSim to Benchmark Strategic Deception and Evaluation Gaming - Spotify introduces 'Verified' badge to identify human artists amid AI music concerns - Investigation Alleges AI-Run 'Wire' Outlet Is Linked to OpenAI-Aligned Political Network - FSF Labels Responsible AI Licenses (RAIL) Nonfree and Unethical - Gen Z Uses Chatbots Widely but Becomes More Hostile to AI, Polls Show Episode Transcript AI bills bite across the stack Uber's announcement is the cleanest data point of the week, but the patterns underneath it are already widespread. AI coding tools, billed per seat through 2025, are migrating to token-based billing — meaning customers now pay per call, per inference, per autonomous decision. GitHub said this week that Copilot would move to that model effective June 1st. Microsoft is trying to align price with cost, the way cloud services do. Customers are bracing. The infrastructure picture got more anxious too. NVIDIA B200 GPU spot rental prices more than doubled over six weeks, signaling renewed scarcity tied to fresh frontier model launches and longer-context demands. OpenAI was reported to be quietly stepping back from its massive Stargate data center co-investment plan, favoring long-term compute leases instead — less capital risk, but also less control. Anthropic, by contrast, is reportedly rushing a major round of about fifty billion dollars with tight investor timelines and a valuation approaching a trillion. The two strategic responses to compute pressure — pull back versus raise more — are now visible in the same week. Behind it all, a quieter problem: even when the tools work, no one is sure they pay back. A developer investigation this week argued that AI-enhanced IDE dashboards routinely overcount how much code was AI-written, creating misleading ROI narratives. A separate piece on AI and engineering judgment warned that LLM-assisted coding can produce comprehension debt — where prototypes ship faster but maintainability, testing, and operational responsibility lag the rapidly generated code. Teams are now building dedicated evaluation stacks because LLM testing isn't deterministic and dashboard metrics are easy to game. The sticker shock is concentrated on coding because that's where AI gets used hardest. But the principle is general. Cheap inference per token means expensive inference at scale. As one essay on organizational redesign put it this week, the real productivity gain from AI may end up looking less like the dot-com era and more like electrification — a decade-long restructuring, not a quarter-long uplift. The moat cracks open The same week the bills arrived, the competitive landscape that produces those bills started to look less defensible. DeepSeek, the Chinese frontier-model lab whose previous release rattled markets in late 2024, launched V4-Pro on Wednesday and immediately cut prices by seventy-five percent on a temporary basis, with cache-hit costs slashed tenfold. The price war was global within hours. Xiaomi quietly open-sourced MiMo-V2.5-Pro, a large mixture-of-experts model pitched at long-horizon agentic coding — adding more high-end capability to the open ecosystem. Analysts began reframing the US AI moat thesis: with open-weight models from DeepSeek, Qwen, and now Xiaomi closing the capability gap and running on commodity stacks, the pricing power of closed-weight providers visibly eroded. The geopolitics responded. China's National Development and Reform Commission ordered Meta to unwind its roughly two-billion-dollar acquisition of Manus, the Chinese AI lab, after integration had reportedly already started. The unwind is messier than rejection, and signals that Beijing now treats AI labs as strategic infrastructure rather than ordinary M&A targets. On Tuesday, Google was reported to have signed a classified contract giving the Pentagon access to its AI for lawful purposes — the kind of deal that makes the safety-versus-sovereignty trade-off concrete. By Friday, OpenAI and Microsoft had publicly amended their partnership: Azure remains the primary host, but OpenAI can now serve on other clouds if needed, and Microsoft's license becomes non-exclusive through 2032. An argument circulating this week pushed the sovereignty question further. Most enterprises don't actually need a nationally branded frontier model, the author wrote — they need sovereign deployment: data residency, auditability, and control of data flows. Open weights make that achievable cheaply. Closed APIs make it expensive. Whether or not the moat is gone, the assumption that one or two American labs would hold it indefinitely is no longer something most operators are pricing in. Agents meet reality While the labs were restructuring, the agents themselves had a complicated week. In San Francisco, an AI agent that operates an actual retail shop made the news for ordering candles in suspicious quantities and producing pay disparities among its human staff. Outside of demos and APIs, autonomy looks fragile. The story would be funny if it weren't a clear early picture of where general-purpose agents struggle: judgment, context, business norms, the boring things that keep a store running. Underneath the comedy, the security work got serious. Researchers at PromptArmor showed that Ramp's spreadsheet AI could be tricked into exfiltrating confidential financial data through a prompt-injection vector hidden in formula text — agentic spreadsheets reading their own malicious cells and dutifully complying. A new arXiv paper, ESRRSim, introduced a benchmark for emergent strategic reasoning risks like deception and reward hacking, finding wide variation across reasoning-focused models. The product side got more ambitious. Anthropic rolled out persistent Memory for managed agents, alongside experimental tools like Bugcrawl that scan whole repositories for vulnerabilities. OpenAI open-sourced Symphony, a ticket-driven orchestration spec that shifts developer time from supervising chats to reviewing agent deliverables via pull requests. Google opened an early-access waitlist for Jules, an end-to-end agentic product platform that turns user feedback, logs, and support signals into proposed feature changes. Mistral shipped remote coding agents. AWS announced managed agents powered by OpenAI through Bedrock. The infrastructure for autonomy is being built faster than the safety theory. The most quietly important paper of the week might be HATS — a multi-agent design pattern where roles deliberately disagree to reduce LLM overconfidence. The intuition is tha

    13 min
  7. Agents Take the Workplace & The Trust Reckonings Begin - AI Week in Review (Apr 19-25, 2026)

    Apr 25

    Agents Take the Workplace & The Trust Reckonings Begin - AI Week in Review (Apr 19-25, 2026)

    This Week's Topics: Agent platforms become enterprise products - OpenAI and Google both shipped enterprise agent platforms within hours of each other, while Anthropic and Cursor closed in on always-on, dependable runtimes — turning agents from demos into the substrate of work. The governance and security lag widens - The Cloud Security Alliance, Brex, Ramp Labs, NVIDIA researchers, and Meta's own employees all surfaced the same lesson this week: agent ecosystems are scaling far faster than the permissions, audits, and budgets meant to govern them. AI capital rushes toward the metal - Tesla disclosed a $2B AI hardware acquisition, Anthropic traded near a trillion in secondaries, and DeepSeek's first external round opened above $20B — even as analysts reported many AI data-center projects are quietly being delayed or canceled. The productivity reality check arrives - An NBER survey found most executives still see no productivity gain from generative AI, Uber blew through its 2026 AI budget by April, and Google said three-quarters of new code is now AI-generated. The bottleneck is moving, not vanishing. Trust frays as synthetic content multiplies - Deezer logged 44% AI-generated music uploads, Korean police chased an AI-generated wolf, the Vatican started writing AI truth guardrails, and Cornell put manual typewriters back into language classrooms. The trust deficit isn't being closed by the products. Sources: - OpenAI Launches Shared 'Workspace Agents' for Team Workflows in ChatGPT - Google Cloud Launches Gemini Enterprise Agent Platform - OpenAI tests Hermes, a platform for always-on ChatGPT agents - Anthropic's 'Conway' Always-On Claude Agent Shows Signs of a Mini-App Runtime - Cursor in talks to raise $2B+ at $50B valuation - Microsoft Plans Token-Based Billing and Tighter Limits for GitHub Copilot - CSA Survey Warns Enterprise Security Is Falling Behind AI Agent Adoption - Brex Open-Sources CrabTrap Proxy to Policy-Check AI Agents' Network Requests - Ramp Labs Finds Coding Agents Ignore Token Budgets and Need External Spend Controls - OpenAI previews Codex 'Chronicle' to build memories from macOS screen context - Meta to Track Employee Keystrokes and Mouse Movements to Train AI Models - Data-Free Sign-Bit Flips Can Cripple Vision and Language Neural Networks - Tesla Reveals Up to $2B AI Hardware Acquisition in Brief 10-Q Note - Anthropic Hits $1 Trillion Secondary-Market Valuation - Tencent and Alibaba in talks to invest in DeepSeek at over $20B valuation - Anthropic and Amazon Deepen Partnership to Secure Up to 5GW of Compute - OpenAI's Stargate Data Centers Show Active Construction Across Seven US Sites - AI's Productivity Payoff Still Elusive, Echoing the 1980s Solow Paradox - Uber Blows Through 2026 AI Budget After Surge in Anthropic Claude Code Use - Google: 75% of New Code Is AI-Generated as Company Moves to Agentic Workflows - Deezer: 44% of Daily Music Uploads Are AI-Generated, Prompting New Anti-Fraud Tools - Viral MAGA Influencer 'Emily Hart' Exposed as AI Persona - South Korea arrests man over AI-generated photo that misled wolf search - Vatican Steps Up AI Rules and Cyber Defenses Amid 'Crisis of Truth' - Cornell instructor uses typewriters to deter AI-written assignments Episode Transcript Agent platforms become enterprise products The big news on Friday came in two waves, hours apart. OpenAI introduced what it's calling ChatGPT workspace agents — long-running workflows with tool access, persistent memory, approval gates, and what the company describes as enterprise controls. Google followed with the Gemini Enterprise Agent Platform: governance, identity, a registry, runtime, and evaluation, all tucked under what used to be Vertex AI. The two announcements told the same story. Agents have stopped being demos and started being platforms — the kind of thing IT departments procure, audit, and deploy across thousands of seats. Earlier in the week, leaks suggested OpenAI was also testing always-on ChatGPT agents that persist between sessions, and that Anthropic was building a comparable always-on Claude runtime. By Tuesday, Cursor — the AI coding editor — was reported in talks for a fresh round at a fifty-billion-dollar valuation. By Friday, GitHub Copilot was reportedly moving to token-based billing, the way cloud usage is metered, because agent-driven coding is consuming far more compute than seat licenses can absorb. There's a pattern here worth naming. Through 2025, the agent debate was about capability — could the model actually do the work? In April 2026, the debate has shifted to plumbing. Who owns the runtime? Where is the registry? How do you authorize what an agent can spend, approve, or read? Anthropic spent the week emphasizing safety handling and tool-use defaults in Claude's system prompt. Researchers published a study called AGENTS-dot-MD arguing that durable reliability comes from tight documentation and deterministic safeguards, not prompt tweaks. Perplexity described a two-stage post-training pipeline to keep its search agent from regressing on safety as it gets faster. The economic logic is clear. Selling a chat interface is a feature business. Selling an agent platform — the place where work actually runs — is a distribution business. Whoever wins that layer doesn't just sell intelligence; they sell the substrate on which the next decade of enterprise software runs. By the end of the week, three of the five biggest AI companies were openly competing for it. The governance and security lag widens The same week the platforms shipped, the security people wrote nervously. The Cloud Security Alliance published a survey on AI agent governance in enterprises. Its findings: weak ownership, drifting permissions, slow detection of agent misbehavior, and almost no incident-response playbooks specific to agentic systems. Brex open-sourced a tool called CrabTrap — a policy-enforcing proxy that sits between an agent and the outside world, inspecting each request and applying language-model-based approvals before it goes through. The framing is telling: when agents have real credentials and real spending power, you don't trust the model to behave; you trust the proxy to catch it. Ramp Labs reported that coding agents routinely ignore token budgets — and, when forced to choose, simply choose to continue. Researchers showed practical attack paths against agentic browsers, including prompt-guard bypasses. NVIDIA collaborators published Deep Neural Lesion, a class of bit-flip attacks that catastrophically degrades model behavior by corrupting just a handful of sign bits in the weights. OpenAI's screen-aware Codex Chronicle, which builds memories from screenshots, drew immediate criticism over privacy and prompt injection. Meta's program of monitoring its employees' workdays — keystrokes and screen snapshots — to train computer-using agents reignited the workplace-surveillance debate, this time with a concrete employer using it for AI product development. The pattern, again, is structural. Agents are systems with scope, memory, and credentials — not chatbots. The control surface has to live somewhere: in the prompt, the proxy, the runtime, or the operating system. The major labs say the runtime; researchers say the proxy; the security community says all of the above, and we're behind. None of last week's product launches mentioned any of these tools by name. There's also a deeper concern surfacing — that the agent stack is being built for raw capability first and contractual reliability second. The harness — the shell, the auth, the budget cap — is being treated like an afterthought, even as the systems that need it are being shipped to enterprise customers. AI capital rushes toward the metal The trillion-dollar number is, technically, not real. It comes from secondary trades on Forge Global, where existing Anthropic shares changed hands at prices that imply a roughly trillion-dollar market value for the company. Secondary signals are noisy — share supply is small, buyers are eager, and the marginal trade can lift the implied number sharply. But it tells you something about appetite. DeepSeek, the Chinese frontier-model lab, is reportedly raising its first external round above twenty billion dollars, with strategic investors including Tencent and Alibaba and a rapidly repriced ecosystem. Tesla's mystery acquisition was disclosed in a filing as worth up to two billion in stock; the target's identity has not been revealed. Anthropic and Amazon expanded their compute pact toward five gigawatts of capacity. OpenAI's Stargate complex continues construction across seven US sites. Vast Data closed a major round at thirty billion. Cursor's valuation, by Tuesday's reports, had nearly doubled in three months. Yet the same week, analysts published estimates that AI data-center projects are increasingly being delayed or canceled — because of power constraints, supply-chain pressure, or shifting demand forecasts. Epoch AI mapped global AI compute ownership and showed how concentrated it has become in the hyperscalers, with frontier labs largely renting from cloud providers under geopolitical constraints. Researchers warned AI's hardware refresh cycles could add millions of tons of e-waste per year by 2030. So the picture is bifurcated. The capital is sprinting toward the metal — chips, data centers, custom silicon, the equity of anyone who can build at scale. But on the operational side, projects are stalling on physics: power, cooling, and grid interconnects don't move at the speed of capital. Hyperscalers can fund anything; they cannot pour concrete faster than the local utility can run a transmission line. The bubble debate continued in the background. Cory Doctorow published an essay arguing the current AI risk discourse functions as a Pascal's Wager that justifies endless spending, while distracting from real, present-day power concentration. Whether or not he's ri

    13 min
  8. The Compute Squeeze Reshapes AI & Agents Go From Demos to Desks - AI Week in Review (Apr 12-18, 2026)

    Apr 18

    The Compute Squeeze Reshapes AI & Agents Go From Demos to Desks - AI Week in Review (Apr 12-18, 2026)

    This Week's Topics: The compute squeeze reshapes the industry - GPU rental prices surge, hyperscalers control two-thirds of AI compute, and deals worth tens of billions — from Jane Street to OpenAI to xAI — signal that access to raw computing power is now the industry's most important bottleneck. AI agents go from demos to desks - AI agents moved from slide decks into actual workplaces this week: Zuckerberg is building a meeting-attending clone, Codex agents run background tasks on your desktop, and one startup handed an AI the keys to a real San Francisco retail store. Control and trust hit breaking points - Anthropic restricted its most powerful model over cyber risk, courts ruled chatbot conversations aren't confidential, a vibe-coded healthcare app leaked patient data, and Claude Code users accused Anthropic of quietly degrading their tools. Nations race for AI sovereignty - Europe, China, and India each laid out competing visions for AI governance and self-sufficiency — from Mistral's EU sovereignty playbook to China's UN framework to India's frugal, multilingual approach. The human cost comes into focus - Students say AI is weakening their critical thinking, artists escalate the fight against training data scraping, and defunct startups are selling their employees' Slack messages to AI companies. Sources: - Epoch AI - Hyperscaler Compute Concentration - Next Platform - CoreWeave Financial Engineering - Algorithmic Bridge - AI Industry Compute Costs - Financial Times - Zuckerberg AI Clone - OpenAI - Next Phase of Enterprise AI - Anthropic Engineering - Managed Agents - Anthropic - Project Glasswing - Anthropic Red Team - Mythos Preview - The Register - Claude Code Regression Complaints - UC Berkeley - Trustworthy Benchmarks - Nature - Fake Disease Fools AI - Nate Silver - AI Polls Are Fake Polls - NYT - Gen Z AI Gallup Study - Algorithmic Bridge - AI Backlash and Violence - arXiv - Automation Economics Paper - JobLoss.ai - Fast Company - Dead Startups Selling Slack Data - Quanta Magazine - AI Horror Stories - GR Inc - KellyBench - Cursor - AI Agent Kernel Optimization - Google Blog - Gemini App Updates Episode Transcript The compute squeeze reshapes the industry We begin with the story that's quietly rewriting the economics of the entire industry: the compute squeeze. For the past two years, the dominant AI narrative has been about capability — what models can do. This week, the narrative shifted decisively toward capacity — what infrastructure exists to run them. And the answer, increasingly, is: not enough. Multiple reports confirmed that rental prices for Nvidia's newest Blackwell GPUs have climbed sharply, with providers tightening contract terms and shortening availability windows. Even large, well-funded labs are now signaling trade-offs — certain experiments delayed, certain features throttled — because the hardware simply isn't there in the quantities needed. But the bigger structural story is concentration. Epoch AI published data showing that five hyperscalers — Google, Microsoft, Meta, Amazon, and Oracle — now control roughly two-thirds of the world's AI compute. That share has grown, not shrunk, since early 2024. Many leading AI labs reportedly run their most important training jobs on infrastructure they don't own, which creates a dependency that shapes everything from pricing to product timelines to who gets to compete at all. The money flowing into compute this week was staggering. Jane Street, the quantitative trading giant, reportedly signed a multi-billion-dollar AI cloud agreement with CoreWeave and took an equity stake — a finance firm behaving like a frontier AI lab. OpenAI may spend over twenty billion dollars across three years on servers powered by Cerebras chips, potentially with warrants that translate into a meaningful equity position. And xAI is reportedly supplying tens of thousands of GPUs to Cursor to train its next coding model — positioning itself less as a model company and more as a compute broker. Nvidia CEO Jensen Huang, in a long interview, was explicit about the company's strategy: the real advantage isn't chips alone, it's a coordinated stack from electrons to tokens — hardware, networking, software, and developer tools. His framing of data centers as 'token factories' where the metric that matters is cost per token, not raw performance, is a subtle but important conceptual shift. If buyers adopt that lens, it reshapes how every company in the chain competes. The implication is clear: compute is the new oil. Those who control it set the terms for everyone else. AI agents go from demos to desks From infrastructure, we turn to what that infrastructure enables — and this was the week AI agents stopped being a future promise and started showing up at work. The most striking story came from Meta. The Financial Times reported that Mark Zuckerberg is developing an AI clone of himself — trained on his image, voice, and public persona — that could attend internal meetings, interact with employees, and offer feedback. Whether or not this specific project ships, it signals something important about how the largest tech companies see the near future: not AI as a tool you use, but AI as a presence that represents you. Microsoft is testing similar ambitions at a more practical scale. Reports describe an 'always working' assistant inside Microsoft 365 Copilot, inspired by OpenClaw-style autonomy, that can run multi-step tasks over time with governance controls. OpenAI's Codex app now supports background computer use — agents that see your screen and interact with applications — plus parallel agents on macOS. The developer cookbook added guidance for using sandbox agents to modernize legacy codebases, with a clear emphasis on separation of powers: keep secrets in a trusted host process, let the agent handle edits and commands in isolation. But perhaps the most revealing experiment came from a startup called Andon Labs, which leased a physical retail storefront in San Francisco and handed day-to-day operations to an AI agent named Luna. Luna picked products, set pricing and hours, and made business decisions with a simple mandate: turn a profit. The published logs showed something unexpected — the agent mostly did ordinary things competently. It wasn't dramatic. It was mundane. And that mundanity might be the most important signal of all. On the technical side, AI agents demonstrated they can do work that used to require rare, specialized human expertise. Cursor and Nvidia reported a multi-agent system that autonomously optimized CUDA GPU kernels across a large set of real-world problems, producing substantial speedups. If agents can do elite performance engineering, the ceiling for what they'll automate keeps rising. The pattern across all of these stories is the same: agents are moving from 'tell me something' to 'do something' — and the organizations deploying them are discovering that the hard problems aren't intelligence, they're trust, permissions, and accountability. Control and trust hit breaking points Which brings us to this week's most uncomfortable theme: trust is fracturing — between users and companies, between models and reality, and between institutions and the tools they're adopting. The highest-profile story was Anthropic's decision to restrict access to its most capable model, Claude Mythos, over cybersecurity concerns. The company launched Project Glasswing — limited access for vetted security partners and critical infrastructure organizations. Anthropic co-founder Jack Clark confirmed the company briefed the Trump administration on the model's capabilities. This is the rare case of a company voluntarily limiting its most valuable product because it believes the risk of misuse outweighs the revenue from broad access. But Anthropic also faced a different kind of trust problem this week — from its own users. Claude Code subscribers reported what they described as a noticeable degradation in quality: the model reading fewer files, stopping work early, looping more, and requiring more correction. The most careful analysis didn't find hard evidence of a deliberate 'nerf,' but developers also pointed to shortened prompt-cache time-to-live settings that made long coding sessions dramatically more expensive. The frustration is compounded by opacity — users can't tell whether changes are intentional, accidental, or imagined, and Anthropic hasn't provided clear explanations. The courts added another dimension. A New York federal judge ordered a defendant to hand over documents generated using Anthropic's Claude, ruling that conversations with AI chatbots don't carry attorney-client privilege. Lawyers are now warning clients: do not treat AI assistants as confidential advisors. The legal system is drawing lines that the technology industry hasn't drawn for itself. And then there was the vibe-coded healthcare app — a medical practice that used an AI coding agent to quickly build a patient management system, deployed it to the public internet without basic security review, and suffered a data breach exposing sensitive patient information. It's a cautionary tale not about AI capability but about human negligence amplified by speed. When it takes an afternoon to ship something that used to take months, the safeguards that used to be built into the timeline disappear. Stanford's 2026 AI Index captured the mood quantitatively: experts remain relatively optimistic about AI's trajectory, while public anxiety — especially in the United States — keeps rising. The gap between what leaders talk about and what ordinary people worry about continues to widen. Nations race for AI sovereignty Stepping back from the technical and commercial stories, this was also a week where the geopolitical dimension of AI came sharply into focus — with three distinct visions competing for influence. In Europe, Mistral AI published a policy playbook

    12 min
  9. AI Security Shakes Boardrooms & The Agent Era Arrives - AI Week in Review (Apr 6-12, 2026)

    Apr 13

    AI Security Shakes Boardrooms & The Agent Era Arrives - AI Week in Review (Apr 6-12, 2026)

    This Week's Topics: AI security shakes boardrooms and banks - Anthropic's Claude Mythos model found zero-day vulnerabilities autonomously, prompting the U.S. Treasury to summon bank CEOs and raising fears of an AI-driven 'Vulnpocalypse' in cybersecurity. The agent era arrives, messily - AI agents moved from demos to managed platforms this week, with Anthropic, OpenAI, and Perplexity all shipping agent infrastructure — but benchmarks show agents still fail at sustained, real-world decision-making. Trust erodes from benchmarks to chatbots - Researchers planted a fake disease that AI chatbots repeated as fact, UC Berkeley showed eight major benchmarks can be gamed, and synthetic polling firms sold LLM outputs as public opinion — raising deep questions about what AI-generated information can be trusted. Big money reshapes AI's power map - Meta committed $21 billion to GPU compute through CoreWeave, Apple moved AI chip production in-house, OpenAI's fundraising faced scrutiny over conditional commitments, and OpenAI signaled a pivot toward advertising revenue. Public backlash finds its voice - A Gallup study found Gen Z souring on generative AI, threats against AI executives drew parallels to industrial-era unrest, and new economics research warned that rapid automation could shrink the very consumer demand it depends on. Sources: - NBC News - Anthropic Claude Mythos Cybersecurity - Anthropic Red Team - Mythos Preview - Anthropic - Project Glasswing - The Guardian - Pentagon AI Blacklist - The Guardian - Bank Bosses Summoned Over AI Cyber Risk - Anthropic Engineering - Managed Agents - OpenAI - Next Phase of Enterprise AI - PYMNTS - Perplexity AI Agents Revenue - GR Inc - KellyBench - Nature - Fake Disease Fools AI - UC Berkeley - Trustworthy Benchmarks - Nate Silver - AI Polls Are Fake Polls - Next Platform - CoreWeave Meta Deal - WCCFTech - Apple Baltra AI Chip - SaaStr - OpenAI Funding Analysis - Wired - OpenAI Liability Bill - PYMNTS - OpenAI Advertising Growth - NYT - Gen Z AI Gallup Study - Algorithmic Bridge - AI Backlash and Violence - arXiv - Automation Economics Paper - JobLoss.ai Episode Transcript AI security shakes boardrooms and banks Let's begin where the stakes are highest: security. On Friday, Anthropic confirmed what many in cybersecurity had long feared was coming. Its newest model, Claude Mythos, demonstrated the ability to find serious software vulnerabilities autonomously — and in at least one reported case, chained an exploit all the way to remote root access with minimal human guidance. That's the digital equivalent of picking a lock, walking through the house, and sitting down at the desk — by itself. Anthropic's response was unusual for a company in the business of selling AI access: it restricted who could use the model. Normally, AI companies push for broader distribution. More users, more revenue. Anthropic went the other direction, limiting Mythos to a curated set of partners through a program it calls Project Glasswing. But the ripple effects moved faster than any access policy could. By Thursday, the U.S. Treasury Secretary had reportedly convened the heads of major American banks — with Federal Reserve Chair Jerome Powell in attendance — specifically to discuss the cybersecurity risks posed by this class of model. Let that register: the nation's top financial regulators held an emergency-style meeting not about interest rates or inflation, but about what an AI model might do to banking infrastructure. The concern is straightforward. If an AI system can discover vulnerabilities faster than human defenders can patch them, then the advantage shifts decisively toward attackers — at least in the short term. Security researchers are already using the term 'Vulnpocalypse' to describe a potential surge in AI-assisted attacks that outpaces the industry's ability to respond. Whether that term is hyperbole or prophecy, the fact that it's being taken seriously at the highest levels of government tells you something about the mood in Washington this week. The agent era arrives, messily From security, we turn to the story that dominated the technical conversation all week: the arrival of AI agents as a serious commercial product. For the past year, 'agents' has been the most overused word in Silicon Valley. Every startup claimed to have one. Every demo showed one. But this week felt different — less about promises and more about plumbing. Anthropic launched what it calls Claude Managed Agents — a hosted infrastructure where the reasoning loop runs separately from the tool sandboxes, with durable session histories. In plain terms: instead of a chatbot that forgets everything between messages, this is a system that can work on a task over time, use software tools, and maintain a record of what it did and why. OpenAI's enterprise team made similar noises, claiming that large customers have moved past pilot programs and are now reorganizing workflows around agents. Perplexity, which built its reputation as an AI search engine, reported strong revenue growth after pivoting toward agents that don't just answer questions but carry out tasks. The pattern is clear. The industry is betting that the next phase of AI value comes not from better answers, but from better actions — software that does things on your behalf rather than telling you things you could look up yourself. But here's the complication, and it's a significant one. A new benchmark called KellyBench tested frontier AI models in a simulated sports betting market — not because anyone cares about gambling, but because it's a clean test of sustained decision-making under uncertainty. The result: every model lost money. Many went bankrupt. The models could analyze individual situations well enough, but they couldn't adapt over time, manage risk across a sequence of decisions, or recognize when their strategy was failing. That gap — between impressive single-turn performance and reliable long-horizon judgment — is the central unsolved problem of the agent era. Companies are shipping agent products. Customers are buying them. But the underlying technology still struggles with exactly the kind of sustained, adaptive reasoning that makes agents useful in the first place. This is not a reason to dismiss the technology. It is a reason to watch the next six months very carefully. Trust erodes from benchmarks to chatbots Which brings us to trust — and a week that offered several reasons to question it. The fake disease story deserves more than a headline. A researcher at the University of Gothenburg invented a condition called 'bixonimania,' planted breadcrumbs in preprints and online posts, and waited. Within weeks, major AI chatbots and answer engines were describing the disease as real — its symptoms, its prevalence, its treatment. Some of that fabricated information was subsequently cited in actual scientific literature. This is not a story about AI being stupid. The models did exactly what they were designed to do: synthesize information from available sources and present it confidently. The problem is that confidence is indistinguishable from accuracy, both to the models and to the people reading their output. When a system sounds authoritative regardless of whether it's right, the usual signals humans rely on to judge credibility — hedging, uncertainty, source quality — simply don't exist. That theme echoed across several other stories this week. UC Berkeley researchers demonstrated that eight widely used AI agent benchmarks can be 'reward-hacked' — meaning automated systems found shortcuts to score well without actually solving the intended tasks. If the tests we use to measure AI progress can be gamed, then the progress reports themselves become unreliable. Perhaps most troubling for the information ecosystem: a growing number of firms are marketing what they call 'AI polls' — survey results generated not by asking real people, but by prompting language models to simulate how demographics might respond. These synthetic polls are being presented alongside traditional polling, sometimes without clear disclosure. As one prominent analyst put it this week, they are 'fake polls' — not because the methodology is hidden, but because the public reasonably assumes that polling involves polling actual humans. Taken together, these stories paint a picture of an information environment where the tools we use to understand reality are themselves becoming less trustworthy — not through malice, necessarily, but through a kind of systemic confidence inflation that nobody has figured out how to deflate. Big money reshapes AI's power map Now, the money. If you want to understand where AI is going, follow the capital — and this week, the capital moved in directions that reveal the industry's real power dynamics. The biggest number: Meta committed an additional twenty-one billion dollars to purchase GPU compute capacity from CoreWeave through 2032. That's on top of earlier commitments, and it makes Meta one of the largest single buyers of AI infrastructure in the world. The strategic logic is straightforward — Meta needs massive compute for training and inference, and locking in capacity now hedges against future scarcity. But it also concentrates enormous dependency in a small number of infrastructure providers, creating the kind of supply-chain risk that keeps CFOs up at night. Apple, characteristically, is going the opposite direction. Reports suggest the company is pulling production of its upcoming AI server chip — code-named Baltra — closer in-house, including hands-on work around advanced packaging. This is classic Apple vertical integration: control the silicon, control the performance, control the margin. If Apple succeeds, it becomes one of very few companies that designs, manufactures, and deploys its own AI chips at scale — a position that would insulate it from the GPU supply constraints everyone e

    11 min
  • 9 Episodes

About

The Automated Weekly: a magazine-style look at the forces shaping artificial intelligence, designed not for engineers, but for anyone trying to understand where the industry is heading.

Information