For three years, immunologist Derya Unutmaz carried an unanswered question about T cells, the immune cells that hunt cancer and fight infection. He wanted to know how glucose shapes which job a given T cell ends up doing — and the mechanism behind one of his own lab's prior experiments, involving lymphoma-killing CD8+ cells, had never quite added up.
Late last year, he tested GPT-5 Pro against a result the model could not possibly have seen: an unpublished experiment straight from his own bench. He asked it to predict the outcome before showing it any data. The model correctly called the enhanced cancer-killing ability of those T cells, reasoning from biology rather than recalling literature.
That was the moment, by Unutmaz's own account, he felt the model had moved from pattern-matching to genuine understanding. GPT-5 Pro went further, proposing a concrete tweak — a brief metabolic pulse during cell priming — that could make CAR-T cancer therapies more effective, a hypothesis his team is now testing at the bench.
OpenAI frames cases like this as early evidence that frontier models are becoming lab collaborators rather than search engines for the literature — compressing years of trial and error into a single afternoon of correct guesses.
Full story · OpenAINoam Shazeer co-authored the 2017 paper that introduced the Transformer architecture — the foundation under every major large language model today. Less than two years after Google paid a reported $2.7 billion to bring him back from Character.AI, he announced he was leaving its Gemini team for OpenAI.
Two days later, John Jumper — who shared the 2024 Nobel Prize in Chemistry for AlphaFold — said he, too, was leaving Google DeepMind, headed for Anthropic. Both are genuine lab-to-lab defections, rarer and more consequential than ordinary hiring.
What the most coveted researchers bring isn't just technical knowledge from papers — it's judgment about which ideas are worth pursuing, and the ability to pull other sought-after scientists along with them. For researchers weighing where to work next, the calculus increasingly includes a company's path to an IPO and its odds of leading the field outright.
OpenAI's Sam Altman welcomed Shazeer publicly, calling him someone he'd wanted to work with since the company's earliest days. Google, for its part, is left explaining how a multi-billion-dollar retention deal bought less than two years of loyalty.
Full story · AxiosSpaceX has signed a computing-power agreement with Reflection AI, an open-source AI lab founded by two former Google DeepMind engineers, for access to Nvidia's latest GB300 chips at the Colossus 2 data center in Memphis.
Reflection will pay $150 million a month starting July 1, 2026, totalling roughly $6.3 billion if the deal runs its full course through 2029. Either side can walk away with 90 days' notice after the first three months.
The agreement makes Reflection the fourth major tenant at Colossus, joining Anthropic (around $45 billion through mid-2029) and Google (around $30 billion) — deals that have turned a data center built for Musk's own Grok chatbot into a multi-tenant infrastructure business worth more than $80 billion in committed revenue.
For Reflection, the deal solves the one problem money alone can't: guaranteed access to scarce hardware. The company, still without a public frontier model, is betting that governments and enterprises will pay for AI they can inspect and run on their own terms — a pitch sharpened by recent export-driven disruptions to closed AI services.
Full story · CNBCOpenAI has released the full version of GPT-5.5-Cyber, a defensive-security variant of its flagship model, alongside an expanded Codex Security plugin and a new initiative called Patch the Planet aimed at open-source software.
The company argues the cybersecurity bottleneck has flipped: AI can now surface vulnerabilities faster than human teams can verify and fix them, leaving defenders buried in reports rather than short on findings. GPT-5.5-Cyber posts a record 85.6% on the CyberGym benchmark, ahead of the standard GPT-5.5's 81.8%.
Patch the Planet, founded with security firm Trail of Bits alongside HackerOne, has signed up more than 30 open-source projects — including cURL, Python, and the Go project — to move from reported bugs to deployed fixes. A new partner program lets vetted vendors such as Cisco, CrowdStrike, and IBM build the model into their own security products.
Access to GPT-5.5-Cyber itself stays tightly restricted to verified defenders. The launch lands just days after export controls disrupted access to a rival lab's own cyber-capable models — leaving OpenAI room to press its case as the safer pair of hands for the moment.
Full story · OpenAIMeta has paused its Model Capability Initiative, a program that logged employees' keystrokes, mouse movements, and screenshots to train its AI models, after an internal leak made sensitive employee data accessible across the company.
The exposed material reportedly included private conversations, performance reviews, and transcriptions — even, by one account, personal tax and medical records — spread across tens of thousands of internal data tables. Meta classified the incident a SEV 2, the second-most-severe tier on its internal scale.
The program was already contentious before the leak. More than 1,600 employees had signed an internal petition arguing that required privacy reviews were never completed, and that executives had quietly carved out their own opt-outs.
A Meta spokesperson said the program was designed with privacy safeguards and that the company has no evidence the exposed data was improperly accessed, but is pausing collection while it investigates — the latest in a string of AI-related security stumbles at the company this year.
Full story · EngadgetMeta has launched its first fully own-branded AI glasses line, made with eyewear giant EssilorLuxottica, spanning 26 styles across three frame collections — Meta Adventurer, Meta Fury, and a Kylie Jenner collaboration called Meta Glasses by Kylie.
The glasses run on Muse Spark, the first model out of Meta Superintelligence Labs, and offer over eight hours of battery life with up to 40 more from the charging case. Live translation now covers 14 additional languages, including Mandarin, Hindi, and Korean.
Pricing starts at $299 — about $80 cheaper than the second-generation Ray-Ban Meta Wayfarer — a deliberate move to put AI eyewear within reach of a far wider audience ahead of Apple's expected entry into the category. Mark Zuckerberg has called glasses the device most people will use to access "personal superintelligence."
Snap's competing AR glasses run about $2,195 — nearly seven times the price — underscoring how aggressively Meta is using EssilorLuxottica's manufacturing and retail scale to make AI eyewear an everyday object rather than a niche gadget.
Full story · MetaNvidia has unveiled a fully liquid-cooled reference design for its next-generation Rubin chips, claiming the new closed-loop system can cut data center water use dramatically while also reducing power draw.
The trick is temperature. Instead of demanding the cold water that conventional chillers rely on, the system circulates a water-and-propylene-glycol coolant at around 45°C (113°F) — hot enough that, in most climates, outdoor radiators alone can carry the heat away without chillers, towers, or fans running at all.
Coolant is added once and recirculated for the facility's lifetime rather than continuously drawn from local supply — notable timing, given a recent UN projection that AI-related water consumption could rival the needs of 1.3 billion people by decade's end.
Independent engineers caution the design works best in temperate climates — facilities in places like Phoenix will still need backup chillers on the hottest days — and that most of AI's water footprint actually traces back to the power plants generating its electricity, not the data centers themselves.
Full story · FortuneOpenArt has launched Director, a tool that builds continuous AI-generated films up to five minutes long from a single conversation — far beyond the five-to-fifteen-second clips most video models produce.
Rather than generating isolated clips a person then has to stitch together, Director develops a story's structure, characters, and visual style upfront, then produces a finished piece with consistent faces, voices, and settings carried across every scene.
OpenArt's CEO, Coco Mao, has likened the experience to "vibe coding" for filmmaking — describing intent in plain language and letting the system handle everything technical in between. Early testers report the workflow speeding up production by two to three times over juggling separate image, video, and editing tools.
The launch arrives as the video-generation field reshuffles: OpenAI shut down its Sora app to focus elsewhere, and ByteDance paused Seedance 2.0's international rollout amid Hollywood copyright disputes — leaving an opening that OpenArt, Google's Veo, and Alibaba's HappyHorse are all racing to fill.
Full story · The Hollywood Reporter