Karpathy joins Anthropic; Google I/O launches Gemini 3.5 Flash

Gemini 3.5 Flash and Gemini Omni unveiled at Google I/O 2026

Google's I/O 2026 keynote on 19–20 May introduced Gemini 3.5 Flash, described as four times faster than comparable frontier models and stronger than Gemini 3.1 Pro on agentic and coding benchmarks; it is now integrated into Google Search's AI Mode and the Gemini app, which counts over 900 million monthly active users. Alongside it, Google launched Gemini Omni, a model that accepts any combination of text, image, audio, or video as input and generates dynamic video output. Google also reported that its SynthID watermarking system has tagged over 100 billion AI-generated assets, with OpenAI, Kakao, and ElevenLabs now adopting the standard and verification tools set to roll out in Search and Chrome.

Andrej Karpathy joins Anthropic to lead pretraining research

Andrej Karpathy, an OpenAI co-founder and the engineer behind Tesla's Autopilot programme, announced on 19 May that he is joining Anthropic to rebuild its pretraining research team under lead Nick Joseph. He will also establish a separate team dedicated to using Claude to accelerate pretraining experimentation. The hire is considered one of the most significant AI talent acquisitions of 2026, reflecting intensifying competition among frontier labs for researchers capable of shaping next-generation models.

xAI agrees to supply Anthropic $45 billion in compute capacity

SpaceX's IPO prospectus, filed 22 May, disclosed that Anthropic will pay xAI $1.25 billion per month to use the Colossus and Colossus II GPU clusters in Memphis, Tennessee, through May 2029 — a total of roughly $45 billion, terminable with 90 days' notice. The deal, an unusual arrangement between two competing frontier labs, places one of the first public-market valuations on frontier-scale compute capacity. The disclosure emerged as a secondary detail in the long-awaited SpaceX IPO filing, which also provided the first public view of xAI's broader finances.

London Mayor blocks Metropolitan Police's £50 million Palantir AI contract

London Mayor Sadiq Khan has vetoed a £50 million contract between the Metropolitan Police and US firm Palantir Technologies that would have used AI to automate the analysis of intelligence reports, email records, and phone data in criminal investigations. Khan's office cited a "clear and serious breach" of procurement rules: the Met had not sought required oversight approval and had not meaningfully considered any supplier other than Palantir. Civil liberties groups welcomed the ruling as a precedent for democratic oversight of AI in public policing; the Metropolitan Police warned it could lead to officer reductions.

EU AI Act compliance deadlines extended, nudification apps to be banned

European Union negotiators reached a provisional agreement on 7 May to amend the AI Act through a package called the Digital Omnibus on AI, postponing key compliance deadlines for high-risk systems. Obligations for Annex III systems — covering employment screening, credit scoring, and similar uses — have moved from August 2026 to December 2027; obligations for regulated-product AI (Annex I) move to August 2028. The package also introduces new prohibitions, effective December 2026, against AI systems that generate or manipulate non-consensual sexually explicit imagery or child sexual abuse material.