OpenAI safety leader resigns, warning labs aren't careful enough
A safety-transparency author is out the door with a farewell essay, and Germany has answered the sovereignty question with a model you can download today.
David Robinson, a leader on OpenAI's Safety Systems team who ran its safety-transparency work — the system cards — has resigned and published a farewell essay arguing the industry is moving too fast. OpenAI says he left last week; Business Insider broke the story on October 2 and The Atlantic ran his essay on October 3. "I agree with other recently departed staff that the companies building this technology aren't being nearly careful enough," Robinson writes, adding that "as the company sprints from one launch to the next, it is failing to achieve the level of care that I believe is needed." His prescription is unusually concrete: frontier labs "need to run like nuclear power plants or busy airports, with layers of redundancy and careful, time-consuming planning, so that the occasional and inevitable human error does not open a door to disaster." OpenAI's response is that it pauses training or holds back models "when we need to slow down." To be clear about what this is not: a resignation, separate from the three safety researchers the company dismissed a week earlier over outside information sharing — OpenAI fires three safety researchers over outside info sharing. Why it matters: the system card is the main public artifact of how careful a lab actually is, and the person who wrote them saying the care isn't there is about as direct a signal as the job can produce.
Aleph Alpha has released Kolibri, a 78-billion-parameter open-weight model built for German-language work in regulated industries. The Heidelberg company shipped Kolibri-1 on October 3, the Day of German Reunification, under the Apache 2.0 license with weights on Hugging Face. It is a Mixture-of-Experts design with 3.46 billion active parameters per token, a native context window of 262,144 tokens that validates out to one million, trained on 768 B200 GPUs across roughly 24 trillion tokens — 21.3% of them German, which Aleph Alpha argues is the only way to get real German rather than translationese. The company's own numbers put it at 75.5 overall on English benchmarks and 70.8 on German, just ahead of comparably sized Qwen and Nemotron models, and the announcement is unusually candid about where it loses — closed-book retrieval and agentic tool-use tasks. One framing correction worth making: "sovereign" here means built and run under German and EU law with no foreign control, not state money — neither the announcement nor the technical report mentions government backing. Why it matters: Europe's sovereign-AI bet has mostly been procurement promises and policy papers; this is a release a developer can pull down today, and it hit the Hacker News front page twice on day one.
What to watch: whether OpenAI refills the safety-transparency role from inside — and how many more system-card authors follow Robinson out.
When the person who writes the safety reports says the company isn't careful enough, should regulators treat that as evidence — or as one dissenter's view? Tell us in the comments.
