Anthropic's "When AI Builds Itself" and OpenAI's democratic-governance blueprint land in the same week — two labs telling us, in their own words, that AI building AI is now the planning assumption, that human judgment is the narrowing bottleneck, and that no coordination mechanism exists for what comes next. Before that in the headlines: Washington is reportedly in talks to take equity in the AI labs themselves.
The labs just told us what they actually believe: AI building AI is the planning assumption now.
Anthropic's "When AI Builds Itself" and OpenAI's "Democratic Governance of Frontier AI" start from the same place — recursive self-improvement is no longer a thought experiment but something both labs see signs of in today's systems, arriving "sooner than most institutions are prepared for." Anthropic ships 8X the code with Claude writing 80% of its own production codebase; humans retreat to taste and judgment; and the only honest admission in either document is that nobody has a coordination mechanism for any of it.
The core of "When AI Builds Itself": Anthropic is delegating a growing share of AI development to AI systems, and taken far enough, that trend points to an AI capable of fully autonomously designing its own successor. "We're not there yet and recursive self-improvement is not inevitable" — but the inflection point is coming fast.
Anthropic engineers now ship eight times as much code per quarter as they did from 2021–2025, and 80% of Claude's production code is authored by Claude itself. Session success rates have climbed well above 60% even on open-ended problems — and above 80% on everything else — from much lower less than a year ago.
Once human- and AI-authored code quality reach parity, humans stop writing and only review — and if they can't review as fast as Claude generates, review becomes the bottleneck to AI development. The remaining comparative advantage: choosing which problems matter, which results to trust, and when an approach is a dead end.
On Anthropic's stall scenario: the bottleneck to AI progress could be energy, chip fabrication, grid expansion, and interconnect bandwidth rather than the intelligence itself — and we're seeing lots of evidence of exactly that right now.
Anthropic's example: Mythos Preview found more than 10,000 high- and critical-severity vulnerabilities across many of the world's most important software systems. But they don't think the stall scenario is likely — "every capability we can measure has so far followed the same curve. We've not yet seen that curve bend."
In the most likely scenario, 100-person companies do the work of 10,000 or 100,000 — but speeding up one part of a process just shifts the bottleneck. Anthropic is living it: human code review is the new chokepoint, and ideas now outrun capacity to pursue them. Spotting and fixing bottlenecks "may become the most important skill for any organization."
— Aaron Levie, Box CEO, on the bottleneck admission in Anthropic's piece. AI lowers the barrier so dramatically that we have far more ideas than we can pursue — more software, more campaigns, more drugs — and the surrounding work to execute still requires people to manage, even when augmented by agents. The infinite-backlog case for why the jobs don't just vanish.
It would "likely be a good thing" for the world to have the option to slow or temporarily pause frontier AI development — but a slowdown that lets the least cautious actors catch up "could leave everyone less safe." A unilateral pause is achievable immediately but "accomplishes much less"; Anthropic says it will convene policymakers, researchers, civil society, and other AI companies in the coming months.
— Corey Quinn, on Anthropic's pause talk. The responses run the gamut: safety advocates are thrilled, Nate Soares thinks they aren't thinking big enough, and Sean Ralston calls it "an insincere and silly sentiment — if they really feel that way then let them act that way."
— Bill Gurley, on All-In, after 30 days reading everything Anthropic has written. Gurley: "I don't know which one I'm more afraid of — the regulatory capture or this Dr. Frankenstein theory." David Sacks piled on: compare it to nukes, warn RSI could end humanity, then race ahead anyway — "you want the government to save us from you."
"Democratic Governance of Frontier AI" starts from the same place as Anthropic: "We also see signs of recursive self-improvement in today's systems, where AI development is itself accelerated by AI" — creating governance challenges "that existing institutions are not equipped to address." Chubby's reaction: "The vibe has changed. Something is happening."
Three policy directions: Congress should adopt and scale the best state regulations rather than preempt them; testing should live in civilian institutions like CAISI rather than the NSA (contra the recent executive order), eventually becoming mandatory; and frontier AI should get a whole-of-government resilience strategy. Dean Ball: civilian, non-classified testing is how you avoid a de facto licensing regime.
The Obernolte–Trahan framework would override state AI laws and require leading labs to implement catastrophic-risk plans verified by third-party auditors. On substance it's less dead-on-arrival than most — but Trahan is taking fire from Northeast Democrats over preemption, House GOP leadership is skeptical, and Speaker Johnson won't commit to a floor vote before midterms. In other words: don't hold your breath.
Notice reports senior US officials have held discussions with major AI companies about the federal government acquiring shares — with Sam Altman said to have pitched the idea directly to the president in early 2025. Talks reportedly center on labs "voluntarily ceding shares" whose returns could fund an AI dividend check to American households. Anthropic is not involved at this time.
— Steve Bannon, responding to the Sanders 50%-stake proposal. The populist right is aligning with Bernie Sanders on forcing the labs to "cough up 50% of the equity to be dispersed to American citizens." The horseshoe theory of American politics is well and truly intact.
— Dan Primack, Axios business editor, on the equity-stake talks. Much of the public reaction asked why taxation isn't the tool instead: Georgetown Law's Peter Harrell warned that ownership "risks giving the government control outside of public view and potentially the wrong incentives."
Individual saved memories are gone, replaced by a continuously maintained, user-editable summary of richer context. On OpenAI's new recall benchmark, success jumped from 41.5% (the 2024 saved-list version) to 67.9% last year to 82.8% now — and a 5X compute reduction means free users get Dreaming for the first time.
— Mark Kretschmann, on ChatGPT's Dreaming update. The more ChatGPT becomes an actual work partner, the less sense it makes to restart from zero every time — projects, preferences, constraints, writing styles, and code-base details should all carry forward. Sounds small, but it changes the product.
Every turn spent getting a model to re-remember context is wasted tokens — exactly the theme of Glean CEO Arvind Jain's "Your Token Spend Is an AI Architecture Problem." A small update on the surface, but one with bigger implications for how companies adapt to the token scarcity era.
CEO C.C. Wei at the annual shareholders meeting: "Customer demand is so high and we can only support so much... We're doing our best to ensure TSMC does not become a bottleneck." Arizona plans have expanded to six fabs, but permitting and construction-worker shortages have the US buildout behind schedule, and "it will be a long time before we can meet customer demand."
— C.C. Wei, TSMC CEO, on memory-chip suppliers' price hikes. Wei would like to raise prices given the shortage but wants to avoid abrupt memory-style hikes. TSMC remains famously relationship-based — Jensen Huang says he's never signed a contract despite being its largest customer.
Bloomberg reports the Airbnb CEO's unnamed startup, focused on user interaction and design, is in early fundraising — and Chesky will hire a leader rather than go founder mode. Saxum's read: instead of the nth coding-benchmark lab, "huge alpha in just having a model that makes great UI/UX experiences."
A checkpoint codenamed Oceanus went to red teamers, which typically precedes wider launch by about seven days, and a dug-up API endpoint shows $16 per million input tokens and $80 per million output — roughly 3× Opus but below Mythos Preview's $25/$125. Andrew Curran is predicting public release around the 16th.
If OpenAI releases 5.6 before Mythos lands, it's a preemption, not a response to Opus 4.8 — implying they don't think it can hang with Mythos. If they believed it was better, they'd wait until right after Mythos shipped to clip Anthropic's momentum. The release sequence will tell us how the labs see the state of the art.