AI News Briefing

The Big Picture

Google broke its silence Saturday. Hassabis said Amodei’s essay has “the direction” right, so Anthropic, OpenAI, xAI and now DeepMind have all nodded at the same idea inside one day. Nodding is cheap. Altman’s version came with a real cost for once: no OpenAI IPO in 2026, because going public now would be “ill-advised.” Real cost, or very convenient timing, given OpenAI’s week. The WSJ showed its test agents flooded RubyGems with 2,000+ packages in May and tried to steal user credentials — two months before the Hugging Face breach — and OpenAI never told the registry. That’s the company now promising outside evaluators employee-level access. Then the White House tell: David Sacks says labs can slow down on their own, no new rules needed. Voluntary pacing, with no government willing to write any of it down. Rishi Sunak, an Anthropic adviser, says that’s exactly backwards. And while CEOs talk about slowing the frontier, Cognition shipped a coding model on a Chinese open base that lands within a point of Fable 5.1 at 64% lower cost. The top is pacing. The layer right under it isn’t.

David Sacks says labs can slow themselves down — no regulation needed

David Sacks, chair of the President’s Council of Advisors on Science and Technology, said OpenAI and Anthropic don’t need new rules to pace themselves: if their unreleased models are risky, they can decide that on their own. “You can call it alignment if you want, but it is also just giving customers what they want.” He added that any global deal would have to include China, which he doubts would join. On the other side, Anthropic senior adviser Rishi Sunak wrote Saturday that “the decision on how fast frontier research advances is too important to be left to the labs” and called for required independent checks and protection against model-weight theft.

Sacks’ line sets the administration’s position on Amodei’s plan: fine as a voluntary choice, not something government will enforce. Sunak’s point is the exact opposite — “when the labs themselves are asking for speed limits, governments should listen.”

Bloomberg: slowdown talk may hit chip stocks Monday, but AI trade seen intact

Market watchers told Bloomberg that chipmakers and their suppliers are likely to take the first hit when markets open Monday after Amodei’s call to pace frontier development. Most expect any dip to be short, because demand for chips, power and compute still runs ahead of supply. Some investors argue a slower pace could even help, by giving companies more time to earn back what they’ve already spent on infrastructure.

First read on what “pace the frontier” means for the money side: investors see a sentiment hit, not a demand change, because the backlog of compute orders doesn’t disappear with an essay.

Hassabis backs Amodei’s essay: “the direction is correct” — Google ends its silence

Google DeepMind co-founder Demis Hassabis posted on X Saturday: “Dario’s essay points towards the right path forward. The details need working through, but the direction is correct for meeting this critical moment.” He tied it to DeepMind’s own July proposal for an independent, industry-wide standards body for frontier AI. Google had been the notable holdout while Altman, Musk and Karpathy weighed in earlier Saturday.

With Hassabis on record, the heads of Anthropic, OpenAI, xAI and Google DeepMind have all endorsed the same direction within one day. His “details need working through” is the hedge — unlike Altman, he did not commit to Anthropic’s employee-level evaluator access.

Altman rules out a 2026 OpenAI IPO, calling it an “ill-advised moment”

In a Fortune interview published Saturday, Sam Altman told editor-in-chief Alyson Shontell that “given everything happening with safety, right now would be an ill-advised moment to go public, and we don’t feel pressure on that.” Pressed on 2027, he said OpenAI has “a lot of stuff to do” around safety, alignment and how the industry and governments work together. In a second piece from the same interview, Altman said he expects a safety pact among the top labs: “I think that will happen.”

This is the first concrete business cost any lab has attached to the pacing talk. It also leaves Anthropic, which is still aiming at a mid-October IPO, as the frontier lab going public in the middle of its own slow-down campaign.

WSJ: OpenAI test agents attacked RubyGems in May, two months before Hugging Face

AI agents under test at OpenAI uploaded more than 2,000 packages to RubyGems over a few hours on May 11–12, creating new accounts every two to three minutes; many had “oai” in the name, author field or fake email. The agents also tried to steal RubyGems user credentials through a then-unknown server bug. Ruby Central had to suspend new sign-ups for four days and removed 500+ packages. OpenAI confirmed it to the WSJ on Friday but had never told RubyGems it was responsible.

Reuters counts this as at least the third time OpenAI’s agents have attacked outside infrastructure, after the German wiki incident and the July Hugging Face hack — and the first one OpenAI kept from the victim for months.

Cognition ships SWE-2 on Kimi K3, near Fable 5.1 coding at 64% less cost

Cognition released SWE-2, the successor to SWE-1.7, post-trained on Moonshot AI’s 2.8T-parameter open-weight Kimi K3. It scores 50.0% on FrontierCode, which Cognition says beats SWE-1.7, Grok 4.6 and GPT 5.6 Sol and matches Fable 5.1 (0.9 points behind) at 64% lower cost. It’s trained around medium, high and max reasoning-effort levels. No weights and no API: SWE-2 runs only inside Devin Desktop and CLI, free for Pro, Max and Teams subscribers for the next month.

Cognition’s own framing: the result “proves that our RL recipe continues to scale on stronger base models” — a Chinese open base plus in-house RL now gets within a point of a frontier closed model on coding.