Summary

The Commerce Department’s directive suspended Fable 5 and Mythos access for foreign nationals, including those inside Anthropic, while Claude was pulled from the market entirely. Thompson thinks the administration may be demanding a technically impossible guarantee from probabilistic software, but Anthropic’s own conduct and rhetoric created the mistrust now driving policy. “Who gets final say?” is the real dispute.
Blocking frontier models may buy “weeks or months, maybe a year,” but it also delays the defensive work needed before comparable capabilities proliferate. Finding a vulnerability and helping patch it can be “the exact same request,” while defenders benefit from possessing their own code. Thompson’s preferred response is “the Manhattan Project we need”: inspect and repair web software and open-source projects at scale.
Anthropic’s safety narrative has created expectations of control that the technology cannot satisfy. After years of comparing AI with nuclear weapons and initially restricting Mythos through Project Glasswing, the company opened access roughly two months later. Sharp’s diagnosis: Anthropic promises something close to perfect security, then policymakers panic when jailbreaking proves unavoidable.
The clash is as much about institutional authority as cybersecurity. Sharp cited reports that Anthropic was slow to disclose 111 organizations with advanced Mythos access plus roughly 50 additional entities, including a South Korean company; Thompson pointed to the Department of War dispute, model “nerfing,” and Anthropic’s red line that “we get to decide how this is used.” The government may be technically misguided—and the directive might not survive court review—but Anthropic cannot credibly position itself above the state.
An AI bubble correction may punish companies that merely “plaster on AI to existing workflows.” Thompson expects the durable value to come from rebuilding products and organizations around probabilistic computing, comparable to replacing ads beside articles with algorithmic feeds. That transition takes years because users must first reset their assumptions about what software can do.
AI will probably be deeply disruptive, but confident forecasts about the exact outcome deserve skepticism. Thompson rejects both “nothing is going to happen” complacency and Anthropic-style prescriptiveness: the internet reshaped society over 25 years in ways no one in 1993 could have mapped, and AI might be larger. The missing ingredient is “epistemic… humility.”
World Cup hydration breaks illustrate how a transparent money grab can also improve the product. Thompson argued that the extra commercial inventory lets exhausted players reset, citing Croatia’s comeback before England ultimately won 4–2; Sharp agreed that “fewer exhausted players is a good thing” but said the pauses should simply be called commercial breaks. The broader media call is Thompson’s enthusiasm for the “eventization of sports”—premium tournaments over routine weekly inventory.

Deep dive

1. Hydration breaks monetize soccer while improving the spectacle

Thompson calls America’s habit of reaching policy goals through awkward mechanisms “the stupidest way possible”: emergency-room access becomes de facto universal healthcare, AI data-center spending becomes a backdoor way to put money in people’s pockets, and player-safety pauses become lucrative World Cup commercial breaks.
His counterintuitive defense is that the break can improve competition. Croatia had been “on their heels” against England, then reset, scored a couple of goals, and made an eventual 4–2 England victory gripping; fewer exhausted players also means more athletes “performing at their best.”
Sharp accepted the product argument but rejected the euphemism: call them commercial breaks and admit the profit optimization. Thompson’s broader viewing preference is the “eventization of sports”—World Cups, tennis’s Big Four, and major golf tournaments, with none of the weekly commitment.

2. Fable export controls expose a conflict built on incomplete information

Sharp’s factual anchor: late Friday, Commerce directed suspension of Fable 5 and Mythos access for every foreign national, “inside or outside the United States”—and, Thompson added, inside or outside Anthropic. Claude was pulled from the market altogether and remained unavailable as they recorded Thursday.
Thompson had briefly experienced Fable as a genuine capability leap and joked about being forced back to “only 5.5 in Opus.” Yet he stressed that non-public information makes the confrontation unusually difficult to assess: “What do we not know?”
The reported catalyst was Amazon jailbreaking Fable and raising concerns with the White House. Thompson cautioned that the public account came through a researcher who saw an Anthropic report—“this is all Anthropic spin”—but found plausible the claim that slightly changing a prompt enabled vulnerability discovery.
His Occam’s razor explanation leaves blame on both sides: the government says stop a behavior, Anthropic says “you can’t stop it,” and officials hear another refusal from a company that has already created “conditions of mistrust.” Thompson therefore wondered whether he had been “a little too hard on Anthropic” that week.

3. Probabilistic software makes perfect restrictions the wrong objective

Thompson’s decisive example: “Find the vulnerability in this software” and “help me patch vulnerabilities in this software” are “the exact same request.” A deterministic system can prohibit a defined function; a general model cannot reliably separate malicious and defensive intent without sacrificing useful capability.
Autonomous driving supplies the analogy. Years of hard-coding every road scenario failed because the variables were effectively unbounded; end-to-end models became viable by generalizing from enormous numbers of situations and developing internal heuristics for events they had never explicitly encountered.
Thompson sees the same category error when people treat AI as “a fancy search engine.” Sharp’s early use of ChatGPT to retrieve sports statistics was “the exact wrong function”: the model’s strength comes precisely from not being deterministically designed for every query.
With Mythos-level capabilities likely to appear elsewhere, Thompson said the White House seems to be putting a “thumb in the dike” when “the wave is going to go over the whole wall.” Every day of restricted access is also a day defenders cannot use the leading model to find and repair weaknesses.

4. Cyber defense requires going through the danger, not around it

Thompson did not minimize the threat: “Almost all software on the web is likely insecure.” His claim is narrower—that access cannot be foreclosed forever, so “the only way to solve the security problem is to go through it” and patch aggressively before the capability becomes broadly available.
Defenders retain an important advantage because they possess the code and can inspect and repair it directly; an outside attacker must infer an obscured system. The required mobilization is “the Manhattan Project we need,” spanning web codebases and open-source projects.
Anthropic deserves credit for Project Glasswing’s defensive work, but its messaging helped trigger the backlash. It spent years likening AI to nuclear weapons, said Mythos was too dangerous for release while it secured banks and big technology companies, then roughly two months later made it widely usable.
Sharp’s pushback sharpened the contradiction: Anthropic created an expectation of perfect security and exclusive stewardship, then appeared surprised when policymakers tried to reclaim control after discovering safeguards could be bypassed. “If you’re talking about a super weapon,” government authority is the inevitable destination.

5. Anthropic’s real problem is its claim to final authority

Thompson’s objection is not that Anthropic worries about safety, but its recurring implication that “only we can handle this. Only we are responsible.” His blunt logical construction: if Anthropic should run AI and AI will run the world, then Anthropic is effectively saying it should run the world.
Sharp asked whether that charge relied too heavily on inference. Thompson answered with the Department of War dispute, where Anthropic’s red line was “we get to decide how this is used,” plus its demonstrated willingness to silently nerf models when behavior violates its preferences.
Sharp added reports of 111 organizations receiving advanced Mythos access and roughly 50 more entities disclosed later, including a South Korean company. Anthropic’s slowness suggested it still had not absorbed that “the government is ultimately in charge,” even if the export-control mechanism might be legally or technically unsound.
Thompson’s Microsoft review supplied the organizational diagnosis: “Your what is excellent. Your how needs a lot of work.” But he ultimately conceded the failure is substantive too—an assertion of authority over customers, companies, and the federal government by people who may not understand “the reality of the world as it actually is.”

6. AI value will emerge from redesigned workflows and humbler forecasts

Thompson expects part of the AI crash to come from businesses trying to “plaster on AI to existing workflows,” becoming frustrated when a probabilistic system will not behave like conventional software. Durable companies will instead rethink the workflow itself, much as internet media moved from ads beside articles to feeds.
That reconstruction takes years because people must reset their expectations before new organizational forms become obvious. The internet transformed society over roughly 25 years; Thompson noted that someone in 1993 would never have predicted its effects on media would eventually help produce a Trump administration.
AI may be even bigger, but Thompson rejects both denial and “the certainty and prescriptiveness” of confident safety elites. Understanding AI does not mean knowing everything about society; his positive prescription is more doubt, better communication, and “humility” about outcomes no institution can map in advance.

‘Hard Fork’ Live, Part 3: Differing Visions of an A.I. Future Why AI Labs With Unlimited GPUs Still Fail — Anjney Midha, AMP