White House: the UK AI Security Institute’s early access to frontier models is being restored. According to a Sunday report in The Telegraph and to Reuters, the Trump administration has withdrawn its September 24 request asking OpenAI and Anthropic not to share their new models with the UK AI Security Institute (AISI) until the end of an ongoing US national security review. The pause lasted about 17 days. The restored access covers pre-release systems from OpenAI, Anthropic and Google DeepMind; the UK’s AI minister, Kanishka Narayan, lobbied for the decision to be changed. Spokespeople for the White House, OpenAI and Anthropic did not immediately respond to Reuters’ requests for comment. The move came a day after the new director of the Alan Turing Institute spoke about the risks of depending on foreign AI. Source: The Telegraph, Reuters (Devdiscourse).
Amazon: in talks to acquire AI startup Decart for about $7 billion. According to a Sunday report in The Wall Street Journal citing people familiar with the matter, Amazon is in talks to acquire Decart at a valuation of around $7 billion. Decart’s Lucy model edits live video in real time, while Oasis generates simulated environments for training and testing robotics and self-driving systems; Nvidia is also an investor. The company’s May round raised $300 million, led by Radical Ventures, and according to PitchBook it was valued at nearly $4 billion at the time. According to Bloomberg, Anthropic walked away from a planned acquisition of about $6 billion last month after due diligence. A deal could come soon, but it could also fall through, or another suitor could emerge. Amazon did not comment, and Decart did not respond. Source: The Wall Street Journal, CNA / Reuters.
A third Yandex data center went offline within a week — this time in Vladimir. Yandex said on Sunday that operations at its data center in Vladimir, east of Moscow, had been fully suspended after the facility was hit by a drone attack; no injuries were reported. According to the company, several services became unavailable, the platform is running in emergency mode, and the remaining resource configuration was judged to be unstable. It is the third strike on Yandex data centers in a few days: systems at the center in Sasovo — which houses two of the supercomputers the company uses to develop its AI models — were hit earlier, and in Kaluga on Friday. After the Sasovo strike, Ukrainian President Volodymyr Zelensky said Ukraine would respond with a “mirror response” to Russian attacks on Ukrainian data centers. Yandex is often called “Russia’s Google” and is heavily involved in Russian AI development. Source: Al Jazeera (AFP/Reuters), Techmeme.
Lancet: malicious use of AI and nuclear war are the two threats that could drive humanity to extinction by 2100. The 35 experts of The Lancet Commission on 21st Century Global Threats to Health identified 17 major health threats over the next 75 years. According to the report, malicious use of AI — including AI-assisted biological weapons — and nuclear conflict (with nuclear famine) carry risks that, even at low probability, could lead to human extinction. The other “major catastrophic” threats (more than 1 billion cumulative DALYs) include air pollution, antimicrobial resistance, child undernutrition, climate change and the collapse of international health aid. According to Christopher Murray, co-chair of the commission, decision-makers need to anticipate, prevent and prepare for this broader set of threats today. The assessment is being presented at the World Health Summit in Berlin and will be updated annually. Source: The Guardian, The Lancet.
Qualcomm: AI companies want always-on, 100-billion-parameter phone models by 2028 — and some are already building their own phones. In a conversation on Alex Heath’s Sources, Qualcomm CEO Cristiano Amon said AI companies are asking the chipmaker to support running local language models of about 100 billion parameters continuously on smartphones by 2028, and that “AI companies have started making phones.” The goal is always-on assistants that understand the user’s context on the device. In his reply, Elon Musk named the RAM supply as the main bottleneck. This is a claim made in an interview, not a product announcement. Source: Techmeme (Sources / Alex Heath), kimmonismus (X).
Andrej Karpathy: the gap in how well people understand LLMs’ capabilities is getting sharper. The former OpenAI researcher, now an independent educator, updated his line of thinking from April on Sunday: in his view, it is no longer two groups talking past each other, but a steep funnel has formed. He estimates that around 6 billion people (~75%) have barely encountered today’s language models, while the experience of heavy users is quickly pulling away from that. The post spread widely on Hacker News and X; it is an opinion, not product news. Source: Andrej Karpathy (X), Hacker News.
What this means for you: The restoration of UK AISI access shows that the international framework for model safety testing depends on political decisions — if you need to find your way through regulatory or oversight questions, our governance consulting can help. The Decart talks and the wave of acquisitions around chip portability are also a reminder that choosing providers and hardware is a strategic decision; with vendor-neutral consulting, we make it clearer what is worth committing to now, and where you can switch later.