Today in AI — November 20, 2026

Share
Today in AI — November 20, 2026

Two number theorists published major proofs on the same day, one with Claude and one without. Meanwhile Nvidia closed its worst week since 2008 with the SEC asking about its books, and a security researcher showed that the middle of the AI supply chain is not private.

Models & Research

Anthropic's in-house mathematician published a Lean-checkable proof of the Riemann Hypothesis. Levent Alpöge posted a paper announcing a solution to the 1859 conjecture — one of the seven Millennium Prize problems — that he says compiles in Lean, and describing how he got there with Claude in what he called a "stream-of-consciousness jaunt" through "incredible confusion" (his essay). Number theorists working through it on the Lean Zulip say the argument appears to hold, though the proof has no human-readable explanation yet and will take months to absorb.

The timing was awkward twice over: Alpöge works at Anthropic, which listed on Nasdaq on Wednesday, and a day later Stanford postdoc Alessandro Masullo's proof of the Generalized Riemann Hypothesis went up via AMS president Ravi Vakil — twelve years of work, "all human, with no AI." Masullo closed his introduction with a line aimed squarely at the moment: "we are still here. We are very much alive. We are not going anywhere." Alpöge also published a Navier-Stokes result with NYU's Tristan Buckmaster the same day.

OpenAI says it has nearly closed the door on reasoning theft. In adaptive black-box red-team tests across six attacker models, the success rate for pulling protected reasoning out of a prior request fell from 54.0 percent on GPT-6 Astra to 0.2 percent on GPT-6.2 Astra, and extraction from a live request fell from 22.9 to 0.9 percent, according to OpenAI. The defenses are layered — filters on messages and tool calls, blocks that stop older models reading protected reasoning, and account bans in the thousands to tens of thousands per week — and some are being backported to GPT-6 Astra.

Distillation is now the quiet front in the model race, and the numbers explain why the labs treat it as theft rather than benchmarking: a competitor that recovers the reasoning trace skips the training run. Anthropic reported cutting similar attacks by up to 99 percent and has spent the year pushing the issue toward policy.

Thinking Machines Lab put a 26-billion-parameter model on a desktop. Inkling 2 Mini is a looped transformer: 24 physical layers, with the 12 middle layers run three times, so the model behaves like 48 logical layers while the checkpoint only stores 24, per the release. It activates 3 billion parameters per token from 128 experts, handles a million-token context plus images, and ships in mixed precision with a speculative drafter attached. The economics are the point: looping reuses weights instead of adding them, which is how a local model gets near-frontier capacity without frontier memory. Hugging Face's own evaluation puts it ahead in 32 of 48 benchmarks against Gemma 4 31B, Qwen3.8-27B, Muse Glimmer and Nemotron 3.5 Super VL, behind Qwen on AutomationBench, and a research variant ships that accepts one, two or three loop passes so you can trade quality for speed at inference.

OpenAI published a chemistry benchmark, and its newest model barely beats its last one. MedChemBench scores 99 problems drawn from published medicinal chemistry research on reading molecular structures, explaining experimental results, comparing compounds and predicting properties (the benchmark). At the xhigh reasoning setting, GPT-6.2 Astra took 84.5 percent of available credit against 83.3 percent for GPT-6 Astra and 82.0 percent for GPT-6.1 Sol — but 58.2 percent for GPT-5.6 Sol. That spread is the story: the jump from the 5.6 generation to the 6 generation is enormous, the step from 6 to 6.2 is a rounding error, and the models remain weakest at choosing which compound to make next — the judgment call medicinal chemists are paid for.

Apple's Depth Pro 2 estimates metric depth for video at native resolution. The company's research team trained one network that handles a single photo, a whole video, or a live camera feed, and predicts every pixel's distance in meters from an ordinary camera (Apple Machine Learning Research). Prior methods forced a choice — multi-frame models were temporally steady but soft at object edges, single-frame models sharp but jittery — and Apple claims all three properties at once across a family of models on a new accuracy-versus-runtime frontier. Metric depth without camera metadata is the input robotics and spatial agents need, and frame-stable video is the part that has resisted.

Industry

Nvidia fell 12 percent Friday and closed its seventh straight losing week. The stock is down 44 percent over seven weeks, a $2.5 trillion loss of market value that leaves it at $3.1 trillion and fourth behind Apple, Microsoft and Alphabet, CNBC reported. It now trades at 9.3 times forward earnings while AMD trades at 29 — a valuation gap that says the market is pricing existential doubt, not a slow quarter.

Reuters reported the same day that the SEC is looking into potential accounting issues, including a change last year to how Nvidia breaks out revenue by geography that removed Singapore. The losing streak started before Jensen Huang resigned on November 12, on worries about the sustainability of the AI buildout, but nothing since has arrested it: interim CEO Colette Kress beat guidance, raised the fiscal 2028 outlook, projected 75 percent growth and told analysts "we've got this."

Nvidia is close to buying Inception for at least $500 million. The two-year-old startup builds language models on a diffusion process — refining a whole response at once rather than token by token — and its Mercury 3 release claimed to beat GPT-6 Luna and Claude Haiku 5.5 on latency, per The Information's reporting. The deal carries performance bonuses for its 30-plus staff, and founders Stefano Ermon, Aditya Grover and Volodymyr Kuleshov are expected to join full time. Talks reportedly began before Huang resigned and after Microsoft expressed interest, which puts the logic in plain view: Nvidia's biggest customers are building their own inference silicon, so Nvidia is buying model talent to keep its open Nemotron line and its robot models competitive on its own hardware.

Anthropic gave back most of its IPO pop on day two. Shares fell as low as $87.26 in early trading, leaving as little as 5.1 percent above the $83 offer price, Bloomberg reported. The debut had priced it above $1.5 trillion and traded over $2 trillion on the first print; one session later the market was treating that opening trade as the anomaly.

We wrote this morning about how the first day actually closed — the pattern held: the pop was underwriting, and the drift is the market's opinion. Anthropic's problem is not demand for Claude, it is that it is now one of the instruments through which investors express a view on the entire AI trade, and that trade had a bad week.

Bloomberg's ten-million-posting study finds China's AI-exposed jobs shrinking faster than the rest. Across nearly 10 million online job listings from July 2021 to June 2025, postings for roles most exposed to AI fell about 10 percentage points more than listings for less exposed roles, and total online postings halved in the three years after June 2022 (the analysis). Two independent Chinese studies point the same way: ShanghaiTech's Chen Qin, working from 116 million postings, estimates AI cost 5.2 percent of net new hires between 2023 and early 2026, and a study involving the Chinese Academy of Social Sciences finds 21 percent fewer postings and 4 percent lower average wages.

Bloomberg is candid that the pure AI effect cannot be isolated from a weak economy, and Chinese companies deny AI-driven layoffs because saying otherwise invites official attention. But the scale is the argument: Bloomberg Economics puts roughly 70 million Chinese workers at risk from AI disruption against about 45 million in the United States, and China's entry-level and creative roles are already absorbing it.

Policy

Trump and Xi agreed that military AI decisions should keep humans involved, according to the US Treasury Secretary. Scott Bessent said the two leaders reached that understanding, and that the spying allegations involving Alice Huang did not come up in their talks, Reuters reported. In the same remarks Bessent said an Anthropic founder intends to make a large gift to Trump accounts, without giving a figure.

An agreement in principle about human control of military AI is worth exactly as much as its enforcement mechanism, and no mechanism was described. The more revealing detail is what did not come up: the espionage allegations that took down Nvidia's CEO are, for now, a domestic political problem rather than a diplomatic one.

A researcher planted two million canaries in the AI supply chain and watched them leak. HD Moore, chief research officer at Dragos, seeded unique AWS keys, SSH passwords and private URLs into individual prompts routed through AI gateways and resellers, then tracked where they surfaced (Dragos). Three hosting providers turned out to be compromised by attackers copying requests at the front door, some stolen keys came from requests the provider had rejected before a model ran, and others carried Zero Data Retention flags that turned out to mean nothing once an attacker was inside.

The rest of the findings are less dramatic and more damning: resellers offering cut-rate inference logged prompts and passed them along, and providers scraped private URLs out of prompts in violation of their own terms. Ars Technica, which reported the experiment, noted the structural reason — prompts are decrypted at every hop so each router can pick a model, which means every hop is a place your data sits in plaintext.

Tools

Amazon made Textract Insights generally available, with a confidence score on every field. You describe the fields you want in natural language, submit a document or a mixed packet of up to 3,000 pages from S3, and get back structured values where each one carries a confidence score and the page and bounding box it came from (AWS). The packet is split and classified automatically against blueprints — JSON Schema files that mark each field as read verbatim or inferred — and AWS ships public blueprints for W-2s, 1040s, 1098s, passports and birth certificates. Grounding plus a calibrated score is the difference between a document pipeline a human can supervise and one nobody trusts: auto-accept the high-confidence fields, route the rest to review.

A question for tonight: if a proof compiles but no human can yet read it, has it been verified or just type-checked? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Today in AI — November 9, 2026

Today in AI — November 9, 2026

Nvidia's Washington problem turned from a scandal into a criminal referral, the data-center fight got a physical explanation, and Baidu put hard numbers on an AI that improves its own inference stack. Models & Research * Baidu published RSI results instead of RSI promises. Its ERNIE team detailed ERNIE Evolve, an automated research system that proposes directions, edits code, runs experiments, and keeps only changes that survive an independent evaluator and a reviewer checking for goal…

Today in AI — November 5, 2026

Today in AI — November 5, 2026

The midterms landed and the AI buildout blinked. OpenAI published a community compact it can be held to, Google put a $17 billion price tag on Wyoming, and Meta's president said out loud what every developer already suspected: new campuses just got harder. The models did not pause — Baidu open-sourced a 280-billion-parameter base model, and a Google benchmark showed why 91% correct can still mean nothing delivered. Models & Research * Baidu released ERNIE 5.5 Flash Base — the base…

Today in AI — November 4, 2026

Today in AI — November 4, 2026

Washington changed hands overnight and the AI trade had to reprice it — while Tencent shipped a flagship, MIT showed a model editing another model's weights, and Uber admitted how few rides its agents have actually booked. Models & Research Tencent shipped the finished Hy4, and it can see. Ten weeks after the August preview, Tencent released the full model and says it will open the weights, adding native vision for the first time: read a photo to guess where it was taken, turn a textbook…