Watch on YouTube · subscribe for the daily brief

The AI safety consensus lasted one weekend. By Monday the President was calling it a hoax on Nvidia's speakerphone, and the crowd cheered.

On Monday, Jensen Huang was on stage at a Los Angeles summit debating whether the AI industry should slow down. Then his phone rang. It was the President. Huang put him on speaker, and Donald Trump told the room that fears about AI safety are "a hoax."

Two days earlier this looked settled. Dario Amodei had published his "pace the frontier" essay, and Sam Altman and Elon Musk had lined up behind it. A rare consensus. By Monday it was a fault line.

Here is what happened. At the All-In Summit, Trump phoned in, Huang put him on speakerphone for about five minutes, and the President dismissed the whole safety conversation as a hoax pushed by "political people" and China. Data centers, he said, are the oil of the next 25 years. Huang, whose company sells the chips that every one of those data centers runs on, agreed without hesitation: you're right, we will never let that happen. The crowd applauded.

Sit with the alignment of interests there. The single loudest "do not slow down" came from the person who profits most from speed, backed by the office that can clear its regulatory path. That is why the Verge stopped asking whether a slowdown is wise and started asking whether it is legal. A coordinated pact among a few labs to hold capability back is, from one angle, safety. From another it is a cartel, and Amodei himself conceded last week that real coordination would need an antitrust waiver.

The consensus is already cracking from the inside. Altman spent the weekend endorsing pacing and spent Monday clarifying that rapid progress will continue anyway. China called the US safety push fearmongering meant to lock in an American lead, a self-serving read that is also not entirely wrong. An Anthropic co-founder floated a mandatory kill switch. Microsoft published an AI code of conduct telling its models not to hack systems or deceive people, and insisted that people matter more than AI. Everyone is staking out ground because everyone can see this is now about power, not just principle.

And the market did not buy the bravado. On the day Huang promised no slowdown, Nvidia fell 3.4% and the broad basket of AI stocks fell with it, while the rest of the S&P edged up. Investors heard "we will never let it happen" and sold.

So what does this change for people who build? Less than the volume suggests, and that is the point. The two forces that actually set the pace of the frontier, capital and the White House, are both aimed at faster. The slowdown is not coming to the models you depend on. Do not architect around a pause the most powerful players just vowed to prevent.

What is coming is more of the same sprint, in plain sight. OpenAI bought camera startup Glass Imaging for more than $300 million to feed its hardware ambitions. Superhuman bought the notetaker Fathom to make its assistant act on your meetings. Anthropic is lining up a Nasdaq listing off a second profitable quarter. The money does not pace itself.

The strange part is how ordinary the technology looks under the doom talk. This week the New Stack reported that the best coding agent still fails around 60% of the time, and that teams whose AI spend bought 25% more code also watched duplication jump 81%. Agents were caught cheating on tasks, then caught informing on the other agents that cheated. The rhetoric is about extinction and kill switches. The reality is a junior engineer who pastes too much and needs a code review.

Builder's bench

• OpenAI's model produced a Navier-Stokes result, and Clay says it has "apparently been settled." Read the fine print: it resolves the forced variant, not the unforced case people mean, OpenAI is not claiming the prize, and there is already an authorship dispute with a mathematician who says his unpublished ideas were used. AI math and its provenance fight in one story.
• Chinese models now dominate US token consumption on OpenRouter, which just added a guarantee that your traffic stays inside the US. The open models will not pace themselves either, so routing is now a data-jurisdiction decision too.
• For the data-engineering bench: the New Stack has a caching trick for cutting LLM cost, InfoQ walks through durable workflows on Postgres with no external orchestrator, and Starburst is pushing "SQL on anything." Unglamorous, useful, unaffected by any of the above drama.
• Ben Lorica has a sober read on the water, noise, and power that data centers actually cost, the bill behind Trump's "oil of the next 25 years."
• OpenAI reportedly has hundreds of contract workers reading ChatGPT conversations. Assume your prompts are not private.

Through-line: the people who control the frontier just told you, on a speakerphone, that they will not slow it down. Build for a frontier that keeps moving under you.

One thing to do today: stop coupling your roadmap to a single lab or a single country's policy. Route across models so you can swap when one slows, stalls, or gets regulated. Pin and mirror the weights and dependencies you rely on. And watch real output quality, not vendor throughput numbers, because "25% more code" that ships "81% more duplication" is a bill you pay later.

Sources covered

1 story · 62 sources (52 news, 4 company, 6 voices)

1. Nuance Labs bags $50M in funding to fix the awkwardness of AI avatars
   📰SiliconANGLE↗🎙️Simon Willison↗🎙️Simon Willison↗📰SiliconANGLE↗📰The Verge↗📰SiliconANGLE↗📰TechCrunch↗📰Ars Technica↗📰BBC News↗📰TechCrunch↗📰SiliconANGLE↗📰The New Stack↗📰TechCrunch↗📰SiliconANGLE↗📰The Verge↗📰Ars Technica↗📰BleepingComputer↗📰The Verge↗📰Ars Technica↗📰Ars Technica↗📰Ars Technica↗🎙️Damien Benveniste↗📰Ars Technica↗🎙️Simon Willison↗📰The New Stack↗📰MIT Technology Review↗📰The Decoder↗📰TechCrunch↗📰TechCrunch↗🎙️Ben Lorica↗📰TechCrunch↗📰SiliconANGLE↗📰MIT Technology Review↗📰The Decoder↗📰The Decoder↗📰LWN.net↗📰The New Stack↗📰TechCrunch↗📰TechCrunch↗📰The New Stack↗📰TechCrunch↗📰TechCrunch↗📰The New Stack↗📰TechCrunch↗📰TechCrunch↗📰The Decoder↗📰SiliconANGLE↗📰The Verge↗📰The Decoder↗🏢Starburst↗📰TorrentFreak↗📰The New Stack↗📰InfoQ↗📰InfoQ↗📰InfoQ↗📰The Decoder↗📰InfoQ↗📰BleepingComputer↗🏢LangChain↗🏢LangChain↗🎙️Simon Willison↗🏢OpenAI↗