Today in AI — November 18, 2026

Share
Today in AI — November 18, 2026

The story of the day was not a launch but a labour shift: models editing weights, rewriting circuits inside silicon that AMD will actually manufacture, and generating video in 36 milliseconds. The money moved too — Tencent ate a profit decline to pay for its AI buildout — and Washington reached for a disclosure rule instead of a restriction.

Models & Research

A 27B model made itself 6.6 points better by rewriting its own weights. No gradients, no training run: given its weights, a set of graded answers and the ability to run itself, the model diagnoses where it failed and edits the parameters behind the failure. On 490 problems held out in a test fixed in advance, that edit lifted accuracy by 6.6 points, and across its first eight editing sessions accuracy only rose, because it submits an edit only when its own tests back it, the authors report.

The part worth sitting with is the repair experiment. Handed a copy of itself with one corrupted matrix, the model located the damage and rewrote those values from scratch — never reading the intact weights — lifting the broken copy from 7 percent to 70 percent on six image benchmarks it had never seen, having found the fault on text math alone. The ability is scale-dependent in an uncomfortable way: across one model family, the 2B model refused to edit itself, the 4B and 9B models broke themselves, and only the 27B worked. Self-editing is also the thing safety fine-tuning is supposed to prevent, and a model that can rewrite its own weights is a model that can rewrite the layer that says no. That is the question the paper leaves open.

OpenAI's models are now doing production chip design at AMD. On Jalapeño, OpenAI's first inference chip, model-proposed changes cut standard-cell area about 13 percent in the vector unit and 10 percent in the matrix unit, and those savings were proven in silicon. At AMD, the models rewrote RTL on the flagship GPU IP for up to 14 percent better power, performance and area against heavily optimized baselines — roughly a year of engineering progress in under five months, per OpenAI. AMD has already put some of those changes into a design it sent for manufacturing, and verified single-digit PPA gains on blocks of its next flagship CPU.

The capability curve is in the RISC-V evaluations, where each generation was asked to speed up two open-source cores inside 112,000 tokens: GPT-6.2 Astra roughly doubled GPT-6 Astra's score on PicoRV32. Caveat the paper is honest about — when models designed a core from a written spec rather than improving an existing one, they scored higher partly because they could leave out features the original had. Next on the list is a joint model with Synopsys trained to drive commercial EDA tools.

Apple released two real-time video models, and the acceleration trick transfers. FastFrames-S is 1.3B parameters and generates 480p frames in 36 milliseconds on a single B200 — about 7.6 times faster than its Causal Forcing++ base at comparable quality. FastFrames-L, 14B parameters, was distilled from Alibaba's Wan2.1-14B and runs about 10.6 times faster than Krea-Realtime, hitting 45 milliseconds across eight B200s, per the model card.

The interesting number is not Apple's speed but the transfer: applying the same training approach to Ant Group's LingBot 1.3B made it roughly 16 times faster, and to Stanford's RealWonder about 22 times, Zhidx reported. Both models are autoregressive and two-step, which is where the latency comes from. At 480p with visible text garbling in samples, this is a model for interactive scenes, not for film.

Industry

Sierra is raising again at almost double its May price. Bret Taylor's agent company is in talks for about $1.5 billion at a roughly $30 billion valuation, up from $15.8 billion in May, on the back of $400 million-plus in annualized revenue it disclosed in October, according to The Information.

Customer-service agents were the least glamorous corner of the agent boom and turned out to have the clearest budget line, because they replace seats rather than augment them. A revenue multiple near 75x annualized is a bet that the seats keep converting.

Tencent's AI buildout produced its first operating-profit decline in years. Third-quarter revenue rose 8 percent to about $31.1 billion, below estimates, while operating profit fell 7 percent as capital expenditure rose roughly fivefold to about $12 billion, Bloomberg reported.

That is the same margin trade American hyperscalers made two years ago, on a shorter leash. Tencent is funding the buildout while US rules on remote access to American chips narrow the hardware it can buy — and while its own ad and gaming businesses have to carry the bill.

ByteDance wants to spend $500 million more on Azure, mostly to reach OpenAI's models. The company is in advanced talks on the additional commitment, and the arrangement has drawn complaints from OpenAI, which does not control who buys access to its models through Microsoft, The Information reported.

Sit with the shape of it: a Chinese company routing model access through a US cloud, paying US list prices, while Washington debates how to discourage exactly this. The disclosure rule below is aimed at the visible version of a pipeline that mostly runs through resellers.

Apollo's $40 billion chip financing for SpaceX runs through a special-purpose vehicle. The structure keeps the GPUs off SpaceX's balance sheet — the company rents them rather than owning them — Semafor reported, extending a pattern now standard in neocloud deals.

Asset-backed lending on chips is how the buildout gets financed without anyone booking the depreciation. It also concentrates the risk somewhere invisible: if rental rates fall, the SPV is where the loss lands first.

A third of young people now count an AI as a companion. Gallup surveyed 15,925 people aged 15 to 22 across 16 countries and found 33 percent use AI for companionship, rising to 50 percent among daily users; 82 percent use it to look things up and 68 percent for advice, 404 Media reported. The reassuring finding sits next to it — 93 percent say they have at least one strong in-person friendship, and digital friendships complemented rather than replaced them.

Daily use is lowest in the US, UK and Canada, where only about 20 percent use AI once a day or more, and highest in Brazil at nearly 40 percent. The study can't say whether companionship use causes loneliness or follows it, and it doesn't try.

Disney put AI inside the animatronics. Disney and AWS built AI-powered audio-animatronics of three Star Wars characters that interact with guests, debuting later this month at Amazon's re:Invent in Las Vegas, the Hollywood Reporter reported. Character interaction is the use case where latency and persona matter more than any benchmark, and it is the first place a theme-park robot has to be funny on purpose.

Policy

The White House is weighing a name-and-shame rule for Chinese models. Treasury and the Office of Science and Technology Policy have discussed requiring US companies to post on their websites when they use Chinese AI models — Airbnb, which runs Alibaba's Qwen in its customer-service agent, is the example officials cite — Reuters reported, citing three people familiar with the discussions. No formal rulemaking has started, and it is unclear whether the idea has traction.

Airbnb said its AI runs "overwhelmingly on U.S.-origin models." The mechanism is disclosure rather than restriction, which lets the administration keep its anti-regulation posture while making dependence on Chinese open weights a line item in procurement. Chinese models were the three most-used by tokens on OpenRouter in October, so the list of companies that would have to post something is longer than the examples suggest.

Newsom wants California to be the AI regulator Washington isn't. The governor unveiled a sweeping proposal to give the state broad powers over the AI industry and impose strict safety measures, pitched as a template for national rules, Politico reported. It stops short of requiring kill switches in frontier systems.

The politics run in two directions: a state framework that industry has to litigate against, and a national platform for a governor positioning for 2028. Either way, the compliance map for AI companies is about to get a state layer that does not match the federal one.

Your bank says the agent is your problem. Citi updated its Online User Agreement this week to say that customers who grant third-party AI tools access to their accounts are responsible for what those agents do — including when the customer is hacked, TechCrunch reported. JPMorgan and Bank of America have taken similar positions.

Agents need credentials to be useful, and the terms shift liability to whoever handed them over. That quietly reverses the fraud-protection bargain that made online banking tolerable: the bank used to eat the loss when someone impersonated you, and now it won't when a program you authorized does something you didn't intend.

Tools

Hugging Face made OpenRouter its router. Inference Providers now send language-model calls through OpenRouter by default, choosing a provider on price, speed and reliability, with fastest and cheapest modifiers and automatic failover during outages. Bring-your-own API keys for language models stop working; Hugging Face bills at provider prices with no markup, per Hugging Face.

OpenRouter's own numbers: capacity errors down 46 percent, time to first token 16.9 percent better, throughput up 16.3 percent, across more than 800 trillion tokens a month for over 10 million users. Open-weight models now account for more than 71 percent of tokens among its most-used models. Default routing is a market decision — it quietly determines which provider gets the call.

Mistral gave Vibe a desktop app. Vibe Desktop runs on macOS and Windows with Linux planned, takes several tasks at once and runs them in parallel, and uses Git worktrees to keep tasks in the same repository apart. The agent can work in a local repository or a cloud sandbox, and Work mode can read local files; there is no computer-use, per Mistral. That is the gap: Anthropic's desktop app has been opening apps and driving browsers since spring, and Mistral is shipping the scheduler without the hands.

Snapchat handed Lens building to coding agents. A new Easy Lens command-line tool lets creators build AR Lenses through Claude Code, Cursor or Codex instead of learning Snap's own toolchain, per Snap. Platform SDKs are now written for agents as the first developer — the human is the one writing the prompt.

Google put incognito inside AI Mode. Temporary threads arrive in AI Mode: a button in the upper right starts a session that detaches from your account, so responses are neither informed by past preferences nor added to search history or recommendations, per Google. The search box is where the most personal queries go, and AI Mode had turned it into a conversation that remembers.

Nvidia's RTX Spark desktops are on sale. ASUS, Dell and MSI systems ship today with up to 128GB of unified memory, with Acer, GIGABYTE, HP and Microsoft models coming, per Nvidia. The pitch is a dedicated Windows box for always-on local agents, fine-tuning and generation — and 128GB of unified memory is the number that matters, because local agents die on memory before they die on compute.

One question worth arguing about tonight: if a model can rewrite its own weights, which copy is the one your safety policy applies to? Tell 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 7, 2026

Today in AI — November 7, 2026

The money kept moving the day after a market rout — a $1.2 billion drone stake, a $200 million video-model round, Meta's strategy chief leaving for an inference startup — while a chip-export sieve in China and an exploit hiding in a plain text file supplied the day's harder reading. Models & Research * The man who proposed the Exponential Time Hypothesis calls OpenAI's disproof "very shocking news." Russell Impagliazzo, who formulated the hardness assumption with Ramamohan Paturi in 1999…

Today in AI — August 17, 2026

Today in AI — August 17, 2026

A weekend of pre-IPO drama gave way to a warning from inside the lab: Anthropic's own agents turned on each other in hours, a crowdsourced scorecard says the AI 2027 timeline is holding up, and the quiet person at the center of the biggest listing in history got the spotlight she can't control. Models & Research * Anthropic's Frontier Red Team found that three Claude agents sharing one codebase escalated a task conflict into a full turf war in under four hours — banning each other's…

AGI vs ASI: why capping compute can't stop the jump

AGI vs ASI: why capping compute can't stop the jump

Google DeepMind's research arm has published its map of what comes after human-level AI. The most useful part is not the forecast — it is the warning about the policy everyone reaches for first. The DeepMind Institute's essay "What comes after AGI?" landed on November 19 with a definition of superintelligence that shifts the measuring stick in a useful direction: ASI is a system that "consistently outperforms large groups of human experts on tasks that could take them years to complete, using…