Andrej Karpathy switches to Anthropic, OpenAI quietly kills Sora, and America's first real AI law is 15 days away: the industry moved fast on June 15.
Andrej Karpathy joins Anthropic: a major signal about where the frontier is moving
Andrej Karpathy, who co-founded OpenAI and built Tesla's self-driving AI program, announced he's joining Anthropic to lead pre-training research on Claude. He'll work under research lead Nick Joseph and build a dedicated team that uses Claude itself to accelerate model research. The move is one of the highest-profile talent switches in recent AI history.
Why it matters: When the people who built the foundations at one company move to another, it tells you where they believe the next breakthrough is coming from. Karpathy betting on Anthropic is a credibility vote in how Claude is being built, not just marketed.
Anthropic calls for a global AI pause, while filing for a near-$1 trillion IPO
Anthropic published a paper titled "When AI builds itself," proposing a globally coordinated slowdown on frontier AI development, citing risks from systems that can improve themselves. The paper landed just days after the company quietly filed for an IPO that could value it near $1 trillion. Both moves are drawing significant attention, and pointed questions about whether they can coexist.
Why it matters: A company simultaneously sounding the alarm about AI risk and racing toward a trillion-dollar public offering is either admirably transparent or deeply conflicted. Either way, it puts the safety-vs-commercialisation tension at the centre of the public conversation in a way that is hard to dismiss.
OpenAI shuts down Sora, its viral AI video app is gone after six months
OpenAI has discontinued Sora, the AI video generation app that surpassed one million downloads in its first week after its public launch in late 2025. The app is being wound down quietly. OpenAI redirected resources to its new $150M Partner Network, an enterprise programme designed to help global organisations deploy AI at scale.
Why it matters: Consumer AI apps that go viral don't automatically become sustainable businesses. Sora's shutdown is a reminder that generating buzz and generating revenue are different things, and that OpenAI is pivoting hard toward enterprise contracts, where the money actually is.
Colorado's AI law takes effect in 15 days, and it changes how AI can be used at work
Colorado's AI Act, the first comprehensive US state AI law, kicks in on June 30, 2026. It requires companies using AI for high-stakes decisions (hiring, lending, healthcare, housing) to conduct bias risk assessments, disclose AI use to people affected by it, and maintain human oversight. Separately, a federal executive order is pushing back, directing agencies to challenge state AI laws that conflict with a national framework.
Why it matters: If you've ever applied for a job, a loan, or insurance and wondered whether an algorithm influenced the decision, this is the first US law that requires an answer to that question. Fourteen other states are watching how enforcement plays out before writing their own versions.
Microsoft releases Phi-4 vision reasoning model, capable AI you can run without a subscription
Microsoft released Phi-4-reasoning-vision-15B, a 15 billion parameter open-weight model that handles text, images, and complex reasoning in math and science. At 15B parameters it's much smaller than frontier models but punches above its weight on benchmarks. Because it's open-weight, researchers, developers, and students can download and run it without paying per query.
Why it matters: Not every AI breakthrough requires a corporate subscription. Open-weight models from large labs are steadily closing the gap with expensive frontier models, making capable AI tools available to people and institutions that can't afford enterprise pricing.
Bollywood goes AI: Indian filmmakers embrace generative tools as content rules take shape
A CNBC report highlights how India's film and media industry is embracing generative AI to meet surging content demand in a $32 billion sector, with productions using AI for scriptwriting, visual effects, and dubbing. Simultaneously, India's government published updated AI content moderation rules in May 2026, setting expectations for how platforms handle AI-generated content. The IndiaAI Mission continues to expand national compute with 38,000+ GPUs provisioned and thousands more planned.
Why it matters: India is doing something few countries are managing simultaneously, building AI infrastructure, writing AI policy, and deploying AI creatively, all at once. The outcome of that three-track experiment will touch 1.4 billion people.
The Anthropic paradox: does "safety-first" survive contact with a trillion-dollar IPO?
Anthropic is the company most associated with AI safety research. It just published a proposal to slow down frontier AI globally. It is also the company closest to a $1 trillion public valuation. These two facts are in real tension. What happens when the financial pressure of a public company meets a research culture that genuinely believes its own product could be dangerous? The answer to that question will shape AI governance more durably than any regulation passed this year.
NAVER bets on gigawatt-scale AI with Nvidia + deepfake job fraud losses up 460% since 2020
South Korea's NAVER is building out sovereign AI infrastructure using Nvidia's DSX platform, starting at 55 megawatts with plans to scale to gigawatt capacity, one of the largest national AI compute commitments outside the US and China. Separately, Experian's annual fraud forecast warns that AI-powered fake job candidates and deepfake video interviews are now a quantified crisis: losses from job scams jumped from $90M in 2020 to over $500M in 2024, and detection is getting harder as the tools improve.