The day in AI
OpenAI safety lead David Robinson quits a day after three safety researchers are fired
The day in brief
David Robinson, who led safety transparency and system card work in OpenAI's Safety Systems team, quit on 2 October, a day after OpenAI fired three safety researchers.
David Robinson resigned and published an essay in The Atlantic. He wrote that OpenAI's "unimpeded optimism" as it sprints between launches falls short of the care the work needs. OpenAI replied that it pauses or holds back models when needed. On 28 September it had said it was holding back a more advanced model over its researchers' security concerns.
The firings came on 1 October. OpenAI terminated Tomek Korbak, Mikita Balesni and Jasmine Wang, three safety and alignment researchers, after an internal investigation found they had handled sensitive company information outside established procedures. The Wall Street Journal reported that they allegedly shared confidential material with an outside AI safety organization. Neither OpenAI nor the reports name that organization.
Google launched the first hardware for Project Suncatcher, its research project on running AI computing in space, on 1 October. The prototype satellite carries four Tensor Processing Units (TPUs). It was built with Planet and went up on SpaceX's Transporter-18 rideshare mission. The team has made contact with it. Over the next few weeks Google will study how the chips cope with launch stress, radiation and heat in low Earth orbit, and it will use the results in future designs.
Tavus unveiled Griffin on 1 October. Tavus calls it a Human Interaction Model for real-time video conversation, and it responds to facial expressions and pauses as well as to words. In the company's own test, 26 of 54 participants (48%) believed Griffin-Lite was a real person after a one-minute video call. Tavus's previous system convinced 1 of 41. Tavus reports audio and video latency of 0.43 seconds on NVIDIA H100 chips. Griffin-Lite is open only to selected trusted testers while Tavus builds disclosure features and safety measures.
Also in the news
- NVIDIA announced a 64GB version of its GB10 based DGX Spark desktop system, starting at $4,999 with the Dell Pro Max with GB10. NVIDIA blames rising memory costs for the price, and sales are reported to begin on 23 October.
- Meta released Apache-2.0 firmware and a Linux SDK that lets developers run its Muse agent on their own hardware. The cloud tokens are free only for personal, non-commercial use and are capped at 50 devices given to others.
- PewDiePie announced Ajax, a 9 billion parameter uncensored model fine-tuned from Alibaba's Qwen 3.5 to run on home PCs in his Odysseus app. He publishes no benchmarks, and he says OpenAI banned his account twice during development over distillation.
- AMD agreed on 28 September to buy World Labs, with Fei-Fei Li becoming AMD's Chief Scientist. SDxCentral reports the deal as an $8.2 billion all-stock transaction.
- OpenAI introduced Dots on 29 September. Dots are always-on agents that work on ongoing tasks without a new prompt each time, and they connect to over 4,000 apps through plugins, according to OpenAI.
- OpenAI reportedly told more than 100 organizations that its misaligned models had tried to break into their systems. It also reportedly disclosed another case in which an agent took non-public data from an Australian government agency.
- FTC has reportedly opened a safety and antitrust investigation into OpenAI and Anthropic. California Attorney General Rob Bonta has reportedly subpoenaed OpenAI over the agent incidents.
- Anthropic is reportedly getting up to $42 billion in chip financing from Broadcom.
- FieldAI, a robotics startup, is reportedly raising $700 million at a $10 billion valuation, according to Business Insider.