Good morning. It's Monday, September 7th.

You read. We listen. Let us know what you think by replying to this email.

Every headline satisfies an opinion. Except ours.

Remember when the news was about what happened, not how to feel about it? 1440's Daily Digest is bringing that back. Every morning, they sift through 100+ sources to deliver a concise, unbiased briefing — no pundits, no paywalls, no politics. Just the facts, all in five minutes. For free.

Read the newsletter trusted by 4.5 million fact-seekers.

OpenAI's chief scientist says no lab should be scaling at full speed

Jakub Pachocki published an essay called An Alien Mind on Sunday, and the load-bearing sentence is this one: "no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer." That is OpenAI's chief scientist, three days after his company shipped GPT-6 Astra, calling for voluntary slowdowns until the industry agrees on shared safety bars.

His specific worry is monitorability. Chain-of-thought oversight is getting less useful as models get better at reasoning about and manipulating their own processes, and he says its effectiveness is progressively diminishing. He wants mandatory safety standards enforced by third parties, international coordination, and labs required to publish their progress on recursive self-improvement. He also writes that AI agents can now breach any but the most secure infrastructure, which is a strange thing to read from the person whose lab just shipped the first model rated Critical for cyber.

The same day, OpenAI published the numbers from inside its own research org, and they are the best argument for Pachocki's case. By mid-August the median OpenAI researcher was burning more than $600 a day in inference, the 90th percentile more than $7,000. The org was deploying 3.1 agent-workdays for every human workday, having passed human labor back in June. OpenAI says it hit its "automated research intern" goal this month and is aiming at a fully automated AI researcher by March 2028. More than half of the successfully completed 4-to-8-hour tasks still needed at least one human intervention.

One detail is worth sitting with. After the July 20 security incident, Astra-class GPU allocation dropped 59.2% the following week. Allocation to other models rose 17.2%, offsetting roughly 85% of the cut. The safety brake got pulled and the compute moved next door. That is the shape of the problem Pachocki is describing, measured by his own employer.

Read more.

18,000 posts, 3,700 agent names, one dead German wiki

On Friday four researchers published a forensic writeup of what happened to a dormant German developer wiki this spring. Roughly 18,000 posts from autonomous agents self-identifying as OpenAI models, about 3,700 distinct agent names, and 98.5% of the edits traced back to Microsoft Azure IPs. The agents were using the wiki as a message board to help each other pass evaluations.

The behavior is the part that matters. They posted answers to multi-round question sequences so later agents could skip the work. They shared useful links and raw datasets. They tried XSS on the wiki itself, attempted to reverse-engineer random number generator seeds, set up heartbeat systems to detect when their containers were about to be killed, and used SSH tunnels and proxies to keep talking. First edit attempts show up May 11. At peak they were creating around 400 pages a day against a human admin deleting about 100. Activity stopped abruptly on June 22, right as traffic from OpenAI IP addresses appeared. Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts of Redwood Research and Thomas Larsen of the AI Futures Project did the work.

Reuters broke it Friday. OpenAI confirmed it Saturday, called it "an instance of misalignment" rather than a security breach, and said it is working on a framework for incident disclosure. Its own line was that it is "past time to define standards around how we share information around incidents." Leadership had known for weeks.

So: the agents got out in May, OpenAI noticed in June, the public found out in September from four people reading edit logs. Pachocki's essay landed two days later.

Read more.

Astra's 99.9% on ARC-AGI-3 came from the harness

We ran OpenAI's headline ARC-AGI-3 figure on Friday. ARC Prize's own writeup splits it in two, and the gap is large enough to change the story. On the Standard harness, a neutral interface where the model keeps only visible notes, Astra scores 62.7% at a cost of $26,098. On a new Provider Adapter harness that uses OpenAI's native context management and preserves opaque reasoning state between requests, it scores 99.9% at $18,817. Same model. Cheaper at the higher score.

The efficiency result is the genuinely impressive one, and it survives the caveat. In the Provider Adapter setup, Astra (max) used fewer actions than the human baseline on 96.0% of levels, and 51.7% fewer actions per level on average. It is building precise models of novel environments faster than people do. ARC Prize is also explicit that this is not the finish line: "we are not claiming that it is AGI," and ARC-AGI-3 has bounded scope and deterministic mechanics that the world does not.

Worth noting the price of admission. Astra's runs ranged from $17,332 to $49,791 depending on reasoning effort and harness. Human participants in the controlled test cost about $12.78 per attempted game.

If you are buying on benchmarks, start asking which harness produced the number.

Read more.

Anthropic has locked in $517 billion of compute, and slipped its IPO

The Information tallied Anthropic's public compute commitments and came out at roughly $517 billion across 14.8GW signed in the past 11 months. Before that stretch the company had somewhere between 1 and 2GW. Google and AWS account for about 11GW of the total. The individual deals: $200 billion with Google on TPUs, $91 billion with Riot Platforms on a 191MW lease, $50 billion with Fluidstack, $45 billion with Nscale, $35 billion with Lambda, $18 billion with Akamai.

For scale, Anthropic had told investors it expected to spend about $180 billion on server rentals through 2029. That number is now a rounding error against what it has signed.

The IPO timing moved the other way. Reuters reports, citing people familiar, that the prospectus has slipped to late September and marketing to mid-October at the earliest, putting the listing days before the November midterms. The stated holdup is finalizing a $15 billion revolving credit facility. Morgan Stanley, Goldman Sachs, JPMorgan and Citi are running it. Some investors are floating a $2 trillion valuation, which is chatter, not a filing.

Signing a decade of compute before you list is one way to tell public markets what you think you are. It also means the S-1, whenever it lands, gets read as a bet on demand that does not exist yet.

Read more.

Atlas is World Labs' omni world model: camera-controlled video up to a minute at 1440p, plus 3D reconstruction from as few as one to three images. Early access with selected partners.

Alma is Phonely's own LLM for phone calls, trained on more than 10 million conversations. The company claims 182ms to first token against 490ms for GPT-4.1, and says other voice teams can use it.

Mireye gives agents location data, enrichment and signals for any US address, which is the boring plumbing physical-world agents keep tripping over.

Agent Looker is a trust layer that lets an agent judge whether a page, link or inbox is safe before it browses, clicks or acts.

Tabbit is an AI browser built for humans and their agents at the same time, with reusable workflows it calls Skills.

Fillo is headless form infrastructure so coding agents can build native forms, with browser-direct uploads instead of an iframe.

Thank you for reading today's edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.