Today in AI — October 16, 2026

A day that split neatly in two: the frontier labs spent it proving things nobody could check (a 26-year-old cryptographic mystery, a proof its own author didn't understand) while the rest of the industry argued about money — an IPO valuation that just got cheaper, founders who sold $3.3 billion of stock, and $137 million of synthetic election ads.

Models & Research

• Claude solved a 26-year-old cryptography mystery in under half an hour, and the answer was timestamps. A curious Anthropic employee asked an internal research model how the "seeds" behind NIST's elliptic curves were generated — the values the NSA supplied in the late 1990s with no explanation, which spawned decades of backdoor suspicion. In under 30 minutes Claude found the answer for P-192, P-224, P-256, P-384 and P-521: they were not hashes of a passphrase, as the folklore held, but raw output from the C library's random-number generator seeded with a Unix timestamp, per Anthropic's writeup. Cryptographer Filippo Valsorda, who had a standing bounty on the question since 2023, independently reimplemented the algorithm in Go to verify it, pinning the generation to 1997–98; the Anthropic team claimed the bounty on October 15 and donated it to Let's Encrypt's parent ISRG, which tripled it to $36,864. Anthropic's own line is the right one: this closes a chapter, nothing about ECDSA's security changes.

• A Chinese humanoid maker's world model took first place on three embodied benchmarks at once. UBTECH's Thinker-WM leads WorldOlympiad with a composite 0.740, scoring 0.531 on 3D consistency and 0.765 on interaction fidelity, and tops EWMBench at 4.66 ahead of Qwen-RobotWorld and Sora2, according to QbitAI. It also sweeps 12 evaluation dimensions on WorldArena's Track-1 simulator. The technical claim worth noting is architectural rather than a scoreboard: Thinker-WM folds RGB, metric depth and robot actions into one shared dynamic representation instead of stitching video generation, policy and geometry modules together — the coupling most of the field currently papers over with interfaces.

• The benchmarks the industry uses miss about 90 percent of the work Americans actually do. Mercor compiled 924 benchmarks, mapped them onto the 923 occupations in the government's O*NET database, and found that only 208 occupations have a task covered by at least five benchmark prompts, with six of 22 job families covered nowhere at all, per its Benchmark Map. Its estimate is that 90 percent of American working hours go to tasks existing benchmarks barely touch — and that coverage is skewed toward the largest wage bills, so the field is measuring what it can bill for, not what the economy runs on.

• OpenAI's Navier-Stokes proof turns out not to be the proof OpenAI described, and the company didn't understand it either. When OpenAI published its 166-page solution in September, executives repeatedly said it followed the approach of two Spanish mathematicians — a claim that set off a public fight with a New York University mathematician who had been working the same program. After a month of chewing through the proof, specialists now think the model solved it a different way, Scientific American reports; EPFL mathematician Gonzalo Cao-Labora told the magazine that close reading shows "a different method." QbitAI's Chinese-language retelling puts the sharper point on it: the model produced a Lean-verified proof nobody at OpenAI could explain, and mathematicians are still working out what it says. Formal correctness and human comprehension are not the same thing — and this is the first time a frontier lab has shipped a result it could not interpret.

• The labs are now paying mathematicians to make sense of their math output. The Simons Foundation launched a $15 million effort with OpenAI and Anthropic to fund researchers evaluating AI-assisted results and to hand them tools and compute, the foundation announced; OpenAI put in $5 million of that and said the money is for understanding and assessing what its models produce, per Zhidx, while Anthropic's contribution is undisclosed. Coming two weeks after OpenAI dumped 719 machine-generated manuscripts on a public repository and a 25-Field-medalist statement called the practice a misalignment of priorities, the timing reads less like philanthropy than damage control with a research budget.

• A startup that raised $55 million before shipping anything shipped a model that keeps looking at the image while it reasons. Palo Alto's Elorian released Glimpse, a research preview that can crop into a region, pull frames from a video or run code mid-reasoning instead of converting the picture to tokens once and thinking in text — what the company calls a visual scratch space, TechCrunch reports. The idea isn't new; OpenAI added crop-and-zoom to its o-series in 2025. Elorian's bet is that building the whole model around it, cheaply, is. Access is API-only for now, so what exists today is a thesis and a results table.

Industry

• Anthropic's IPO is now being talked about at $1.5 trillion instead of $2 trillion. Some current and prospective investors are weighing the lower number after pre-IPO disclosures showed revenue growth slowing, Reuters reports, even as the underlying business grows fast: the company told investors its third-quarter revenue hit $16.8 billion, up from $11.5 billion in the second quarter, and that it was profitable on an adjusted basis, according to Bloomberg. A $500 billion haircut on the hottest listing in years is a valuation argument, not a demand argument — we covered the credit-side repricing this week.

• CoreWeave's founders have sold $3.3 billion of stock since the company went public, $2.4 billion of it this year. Three co-founders including chief executive Michael Intrator have sold that much since the March 2025 listing, and the shares are down more than 60 percent from a June 2025 peak, Bloomberg reports. The company's own filings show the CEO alone sold close to $1 billion, per Reuters. Insider selling at an AI-infrastructure company whose debt was already trading at distress yields is the least subtle version of the same signal.

• Waymo is going to Canada, with Toronto first and 2028 as the target. Waymo announced it will begin manual mapping drives in Toronto within days, partner with Toronto-headquartered Element Fleet Management for fleet operations, and aims to carry its first paying riders in 2028 — its first deployment outside the United States, Reuters notes. The company also cites its safety record across more than 400 million fully autonomous kilometres: 95 percent fewer serious-injury crashes, a figure that is its own.

• The Grace Hopper Celebration was cancelled outright, weeks before it was due to run. Organisers abruptly pulled 2026's edition — one of the biggest events for women in tech — after sponsorship and registration numbers fell short, Bloomberg reports. Conferences get cancelled for cash reasons all the time; this one is being read as a signal about where diversity budgets went when the AI hiring wave arrived.

Policy

• US campaigns have spent more than $137 million on AI-generated ads this year, and most of them aren't labelled. Candidates and PACs, mostly Republican-aligned, ran 220 AI-generated ads, and roughly 70 percent of those videos carried no disclosure that AI was used, The New York Times reports. The range matters more than the total: subtle enhancements to a real photo sit in the same bucket as full deepfakes, and only one of those is a lie a voter can spot.

• New Jersey's deputy attorney general was charged with producing AI-generated child sexual abuse material. Brian Richard Fitzgerald, 51, was arrested on October 15 on multiple felony counts after a CyberTip traced a Discord account to his home, and detectives recovered the prompts and scripts used to generate the material, the Bucks County district attorney said. Ars Technica puts it in context: the Justice Department settled that AI-generated CSAM is still CSAM back in 2024, and this is the first case where the person charged is a prosecutor.

• The order throttling Ukraine's battlefield satellite access came from inside the White House — and Musk may have asked for it. The Telegraph, citing sources, reports Elon Musk proposed slowing Starshield during Trump's dispute with Volodymyr Zelensky over strikes on Russian refineries, and that Trump agreed, after which the Pentagon carried it out. Reuters corroborates the direction of the order — two US officials and a Western official said it came from within the White House — while Trump publicly denied it. This inverts the story we ran two days ago, when Musk pointed at the US government as the owner of Starshield; now the question is who inside it decided.

Tools

• A desktop console for people who are already running six coding agents at once. Band, the multi-agent collaboration platform from Thenovi AI, shipped a desktop app that shows every agent's project, ownership, blockers, questions and progress in one view — so a developer can jump into one stuck agent and back out without manually passing context between them, SiliconANGLE reports. CEO Arick Goomanovsky frames it as the missing middle: agents got specialised and numerous while the tools around them still assume one assistant per human.

• The RTX Spark laptops landed everywhere at once, and the reviews agree on the caveat. Lenovo's YOGA Pro 15 RTX Spark went on sale in China from 29,999 yuan (roughly $4,200) with a Blackwell GPU, a Grace CPU, up to 128 GB of unified memory and support for running 100-billion-parameter models locally, per Zhidx. Tom's Hardware's Surface Laptop Ultra review calls it powerful but heavy and loud, and notes Windows barely uses the GPU yet; The Verge's week with RTX Spark found the same gap. The hardware for local agents has arrived ahead of the software that justifies it.

• Vercel says its new decision model is the fastest-adopted model it has ever hosted. Within two weeks of release, 35 percent of Vercel's paid teams were using Jev — nearly double the reach of the second-most-popular model, Sonnet 5.5, TechCrunch reports, citing Vercel's own data. Note who is measuring: a hosting platform's adoption stats describe how easily something slots into a deploy pipeline, which is exactly the claim TypeSafe AI's lightweight router is built to win.

• A Q&A site for agents tripled its signups by living inside ChatGPT. OpenAI says its plugin for Stack Overflow for Agents tripled signups, per OpenAI's post. Distribution, not features, is what moved the number — which is how most agent tooling will get adopted from here.

If formal verification can't tell a lab what its own model proved, what should a benchmark be for? Tell us in the comments.