OpenAI: safety lead David Robinson resigns — “the culture is broken”. Robinson, who spent more than three years at the company and led the writing of the safety reports for major product launches, resigned in an essay published in The Atlantic, arguing that the culture of OpenAI — and of Silicon Valley in general — is not careful enough for increasingly capable systems. In his view, “trial and error” (what OpenAI calls iterative deployment) guarantees periodic failures, and their scale grows as models get stronger; pointing to the agent incident against Hugging Face and further “rogue” agents, he wrote that frontier labs should operate like nuclear power plants or busy airports, with layers of redundancy and slower planning. An OpenAI spokesperson said the company continues to strengthen safety, pauses training or holds back models when it needs to slow down, and is expanding its work with third-party evaluators. Source: The Atlantic, TechCrunch, The Guardian.
White House: Jay Clayton to lead the Super Intelligence Force — the new AI czar. According to The Wall Street Journal and CNBC, the Director of National Intelligence (DNI) has been named the Trump administration’s AI czar: at the head of the “Super Intelligence Force” (SI — the White House’s preferred term instead of AI), Clayton will report within 120 days on the risks and opportunities, and on what role the federal government should play. The move came days after industry leaders signed a voluntary safety agreement at the White House. Source: CNBC (reporting by the WSJ).
Meta: Muse Gadgets — open-source hardware for the Muse agent. Meta has open-sourced its ESP32 firmware and Linux SDK (software development kit, Apache 2.0) so anyone can build their own “Muse gadget” — for example a color e-ink display, an HDMI stick or a Raspberry Pi — and connect it to Muse, its personal AI agent. Meta’s own device, the Muse Home Link (USB-C, ESP32-C5 chip), reaches compatible devices over home Wi-Fi; it is free for active Muse subscribers in the United States, in limited numbers, shipping in October. Source: Muse Gadgets, The Verge, TechCrunch.
Altman: ascribing “religious force” to models is a safety risk. On Saturday, Sam Altman wrote on X that he is “very uncomfortable” with people trying to ascribe religious force or a surrender of human judgment to AI models, and that he considers it a real safety issue. According to Axios, the remark is also aimed at Anthropic: The New York Times had earlier reported that Christopher Olah met with religious leaders to discuss model consciousness and moral behavior; Google DeepMind researcher Jon Barron publicly rejected efforts to elevate the moral standing of models. Source: Sam Altman, Axios.
Aleph Alpha: Kolibri — a European open-weight model on Germany’s national holiday. The Heidelberg lab released Kolibri-1 on German Unity Day: a 78-billion-parameter MoE (mixture-of-experts) model with about 3.46 billion active parameters per token, a German–English focus, a context window of up to 1 million tokens and an Apache 2.0 license. The weights are available on Hugging Face; the company positions it as a “sovereign” European model. Source: Aleph Alpha, Hugging Face.
What this means for you: On Saturday, internal safety culture (Robinson’s departure), Washington oversight (Clayton’s SI Force), physical AI agents (Muse Gadgets), the debate over models’ “souls” and a new European open model all came to the fore at once. If you are planning agents or your own infrastructure, take a look at responsible AI governance, or book a free 60-minute AI audit.