Today in AI — November 21, 2026

Share
Today in AI — November 21, 2026

Wall Street spent the day digesting the AI trade's worst week since 2020, while the labs kept shipping: a general-purpose model doubled its score on a real robot benchmark, an Nvidia paper argued the context-window race is overkill, and the UN closed a week of talks on autonomous weapons by agreeing to keep talking.

Models & Research

GPT-6.2 Astra doubled a general-purpose robot benchmark, and knotted a rope. XDOF, a robotics infrastructure company, runs Ember-0: sixteen two-armed manipulation tasks on real hardware, six workstations, thirty repeats per task. GPT-6.2 Astra averaged 41.2 percent across the sixteen, up from 19.6 percent for GPT-6 Astra two months ago, and it became the first model to knot a rope and unfold a T-shirt — 143 earlier valid attempts at the shirt had failed for every model tested (QbitAI). The most lopsided gain is the fiddly one: picking up bread with tongs went from 10 percent to 90. Average prediction latency also fell from 15 seconds to 10.1, which is still slow for an arm waiting on a decision, and that is why XDOF thinks the benchmark may saturate before the hardware does.

Nvidia researchers argue the context-window race is overkill. A paper published this month, titled "You don't need long context," finds frontier models keep nearly all their accuracy on long-horizon and long-input tasks if the context is compacted frequently rather than held in one window; periodic summaries and targeted lookups did about as well as feeding everything in, per The Decoder's write-up. The window itself grew from GPT-4's 8,192 tokens in March 2023 to roughly a million today, about 120 times larger, and every token in it costs attention and KV cache. The practical read: compaction and retrieval are engineering problems with cheap fixes, and million-token pricing is a product decision, not a physical constant.

Tencent's Hunyuan team put small models on the roadmap. In a Reddit AMA overnight, the team behind Hy4 and Hy4 flash — chief AI scientist Shunyu Yao among the accounts on the thread — said a Hy mini, a lighter Flash-Lite and a model sized for 24 GB of VRAM are all planned, and that the next Hy generation will be significantly larger, QbitAI reported. Flash carries about 156 billion parameters and activates around 12 billion per token, with roughly 63.5 billion more in an Engram lookup table that can be pushed to system memory; sliding-window attention cuts the KV cache to about a quarter. Asked which post-training behavior is hardest to fix, the team answered with one word: overthinking.

Group-chat memory is the new agent benchmark, and recall does not equal service. Tencent Yuanbao and King's College London built GroupAssistBench from 50 real group chats and 80,299 messages: 2,395 questions split between recalling a fact and actually serving the group, the paper is on arXiv. One split tells the story — EverMemOS beat full context on recall, 0.621 to 0.570, and lost badly on service, 0.247 to 0.431. Tracking 497 paired memories, the authors found 36 percent of what the system had already recalled never reached the serving model's input at all.

ByteDance argues most of the training stack is unnecessary. A ByteDance team working with several US university co-authors published a three-stage pipeline, arguing that agent-training recipes have grown so nested that nobody can say which phase produced which capability, and backed the claim with an open model built that way, The Decoder reports. The claim is testable, which is the point: if three stages match a ten-stage stack, most of the field's post-training complexity is cost without a receipt.

Industry

The AI trade closed its worst week since 2020, and the reassurance started inside the hour. The Nasdaq 100 fell 5.6 percent Friday for a 10.4 percent weekly loss, its worst run in more than six years; the S&P 500 and Dow closed in correction and the Nasdaq composite is negative for the year, the Financial Times reported. Nvidia's seven-week slide accounts for $2.5 trillion of lost market value, and Anthropic's stock fell again days after an IPO that cleared a $1.5 trillion valuation — we wrote about how the first session actually closed.

Oracle told the market its OpenAI contract is fully funded, on a Friday, after the close. "Oracle has secured all the funding we need to fulfill our contract with OpenAI," the company posted. "With our final capital raise earlier this month, we became the first hyperscaler fully funded for this investment cycle." It added that any media reports "that will come out and say the contrary are incorrect and wrong," Reuters reported. A denial that specific, timed to land after the closing bell, is a statement about what Oracle expects to be published next as much as about its balance sheet.

Masayoshi Son explored a trillion-dollar buyout of SK Hynix, then shelved it. The SoftBank chief executive approached Wall Street institutions about financing a purchase of the memory maker, and SoftBank now says the idea is "no longer under active consideration," per The Information. Financing of that size and Korean government approval made it a long shot, but the fact that it was explored at all shows how far SoftBank will reach for a position in AI memory.

Databricks is back at a $225 billion valuation. The company is in early talks to raise several billion dollars at that mark, after closing a $5 billion round at $190 billion in August, the Wall Street Journal reported. That markup is the counterweight to the public selloff: private buyers of AI infrastructure are still bidding, and the gap between the two markets is where 2027's IPO pipeline gets decided.

Ubisoft spun its AI gameplay work into a subsidiary. Motherbrain, disclosed in the company's half-year results, is the team behind the Teammates project and is building "smart NPCs with memory and AI game masters that manage large-scale systemic worlds," GamesIndustry.biz reported. The same week Ubisoft cut 166 jobs in its Montreal and Quebec studios — the industry's two instincts, invest in generative tools and shrink the teams that used to do the work by hand, arriving in the same filing.

Policy

The UN could not agree on rules for autonomous weapons. A week of talks in Geneva ended with countries agreeing only to "start discussing options," with two years to propose regulations and room for some of them to remain optional, NPR reported. The International Committee of the Red Cross had asked for legally binding rules within a year, backed by the UN secretary-general. Russia delayed the agenda on day one and the US proposed a five-year window for nonbinding, optional guidelines, which is the shape of a process designed to produce nothing before the weapons are already deployed.

The chip supply chain warns the new US export rules reach past AI accelerators. Industry sources told Nikkei Asia that the latest controls catch a broader range of chips, including the systems-on-chip used in consumer electronics, Nikkei Asia reported. If that reading holds, the controls stop being a frontier-compute lever and become a general export regime, with the compliance burden landing on device makers who never modeled China risk.

China's AI capex doubled; AI revenue did not. Rhodium Group estimates China's biggest technology companies will spend $139 billion on capital projects in 2026, with revenue from their AI labs under 8 percent of that figure, the New York Times reported. Set that against Anthropic and OpenAI raising tens of billions in public markets and the argument writes itself: export curbs matter less than the return problem underneath them.

Tools

OpenRouter told customers their MiniMax prompts may have been read. The notice, sent by email rather than posted, says requests sent to MiniMax models may have been read by a third party; three hosts — AtlasCloud, Novita and StreamLake — told OpenRouter they forwarded those requests to MiniMax, and test links planted in prompts were opened 40 to 130 hours later, PrivacyAffairs reported. OpenRouter has pulled the Zero Data Retention label from the affected endpoints and says the problem sits with MiniMax and the hosts. MiniMax's only explanation so far, given to StreamLake: content moderation ran on its endpoints before October 22, and zero retention has been enabled since — which does not explain links opened days later, or the repository-cloning attempts a canary test caught.

Kling Image 4.0 pushes generated images toward finished commercial layouts. Kuaishou's new model renders at native 8K, holds up to 2,000 characters of text in one image, generates as many as ten coherent frames at once, and edits text already inside a picture without repainting the rest, Leifeng tested it. It moves to a unified autoregressive architecture for understanding and generation, ships more than 70 preset fonts, and handles instructions up to 15K tokens. Kling's second-quarter revenue passed ¥850 million, up more than 200 percent year over year, so the image model's job is retention: keeping creators in the app between videos.

The UN gave itself two years to propose rules on autonomous weapons. Caution, or capitulation? 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…