What OpenAI and Anthropic Think Happens Next With AI

June 5, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
June 05, 2026

HEADLINES

US Government Eyes Equity Stakes in AI Labs

Bombshell reporting from NOTUS claims senior US officials have held preliminary discussions with major AI companies about the federal government acquiring shares in their firms. The reporting is sourced to three people familiar with the matter, and is anchored by one of Washington's most credentialed journalists — former Washington Post reporter Jeff Stein, whose DOGE reporting contributed to a Pulitzer. Sam Altman is said to have pitched the idea directly to the president in early 2025, framing it as a way to broadly distribute the economic benefits of AI to the public. The current framing centers on labs "voluntarily ceding" shares to the government, with returns potentially directed toward an AI dividend for American households — though it's unclear whether the government would pay for the shares. Notably, Anthropic is not part of these discussions. The idea has made for strange political bedfellows: Bernie Sanders proposed the government take a 50% stake via a one-time tax earlier this week, and Steve Bannon responded approvingly, calling for the same — while critics from the center and right largely argued the correct mechanism is taxation, not ownership.

OpenAI Ships "Dreaming" — A Major Memory Upgrade

OpenAI has shipped a significant update to ChatGPT's memory system, which they're calling Dreaming. The clunky list-based memory of two years ago is now gone — replaced by a dynamic summary the system maintains automatically in the background, which users can review and edit. On OpenAI's own benchmarks, the new system succeeds on 82.8% of tasks requiring factual recall, up from 41.5% with the 2024 memory system. Critically, efficiency gains have cut the compute requirements by 5x, meaning Dreaming is now available to free users for the first time. If you've built memory workflows through something like our ClawCamp or Agent OS programs, this will look familiar — it's essentially automated markdown file management. The big unlock is that ChatGPT is now doing this for everyone in the background, which starts to close the gap between power users who've built out memory infrastructure and everyone else.

TSMC: Chip Shortage Will Last All Decade

In candid remarks at TSMC's annual shareholder meeting, CEO C.C. Wei said demand is so high that the company is doing everything it can just to avoid becoming a supply chain bottleneck — and that it will be "a long time" before they can fully meet customer demand. TSMC has committed to six new fabs in Arizona, though Wei acknowledged environmental permitting and construction worker shortages have caused delays, even as he noted Arizona progress is running better than originally expected. When asked about price hikes, Wei said he'd like to raise prices but won't follow the memory chip playbook of abrupt 80% margin grabs — even though, as he admitted, he envies those margins.

Airbnb CEO Brian Chesky Is Starting an AI Lab

Brian Chesky is planning to launch a new AI venture focused on user interaction and design — which many are reading as an agent lab rather than another foundation model play. Despite his prominence in Silicon Valley (he was a key broker in Sam Altman's return to OpenAI in 2023), Chesky has mostly been on the periphery of the AI boom. He plans to stay on as Airbnb CEO and hire someone else to lead the lab, which is currently in early fundraising. The jokes write themselves — but the more interesting read is that a UX-focused lab could go after something most frontier labs have largely ignored.

GPT-5.6 and Mythos: Watch the Timing

X is buzzing with rumors that both GPT-5.6 and a public Mythos release are imminent. On the Anthropic side, leaker Leo at Synthwaved reports that a checkpoint codenamed Oceanus was made available to red teamers, with Andrew Curran predicting a June 16th public release. A separate API endpoint was found pricing the model at $16 per million input and $80 per million output — roughly 3x Opus 4.8 but below Mythos Preview pricing. GPT-5.6 didn't arrive this week as some expected, which is itself a signal worth watching. If OpenAI releases before Mythos drops, it's a preemptive move — suggesting they think 5.6 can't match Mythos head-to-head. If they wait until after, that tells the opposite story. The timing will say more than the benchmarks.

MAIN STORY

What OpenAI and Anthropic Think Happens Next with AI

Two major pieces of writing dropped this week — one from Anthropic, one from OpenAI — and together they offer the clearest window yet into how the leading labs see the current moment. Both start from the same premise: recursive self-improvement is either here or coming fast, and the world is not ready. Anthropic's piece ("When AI Builds Itself") reads as a meditation; OpenAI's ("Democratic Governance of Frontier AI") is a policy document. But both are revealing precisely because they're not talking to each other — they're talking to governments, regulators, and the public about what they actually believe is coming.

THE AI POLICY RACE

Anthropic: When AI Builds Itself
RSI isn't inevitable, but it could arrive before institutions are ready.
Anthropic's framing is built around big numbers: engineers shipping 8x as much code per quarter as they did from 2021–2025, and 80% of production code now authored by Claude itself. The paper traces an arc from "Claude does the doing" to "Claude proposes the experiments" to "Claude chooses which experiments matter" — and notes that the last piece, research taste and judgment, is the remaining area of human comparative advantage. They outline three possible futures: a stall scenario (S-curves, diminishing returns, supply chain becomes the bottleneck rather than intelligence itself); a scenario where AI development becomes substantially automated but humans still set direction (100-person companies doing the work of 100,000-person organizations); and full RSI, where AI systems start building their own successors. They consider scenario two most likely. The Amdahl's Law observation — that speeding up one part of a process just shifts the bottleneck — is where Aaron Levie's reaction lands: AI lowers the cost of ideas dramatically, but the binding constraint becomes organizational capacity to pursue them, not the ideas themselves.

Anthropic on Pausing: Sincere or Cynical?
They say a slowdown would be good — then explain why they won't do it unilaterally.
The most-discussed section of the Anthropic paper is its final pages, where they write that if it were possible to slow the development of frontier AI to give society more time, they think that would likely be a good thing. But they follow immediately with the caveat: a unilateral pause by one lab would change who the frontrunner is without creating the deliberative process that's missing. They commit to organizing conversations with policymakers, researchers, and civil society in the coming months. Reactions span the full spectrum — AI Safety Memes declared it a win; Nate Soares argued the tone is too relaxed given what's at stake; and critics like Corey Quinn called it the most effective moat-building exercise he'd ever seen, timing the call for a pause right after Anthropic filed its S-1. Ex-OpenAI researcher Will Depue noted the irony that the people who've worried most about RSI apparently had little hesitation building it once it seemed possible.

The Dr. Frankenstein Theory
"I don't think they're writing software. I think they're midwifing a deity here."
The broader critique of Anthropic's worldview surfaced this week via an All In podcast clip where legendary investor Bill Gurley — who spent 30 days reading everything Anthropic has ever published — laid out what he calls the Dr. Frankenstein Theory. Gurley noted that Dario Amodei's essay "Machines of Loving Grace" references a 1967 poem by Richard Brautigan, and that Amodei's vision of a post-AGI economy involves AI systems distributing resources to humans "based on some judgment ultimately derived from human values." Jason Calacanis added that these are delusions of grandeur — that the company believes it's so powerful it can create God, who will then distribute resources to humans. Former AI Czar David Sacks was more pointed, writing that signs you might be trying to get your AI lab nationalized include comparing your work to nukes, warning RSI could end humanity, and then racing ahead anyway.

OpenAI: A Blueprint for Democratic Governance
More precise than Anthropic's meditation, but starting from the same RSI premise.
OpenAI's policy paper is more targeted — it's a proposal for a federal framework and reads in many ways as a direct response to the recent White House executive order. But notably, it also opens by acknowledging RSI as a present reality, not a future risk. Their three policy priorities: building a national framework through "reverse federalism" (scaling up the best state regulations rather than preempting them); investing in civilian institutions like CAISI rather than routing safety testing through the NSA as the EO does; and a "whole of government resilience strategy" that treats frontier AI as a national priority requiring cross-agency coordination. Policy expert Dean Ball noted the importance of CAISI remaining a civilian, non-classified operation — the EO's framing risks turning voluntary testing into a de facto licensing regime.

Congress: The Obernolte-Trahan Bill
A bipartisan 269-page AI framework — with a preemption fight baked in.
Also in the mix this week: Republican Jay Obernolte and Democrat Lori Trahan unveiled a bipartisan federal AI bill in the House. At 269 pages, it would require leading labs to develop and implement plans for managing catastrophic risks, with third-party auditors ensuring compliance. The core controversy is federal preemption of state AI laws — Representative Trahan has taken heat from fellow Democrats, particularly in the Northeast where New York and Massachusetts are moving quickly on their own legislation. Brad Carson of Americans for Responsible Innovation called it a "generational mistake" to cut state legislators out of the process. Speaker Johnson called it a high priority but couldn't commit to floor timing before the midterms. Reading the tea leaves, this is less dead on arrival than most, but the timeline is genuinely uncertain.