News Digest

Key Highlights Anthropic put a clock on recursive self-improvement. In a report titled “When AI builds itself,” the company disclosed that more than 80% of its merged code was authored by Claude as of May, with an 8x rise in daily code submissions since 2024 — and signaled it would slow frontier research if rival labs did the same. OpenAI flagged comparable early-RSI indicators in its own governance framework. OWASP rewrote its Top 10 for the vibe-coding era, shifting from “outdated components” to a broader software supply chain focus and adding two new awareness items: memory safety and vibe-coding — a formal acknowledgment that AI-generated code is reshaping the application-security threat model. Anthropic is heading for an IPO as co-founder Daniela Amodei pointed to the enormous upfront cost of training models; annualized revenue hit $47B in May, up from ~$9B at the end of 2025 — a striking counter to the week’s enterprise-ROI skepticism. The infrastructure land-grab intensified: AirTrunk committed $30B to build 5GW of AI data centers in India by 2030, while Meta began housing AI chips in weatherproof tents in Ohio to compress build timelines — “the AI race has officially entered its Mad Max phase.” Apple’s WWDC lands Monday with a long-awaited Siri revamp reportedly powered by Google’s Gemini, App Store AI-agent integration, and natural-language photo editing — and Apple just approved Poke as the first AI agent on its Messages for Business platform. Analysis & Opinion Anthropic Confronts the RSI Clock — The Rundown Anthropic’s “When AI builds itself” report makes the recursive-self-improvement (RSI) debate concrete: over 80% of the company’s merged code is now written by Claude, daily code submissions are up 8x since 2024, and the authors warn of a trajectory where “each new version of Claude could be built by the version before it, without human involvement.” Co-author Jack Clark frames this as a near-term governance problem, not a sci-fi one, and OpenAI simultaneously flagged comparable early-RSI signals in its own framework. The most notable proposal: Anthropic says it would be willing to decelerate frontier research if competing labs adopt the same measure — an explicit nod to the coordination problem at the heart of any AI “pause.” The risk it highlights is structural: multiple labs (including MiniMax’s M2.7 and newer startups) report self-improvement capabilities, so even a well-intentioned unilateral slowdown does little without industry-wide buy-in. Anthropic stresses RSI hasn’t materialized and remains uncertain, but the report reads as an attempt to set the terms of debate before the capability arrives. Expect upcoming policy discussions on methodology, systems architecture, and slowdown scenarios. ...

Key Highlights Failing grades surged in UC Berkeley CS courses as instructors point to AI-driven academic dishonesty — 35.3% of CS 10 students and 10.6% of CS 61A students received F’s this spring, versus under 10% in prior years — a stark data point in the debate over how generative AI is reshaping (and eroding) foundational learning. The U.K. forced Google to let publishers opt out of AI Search, with the CMA calling the Search Console toggle a “world first” — reframing yesterday’s Google “publisher controls” announcement as regulatory compliance rather than voluntary goodwill. Alphabet raised a record-breaking $85B for Google’s AI business — an oversubscribed offering that topped Petrobras’s 2010 record — signaling investors’ near-insatiable appetite for AI infrastructure exposure. NVIDIA shipped Nemotron 3 Ultra, a 550B-parameter (55B active) open MoE model purpose-built to orchestrate long-running agents at lower cost and reduced goal drift. Theo and Sean Goedecke argue your prompts are tech debt — AGENTS.md/CLAUDE.md files decay silently with every model upgrade, making a January-tuned prompt actively harmful by February. Analysis & Opinion Failing Grades Soar With AI Usage, Dwindling Math Skills in Berkeley CS Classes — Daily Californian The share of failing grades in several UC Berkeley computer science courses spiked far above historical norms in spring 2026, breaking the department’s own grading guidelines: 35.3% of CS 10 students and 10.6% of CS 61A students received F’s, against a guideline of roughly 7% D’s-and-F’s and a historical ceiling under 10%. Teaching professor Dan Garcia attributes the “primary driver” to a “vast increase in academic dishonesty” tied to students leaning on LLMs, compounded by weaker mathematical preparation and understaffing. The episode is a concrete signal of a tension educators have warned about: when AI can produce passing work, students may skip the struggle that builds durable skill, then fail when assessments demand genuine understanding. It also raises hard policy questions — whether to redesign assessment around in-person or oral exams, how to detect misuse fairly, and whether grading curves should even hold in an AI-saturated classroom. Expect this to become a recurring data point as more institutions report semester outcomes. ...

Key Highlights Uber capped employee AI tool spending at $1,500/month after burning through its entire annual AI budget in four months — the bluntest signal yet that enterprise AI ROI remains hard to prove, with Uber’s own COO admitting it’s “very hard to draw a line” between tool usage and business outcomes. Microsoft made a major move toward AI independence at Build 2026, unveiling seven proprietary MAI models, a “Scout” Autopilot agent in Teams, and an AI-designed quantum chip (Majorana 2) — loosening its reliance on OpenAI. NVIDIA and Microsoft unified the agentic AI stack from Windows devices to cloud — RTX Spark and DGX Station hardware, the NemoClaw agent blueprint, and the OpenShell security runtime — with Anthropic’s Claude models now running natively on Blackwell systems in Azure. Google gave website owners a Search Console toggle to control whether their content appears in generative AI Search, as AI Overviews reaches 2.5 billion monthly users and AI Mode passes one billion — a meaningful concession on publisher control. Cursor shared a year of lessons building cloud agents, which now generate 40% of its internal pull requests and process 50M+ daily actions after a migration to Temporal for reliability. Analysis & Opinion Uber Caps Employee AI Spending After Blowing Through Its Budget — TechCrunch Uber has imposed a $1,500 monthly per-employee cap on AI tools including Claude Code and Cursor, tracked through an internal usage dashboard, after its CTO disclosed in April that the company had exhausted its entire annual AI budget in just four months. The reversal is striking because Uber had previously urged staff to “use AI as much as possible,” even running internal leaderboards to gamify consumption. COO Andrew Macdonald conceded it remains “very hard to draw a line” between AI usage and tangible business outcomes. The episode crystallizes a sector-wide anxiety: enterprises are spending heavily on AI tooling while ROI stays largely theoretical, and unmetered “use it all you can” policies collide quickly with real budgets. Expect more companies to move from encouragement to metering as finance teams demand accountability. ...

Key Highlights Florida’s attorney general sued OpenAI and Sam Altman in a first-of-its-kind state action, alleging ChatGPT was linked to violent incidents and that the company ignored safety warnings while racing to win the “AI arms race.” NVIDIA’s GTC Taipei keynote declared “agentic AI has arrived” — Jensen Huang unveiled Vera Rubin (in full production), the Vera “CPU for agents,” RTX Spark AI PCs with Microsoft, and Cosmos 3 for physical AI. Coverage from TechCrunch, The Rundown, and NVIDIA’s own blogs all converge on the same agent-centric pivot. Two sharp takes on what AI does to engineering careers: Theo (t3.gg) argues AI raises the floor for weak engineers but will widen the gap and crush the unmotivated bottom 30%, while Jensen Huang insists AI is increasing software hiring (GitHub commits nearly tripled in early 2026). Business milestones: Anthropic confidentially filed to go public, and Alphabet plans to raise $80B (including $10B in stock to Berkshire Hathaway) to fund its AI buildout. A new coding benchmark (DeepSWE) exposes how contaminated, badly-prompted benchmarks like SWE-Bench Pro misled model comparisons — and shows a far larger gap between frontier and open-weight models than older benches suggested. Analysis & Opinion Florida Sues OpenAI, Sam Altman in First-of-Its-Kind Lawsuit — TechCrunch Florida AG James Uthmeier filed an 83-page complaint against OpenAI and CEO Sam Altman, alleging ChatGPT has been linked to multiple violent incidents in the state. The suit claims defendants prioritized winning “the AI arms race and amass[ing] large fortunes” while ignoring internal and external safety warnings and putting children at risk. It specifically alleges the chatbot “aided and abetted” mass shooters and “encouraged” vulnerable people toward suicide. As the first state-led action of its kind, it could set a template for how attorneys general pursue AI product-liability and child-safety claims — a meaningful escalation of regulatory risk for frontier labs. ...

Key Highlights A widely-shared take argues teams should stop blindly committing auto-generated AGENTS.md files from /init — treating them as a living list of unfixed codebase smells, scoped hierarchically per module, rather than a monolithic root-level config. A tinkerer fit a 2017-era datacenter GPU (Tesla V100) into a gaming PC for ~£200, reaching 32GB of VRAM and running a 27B-parameter model at 32 tokens/sec — a reminder that older server silicon can still beat consumer cards on memory bandwidth for local inference. Quiet day across the major labs: no new posts from OpenAI, Anthropic, Google, or NVIDIA since the I/O 2026 wave earlier in the week. Analysis & Opinion Stop Using /init for AGENTS.md — Addy Osmani Osmani argues the common ritual of running /init, accepting the auto-generated AGENTS.md, and committing it unscrutinized may actually hurt agent performance. His fix: treat the file as a living list of codebase smells you haven’t fixed yet, and use hierarchical, module-scoped context files so agents get precisely-scoped information instead of one project-wide document. He notes the research is genuinely mixed — two 2026 studies reach opposite conclusions on whether context files help or just add token overhead. ...

Key Highlights Jensen Huang pushes back on AI layoffs. In a wide-ranging CNA interview, Nvidia’s CEO calls the AI-job-loss narrative “lazy” and “irresponsible,” arguing there will be more jobs in five years, not fewer, and framing AI as a “five-layer cake” (energy → chips → infrastructure → models → applications) that reinvents every industry. He also addresses US–China competition, dual-use risk, and the case for cooperation over decoupling. Anthropic ships Claude Opus 4.8 — and the reviews are in. The Rundown reports Anthropic has eclipsed OpenAI on valuation and benchmark performance, while Theo’s hands-on review calls it a “modest but tangible” improvement that tops coding benchmarks but burns tokens aggressively through the new Ultra Code / dynamic-workflows feature (one prompt maxed out a $100/mo tier in under 30 minutes). Google I/O 2026 lands. Google unveiled Gemini Omni (generate video from any mix of image/audio/video/text), Gemini 3.5 Flash (claimed to beat 3.1 Pro on most benchmarks at 4× speed), and an expanded Antigravity agent ecosystem — Antigravity 2.0 desktop app, a CLI, an SDK, and Managed Agents in the Gemini API. Inference efficiency is the new race. Kog AI’s engine hits 3,000 output tokens/s per request on standard datacenter GPUs via a persistent “monokernel,” targeting single-request latency for agentic workflows, while the open-source tiny-vLLM ships an educational C++/CUDA inference engine for Llama 3.2. Analysis & Opinion Anthropic just eclipsed OpenAI — The Rundown Anthropic released Claude Opus 4.8, which tops nearly all major benchmarks (agentic coding, financial analysis) and pairs the launch with a $65B raise that reportedly pushes its valuation past OpenAI’s. The model holds pricing flat versus its predecessor while improving honesty and reducing hallucinations, and Anthropic teased a forthcoming “Mythos-class” model. The piece frames Anthropic’s safety-first positioning as now paying clear commercial dividends. ...

Key Highlights Anthropic eclipses OpenAI on two fronts at once: a $65B Series H at a ~$965B post-money valuation (likely its last private round before an IPO), landing the same week it shipped Claude Opus 4.8, which beats GPT-5.5 and Gemini 3.1 Pro on agentic coding and financial-analysis benchmarks at unchanged pricing. The AI jobs apocalypse gets walked back: both Sam Altman (“pretty wrong”) and Dario Amodei are softening their earlier predictions of mass white-collar displacement, reframing AI as productivity-expanding — just as both companies head toward ~$1T IPOs. The “AI sticker shock” reckoning deepens: Glean now pitches cost reduction as its primary selling point on its way past $300M ARR, while enterprises scrutinize ROI and one client reportedly burned $500M in a month with no usage controls. Identity and security are the next agentic bottleneck: 1Password’s CTO warns that today’s identity standards break down when “ephemeral agent swarms” make attribution to a single user impossible — supply-chain and credential management are becoming the hard part of agentic deployment. Infrastructure is being rebuilt for machines, not people: AI agents that spawn sub-agents and vanish are driving serverless redesigns (AWS OpenSearch), and a chip startup (XCENA) just raised $135M betting the real bottleneck is memory, not compute. Analysis & Opinion Anthropic just eclipsed OpenAI — The Rundown Anthropic surpassed OpenAI in both valuation and headline model performance this week. Claude Opus 4.8 outperforms GPT-5.5 and Gemini 3.1 Pro across agentic coding and financial analysis at the same price as its predecessor, with improved honesty and reduced fabrication. The company simultaneously closed a $65B round at a ~$965B valuation — exceeding OpenAI’s — and teased a more advanced system, “Mythos,” within weeks. OpenAI leadership dismissed the safety-first positioning as “fear-based marketing,” but investors and users appear to be validating the approach. Opus 4.8 also adds a faster, cheaper mode and parallel sub-agents in Claude Code for long-running tasks. (See the companion funding story in Products below.) ...

Key Highlights Pope Leo XIV’s first encyclical Magnifica Humanitas makes the Catholic Church a moral authority on AI, demanding human-friendly systems, banning algorithmic lethal decisions, and warning that “a moral AI means nothing if that morality is determined by a few” — Dario Amodei echoed the same theme in his Oprah interview, arguing trust is in short supply and Anthropic refused Pentagon contracts allowing autonomous weapons or domestic mass surveillance even at risk of company-ending consequences. Anthropic is reportedly headed for its first profitable quarter at ~$10.9B in Q2 revenue, driven by Opus 4.5’s enterprise traction, AWS hosting reach, a stealth price hike via re-tiered model names (Opus 4.5 replacing Sonnet pricing-wise), and a 30-50% more verbose tokenizer in 4.7 — all while enterprise sales contracts have shifted from seat-based to API-priced billing. Cognition raised $1B at a $25B pre-money valuation ($492M ARR, 50% MoM enterprise growth) and Cursor announced a 10x compute scale-up with xAI/Colossus 2 for a from-scratch model, signaling that independent AI-coding labs may leapfrog the frontier labs rather than be absorbed by them. NVIDIA reported 85% revenue growth and 92% data-center revenue growth; Jensen Huang argued AI is creating jobs, called the AI-causes-layoffs narrative “too lazy,” and Anthropic published an essay describing how it now routinely deploys agents with substantial permissions after finding humans approve ~93% of permission requests with declining attention. Recursive self-improvement (RSI) is emerging as the new AGI fixation with Richard Socher’s Recursive Superintelligence, Karpathy’s Auto-Research (now folded into Anthropic), and Adaption’s AutoScientist — even as enterprises increasingly kill AI deals over operational instability rather than model quality, per Databricks. Analysis & Opinion How we contain Claude across products — Anthropic Anthropic’s engineering team describes a shift from refusing meaningful agent permissions to routinely deploying agents with substantial system access. The thesis: as capability grows, the cost-benefit calculation flips when robust safeguards are in place. The piece reveals a striking empirical finding — users approved roughly 93% of permission requests, with attention declining over repeated prompts, undermining human oversight as a primary safety layer. The team now leans on environmental containment (sandboxes, VMs, network controls) as the hard boundary, treating human approval workflows as a soft layer that degrades with use. The framing is honest about safety theater and points toward a future where capability headroom is paid for in containment infrastructure rather than reviewer vigilance. ...

Key Highlights Karpathy lands at Anthropic to lead a recursive self-improvement pre-training team — and All-In’s panel argues continual learning + recursive self-improvement are the two “final frontiers” that could pull the timeline forward sharply. Chamath’s framing: an order-of-magnitude per-year quality jump may turn out to be the conservative case once these two unlock. SpaceX’s S-1 reveals “Elon Web Services” is the real story, not Starlink. Anthropic is paying SpaceX $1.25B/month — a $45B / 3-year deal — to rent Colossus 1 and parts of Colossus 2 (with 90-day cancellation for either side). Composer 2.5 (Cursor’s model, trained on Colossus 2 in ~3 weeks of RL) now sits Pareto-dominant on the coding frontier; the panel reads it as proof that Cursor’s coding-token corpus + XAI compute is a real new pole in the model race. The “America turns on AI” thread is hardening into a policy problem. Three commencement speeches (Eric Schmidt’s included) got booed for AI-job framing; a planned Trump AI executive order was scrubbed at the last minute after the attendee list (frontier-lab CEOs + hyperscalers) leaked; Chamath and Gavin Baker both flag a coordinated anti-data-center / anti-AI sentiment campaign and call on the industry to lead with end-user benefit stories rather than CEO doom takes. Quiet day on the blog side — every curated source’s latest content predates the 2026-05-23 cutoff and was already captured in earlier digests. No new written posts to summarize. Interviews & Conversations SpaceX’s $2T Case, Nvidia’s Shock Selloff, America Turns on AI, Trump Pulls AI Order, Bond Crisis? — All-In Podcast (1:42:00) Episode 274 with guest Gavin Baker (Treaties Management). The week’s hinge story for the panel is Andrej Karpathy joining Anthropic to lead a new pre-training team focused on recursive self-improvement. Baker argues recursive self-improvement + continual learning are AI’s “two final frontiers” — and that if they unlock, Chamath’s repeated “10x per year” line “might seem conservative.” Anthropic was EBIT-positive last quarter per the WSJ; combined LLM-app ARR across OpenAI, Anthropic, Gemini, Cursor, XAI, and open source is on a path to $200–400B by year-end at ~80% gross margins on inference, which the panel treats as the end of the “circular funding” critique. ...

Key Highlights Jensen Huang’s post-earnings victory lap doubled as the clearest pushback yet against AI-doomer framing. Coming off an 85% revenue jump and 92% data center growth, Huang told Fox Business that telling young graduates AI will erase their jobs is “a disservice to society” — the actual risk, he argues, is losing your job not to AI but to “someone who is an expert in using AI.” The China-chip question is now framed by Nvidia as a capacity argument, not a security one. Huang’s line: “China obviously has all the chips they need. That’s the reason why they don’t need ours.” Huawei had a “record year” and is exporting its stack. The implication for US policy is that export controls aimed at slowing Chinese AI are running into a domestic-supply ceiling that’s already been built. Quiet news cycle on the blog side — every curated source’s latest post is dated 2026-05-22 or earlier, all of which were captured in yesterday’s digest. The I/O 2026 / Computex / Nvidia-earnings news wave has crested for now. Interviews & Conversations ‘DISSERVICE TO SOCIETY’: Nvidia CEO PUSHES BACK on AI ‘doomers,’ says tech creates jobs — Fox Business (19:28) Part two of Maria Bartiromo’s interview with Jensen Huang, taped the day after Nvidia’s record quarter (85% revenue growth, 92% data center growth, $80B buyback). Huang’s frame for the entire AI stack is a “five-layer cake” — energy, chips, infrastructure, models, applications — and his core policy ask is that the US lead at every layer rather than narrowly defending one. ...