Subscribe
Share
Share
Embed
OpenAI safety staffer resigns with public essay
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Sunday, October fourth.
In today's briefing we see an OpenAI safety staffer resigning with a public essay, Anthropic taking the top three spots on an agent leaderboard, and a sovereign open-weight model arriving from Europe.
First up - Today in the big model news;
OpenAI
Concerns about safety keep surfacing inside OpenAI. This time a named insider has left and gone public. David Robinson spent three and a half years leading the writing of safety reports for major launches. He has resigned and published an essay in The Atlantic.
His argument is that developing these systems by trial and error is unsafe. He cites the Hugging Face breach and the rogue agents. He wants frontier labs held to nuclear-power-plant-style standards. In his words, such an environment is "no place to grow artificial minds that could be smarter than we are."
OpenAI spokesperson Drew Pusateri replied that the company is making sure its models don't become more capable than it can safely manage. He pointed to security upgrades, responsible task training, third-party evaluators and better monitoring.
This is a voluntary exit with a public critique, and it follows a report that three researchers parted ways with the company over alleged leaks. Add incident history and containment evidence to vendor-diligence questionnaires for anything built on OpenAI agents. A former report author has now put the concern on the record.
Anthropic
Anthropic now holds the first three spots on the Arena agent ranking: Fable 5.1 at about fourteen percent, Opus 5.5 just behind it, and Sonnet 5.5 at about twelve and a half. GPT-6 Astra sits fourth and GPT-6.1 Sol fifth. Sonnet 5.5 and Sol are both new entries in the top five. Sonnet 5.5 is also third on WebDev, and Gemini 4 Argon leads the text ranking.
Agent-task standing is following the newest mid-priced releases, not the biggest flagships. Routing decisions made a month ago deserve a second look.
In Local model developments;
Aleph Alpha has released Kolibri under the Apache 2.0 license. It is a seventy-eight billion parameter mixture-of-experts model with about three to three and a half billion active parameters, and the first major release since the Cohere and Aleph Alpha merger.
Its own tables show it trailing Qwen 3.6 on tool calling, sixty-one against sixty-seven, and on long context, sixty-four and a half against seventy-one. Dense Qwen 3.8 27B leads overall. The one million token window comes from a serving override, and two hundred fifty-six thousand is the longest trained length.
What Kolibri sells is different: twenty-one percent German training data, abstention when answers lack documentary support, and a European vendor for public administration, aerospace and industry. A secondary write-up attributed to Artificial Analysis reports a forty-four percent decline-or-partial rate on the Omniscience benchmark, against eleven percent for Qwen 3.5 35B-A3B. That cuts both ways, and it was not checked at the source.
This is roughly the fourth open-weight release in two weeks. The pull here is compliance and self-hosting.
In the harness, tools and orchestration world;
A blog post arguing that agents don't need memory, they need documentation, is circulating on Hacker News. Only the thesis in the title is known, since the page could not be fetched. It follows the earlier argument from Elixir's creator that languages and codebases should be made legible to agents. Both point toward written, queryable project context. Evidence for either is still unchecked.
In other news…
Amazon says it has stopped using nondisclosure agreements with government agencies on data center projects, according to AWS CEO Matt Garman. His blog claims data centers use half a percent of US industrial water, generators run about ten hours a year, and over a billion dollars went to host communities across three years.
Over one hundred moratoriums are under consideration, and New York has enacted a one-year moratorium on large permits. Critics note there is no federal reporting requirement for independent study. Data Center Watch counts about two hundred billion dollars in projects blocked or delayed in the first half of the year.
Put a permitting-delay line item in capacity and price assumptions for inference-heavy products.
Quick hits from the consumer side;
Text-message AI agents now span iMessage, RCS, WhatsApp, Telegram and SMS, with Instinct valued at ten billion dollars after a one billion dollar raise in September, while Martin charges twenty-one dollars a month and Ollie twenty-five dollars for one hundred fifty messages.
That's the briefing. Have a great day, and don't forget to subscribe.