hot events · 2026-05-11

Thinking Machines, the company Mira Murati founded after leaving OpenAI, is finally showing product direction beyond funding headlines. Both The Verge and TechCrunch got a demo of an interactive AI that processes audio, video, and text simultaneously—it listens while you talk, watches what you show, and responds in text. The two outlets align closely on the real-time, multimodal angle, which suggests this is the core message the company wanted the demo to convey. I'd take it with a grain of salt for now. We're looking at two reports from the same demo, with no public benchmarks, latency numbers, model size, or pricing. TechCrunch's headline emphasizes "actually listens while it talks," which reads more like an experience claim than a technical spec. Real-time multimodal interaction isn't new—Google Astra and GPT-5 are chasing the same thing. The real differentiators would be how low the latency is, how natural the interruption handling feels, and whether the cost structure can support scale. None of that is public yet.

Three outlets track Google’s line closely: criminal hackers used AI to help discover and weaponize one zero-day. This reads like controlled disclosure, not independent convergence, because the date, target, model, tool, and actor names are withheld. I buy the direction of risk; I do not buy the completeness of the story. Google says the flaw hit a “popular open-source, web-based system administration tool,” bypassed two-factor authentication, and still required valid credentials. That is not a magic break-in button. It is AI moving vuln discovery and exploit scripting earlier in the kill chain. Against Anthropic’s Mythos claim last month of finding thousands of zero-days, the capability curve is ugly enough already. The disclosure style also helps Google push the regulatory narrative while keeping the evidence mostly unverifiable.

Two sources track the same event, and both run on OpenAI’s own announcement: DeployCo, the Tomoro acquisition, about 150 FDEs, and more than $4B in initial investment. This is official amplification, not independent discovery. I buy the direction; I don’t buy the clean story. Enterprise AI has not stalled because demos are weak. It stalls on permissions, data plumbing, workflow ownership, audit, and liability. OpenAI pulling in FDEs, Bain, McKinsey, Capgemini, TPG, and 19 partners is an admission that API-led self-serve growth hits a wall inside serious companies. Palantir already proved heavy deployment can reach core operations, but it also drags in long cycles, custom work, and margin pressure. That is the trade OpenAI is choosing.

Gowers wasn't messing around. He fed ChatGPT 5.5 Pro a set of open problems in additive number theory — the kind of material typically handed to new PhD students as a warm-up. The AI thought for 17 minutes and produced a theoretically optimal quadratic upper bound, combining Sidon sets and arithmetic progressions in a way Nathanson himself hadn't considered. Then it got wilder: Gowers asked for a harder variant, and the AI independently pushed the bound from exponential to sub-exponential, inventing a k-dissociated set construction along the way. MIT student Isaac Rajagopal reviewed it and confirmed the reasoning was correct and genuinely novel. Gowers contributed zero math — just "try this direction" and "write it up in LaTeX." Both sources agree because they're drawing from the same original blog post by Gowers, so the core facts are solid. But I'd discount this a bit: we only have Gowers' account and Isaac's review. No formal peer review yet, no independent verification from other mathematicians. arXiv won't accept AI-generated content, so the result currently lives as a blog link. Gowers' real concern isn't that AI is strong — it's that the PhD training pipeline just lost its first rung. The old entry bar was "prove something nobody has proven." The new bar is "prove something the AI can't." He offers two buffers: PhD students can collaborate with AI, and fields outside combinatorics may be harder for current models. But he admits this judgment might expire in months.

Two sources picked up the SpaceXAI trademark filing, but the accessible body is only a CAPTCHA page and headlines. I don’t buy the “officially announced” framing: the disclosed facts stop at a trademark application, with no filing number, class list, date, or clean SpaceX/xAI org link. The useful hook is not “Musk starts another AI company.” It is satellite data and orbital computation. SpaceX owns Starlink network telemetry, launch data, ground-station links, and orbital operations data; that is a different asset from Grok’s web-and-chat distribution. If the trademark classes really cover data processing, orbital scheduling, or edge inference, SpaceXAI is more likely a claim on aerospace data workflows than a consumer model brand.

Cerebras’ $35 billion IPO case rests less on beating Nvidia and more on OpenAI underwriting the revenue curve. The concrete hook is huge: OpenAI signed a 750MW compute purchase through 2028, with outside estimates above $20 billion. It also provided a $1 billion operating loan at 6% interest, tied to warrants for about 33.5 million common shares. That makes the story cleaner and more fragile at the same time. Cerebras posted $510 million in 2025 revenue and $87.9 million in net income, after losing $485 million in 2024. G42 concentration dropped from over 87% to 24%, but the customer-risk problem did not vanish; it moved to OpenAI. The WSE-3 inference pitch has substance, with 44GB on-chip SRAM and 21PB/s bandwidth. Investors are still mostly buying OpenAI credit, not independent demand proof.

Qwen-Image-2.0 is betting on document-grade image generation, not another poster-demo leaderboard run. The concrete hook is the stack: Qwen3-VL as condition encoder, a multimodal diffusion transformer, and instruction inputs up to 1K tokens. That length matters for slides, posters, multilingual text, and layout constraints. I care about this because image models spent the last year stuck at “looks good, breaks on control.” GPT-4o image also landed hardest on text, layout, and instruction following, not pure aesthetics. The weak spot here is evidence quality: the snippet claims large human-rated gains over the prior model, but gives no benchmark names, sample size, pricing, release weights, or failure cases. Without those, this reads like a strong technical direction with unpriced proof.