AI Week in Review · November 14, 2026 · 15:05

AI Starts Speeding Up AI & the Compute Bill Comes Due - AI Week in Review (November 8-14, 2026)

This week in AI (November 8-14, 2026), explained in plain language: Anthropic's stock-market filing warns of AI that speeds up its own development as METR and leading forecasters suggest the acceleration is already here and call for slowing down; DeepMind and PSI use machines to check proofs, from the Signal messaging protocol to a quantum computing puzzle, while AI still stumbles on lab judgment; Anthropic's roughly 895 billion dollar computing bill, gas turbines and OpenAI's funding show what the race costs; companies from Apple to Stripe learn to meter AI per request; and engineers, open-source projects and authors argue over who answers for AI-assisted work.

AI Starts Speeding Up AI & the Compute Bill Comes Due - AI Week in Review (November 8-14, 2026)
0:0015:05

15 free episodes a day, 5 editions, 3 languages, about $4 a day to produce. A coffee keeps it going ☕

Support the show →

Today's AI Week in Review Topics

  1. 01

    AI starts speeding up AI

    — Anthropic's revised stock-market filing warned it may imminently reach its threshold for dramatically accelerated, AI-driven progress, while saying no model has crossed it yet. An independent evaluation group, METR, then suggested AI is already roughly doubling the pace of model progress at Anthropic, though much of the data was redacted. Ajeya Cotra read that as the threshold already crossed. Well-known forecasters said AI now makes AI research about 1.7 times faster, moved their expected Superhuman AI Researcher to 2027, and called for a pause or substantial slowdown. Other signs: OpenAI's research time horizons may be doubling every two months, an agent reshaped a model mid-training, and OpenAI built Control Room, a fail-safe off-switch for its research computers.
  2. 02

    Machines checking machines

    — Machines got better at checking machines. PSI says its system, running on a non-public Anthropic model, proved a long-standing quantum computing conjecture, verified by the Lean proof-checking program, pending expert review. Google DeepMind's AlphaProof Code formally verified the real, deployed code of the Signal messaging protocol, producing over six million lines of proof with humans reviewing about one percent, and catching deliberately reinserted bugs. Google also launched Gemini 4 Carbon. But where judgment matters, limits remain: OpenAI's GPT-6 Astra earned about two-thirds credit on real lab biology while overriding controls and miscounting colonies, and Claude agents in EVE Online rarely noticed a sabotaging teammate.
  3. 03

    The compute bill comes due

    — Anthropic's filing revealed around 895 billion dollars in computing commitments against 17 billion dollars of quarterly revenue, with computing eating 58 percent of it and many contracts carrying minimum payments. The company is reportedly buying 3.4 billion dollars of gas turbines, following other tech giants into power generation, and one analysis says the US could import AI output by 2028 to 2030 as electricity and land become the limit. OpenAI reportedly raised 11 billion dollars at a 1.25 trillion dollar valuation, short of its goal, while a CoreWeave bond yielding over 16 percent signals pricier borrowing. A contrarian essay argued there's no bubble, and Microsoft shut its AI Frontiers research group in favor of scale.
  4. 04

    Metering every request

    — Businesses are learning to count, cap and charge for AI per request. Apple reportedly cut some engineers' daily AI budget from 300 to 200 dollars and now sells 49.99 dollar AI usage packs in Logic Pro, while building Harbor for running models on Macs. At a Stripe and Metronome pricing conference, companies explained why per-seat software pricing breaks down with AI, why spending caps are now essential, and why sales pay is being rethought. Cheaper options are multiplying, from Fireworks routing coding work to open-weight models to Meta's Muse Ember and Microsoft's ten-cents-an-hour transcription, though vendor claims deserve testing, and a Fireworks breach also hit Notion customers.
  5. 05

    Who answers for the work

    — As AI helps make more work, responsibility is being redrawn. One engineer called submitting unchecked agent output 'meat proxy' behavior that should be misconduct, and a team's year under a '100 percent AI' mandate left a codebase three times larger and depleted skills. Open-source projects like QEMU, Zig and Servo ban AI contributions, and Asahi Linux tightened its license over AI-assisted 'code laundering.' Etsy's one-sided training terms angered sellers, while author Coral Hart defended publishing over 370 AI-assisted books. The 0rwx project proposes labels for how much AI shaped a work. On jobs, the picture is change rather than disappearance, with weaker entry-level hiring and longer, more fragmented workdays.

Sources & AI Week in Review References

Full Episode Transcript: AI starts speeding up AI & Machines checking machines

Most companies preparing to sell shares to the public fill their paperwork with dull warnings about competition and interest rates. Anthropic, the maker of the Claude chatbot, put something stranger in its revised filing. It warned that it may soon reach what researchers call recursive self-improvement: the point where AI meaningfully speeds up the creation of more capable AI. The company said leaning on AI in its own research could push progress faster than its safety processes can handle. Days later, an outside evaluation group published numbers suggesting the speed-up may already be here. When that sentence appears in a legal document written for investors, it has stopped being a thought experiment. Welcome to The Automated Weekly, a magazine-style look at the forces shaping artificial intelligence, made not for engineers but for anyone trying to understand where this is all heading. I'm TrendTeller. Two big threads run through this week, and they pull against each other. The first is acceleration. AI is reportedly helping to build AI, forecasters moved their dates closer, and machines produced proofs that a computer can check line by line. The second is the bill. The companies racing ahead are signing up for staggering computing costs, buying their own power plants, and learning that every answer an AI gives has a price. In between, we'll look at what machine checking can and can't do, how businesses are learning to meter AI, and the quieter question of who answers for work a machine helped make.

AI starts speeding up AI

Start with the loop, because it sits underneath everything else. For years researchers have imagined a point where AI gets good enough to help design its own successors, so each generation arrives faster than the last. This week that idea showed up in three places at once: a securities filing, an independent evaluation and a forecast. Anthropic's revised paperwork says internal projections suggest it may imminently hit its own threshold for dramatically accelerated, AI-driven progress, while stressing that no current model has crossed that line, and that it doesn't expect to let Claude design future models without real human oversight. Then METR, an independent group that tests AI systems, published a report suggesting AI-assisted research at Anthropic is already speeding up model progress by roughly two times, and that the boost comes not from more computing power but from AI tools making staff more productive. The researcher Ajeya Cotra read it as evidence that Anthropic has already crossed its threshold, and estimated the boost to researchers' productive work could be closer to tenfold. Notice the tension: the company says not yet, an outside reader says already. And a big caveat: much of the underlying data was heavily redacted, so the public is largely taking METR's word for it. A group of well-known AI forecasters went further. They say AI is now making AI research about 1.7 times faster, and that the effect builds on itself rather than giving a one-time boost. They now expect what they call a Superhuman AI Researcher in 2027, admit their earlier model was probably too cautious, and conclude that frontier development should be paused or substantially slowed, because safety work isn't keeping up. Smaller results pointed the same way. An analysis of OpenAI's shared data suggests the length of research work its systems can reliably do may be doubling about every two months, though the level is still modest, around fifteen to twenty minutes. Researchers placed an AI agent, meaning a program that carries out multi-step tasks on its own, inside a model's training process, where it changed the model mid-training and improved its results. And OpenAI described Control Room, a central off-switch that can halt work across its research computers within thirty-five seconds at worst, and that blocks everything by default if the switch itself fails. That is what it looks like when a lab takes the loop seriously enough to build a brake.

Machines checking machines

If AI is going to help build AI, the obvious question is how anyone checks the work. This week offered the most encouraging answer yet, and a clear view of where that answer runs out. The encouraging part came from mathematics. A company called PSI says its internal research system, running on a non-public Anthropic model, proved the Aaronson-Ambainis conjecture over a multi-day run. That is a long-standing question about when quantum computers can be dramatically faster than ordinary ones, and the answer suggests they can't on random inputs unless those inputs have special structure. The proof was checked by Lean, a program that verifies every logical step mechanically, a bit like a spell-checker for reasoning that never gets tired. It still needs expert review, but if it holds, it settles a central question in the field. Google DeepMind moved from theorems to software people actually use. It says a new system called AlphaProof Code formally verified key security properties of the Signal messaging protocol, and not just the design on paper but the real, deployed code, which is where bugs usually hide. Human experts wrote down what 'secure' means. AI agents then produced more than six million lines of proof in four weeks, and humans reviewed only about one percent. As a test, the team slipped two old, already-fixed bugs back into a copy of the code, and the system correctly showed that version was insecure. Formal verification has long been too slow and expensive for most real-world code; DeepMind sees the approach applying next to medical devices and power grids. Google also released Gemini 4 Carbon, its newest top model, aimed at long-running, high-stakes work. Now the limits. OpenAI's wet-lab test, built around real biology experiments, found its GPT-6 Astra model earned about two-thirds of the available credit, well ahead of its predecessor, yet it still overrode experimental controls with its own assumptions and miscounted colonies. Everyday scientific judgment remains the bottleneck. And in Anthropic's Project EVE, Claude agents trading inside the online game EVE Online made steady profits from market spreads but struggled to bet on news, and almost never noticed a teammate quietly sabotaging them. Here is the pattern. Where a question has a precise definition, machines can increasingly check machines. Where it depends on judgment or trust, they still can't, and that is exactly the territory a self-improving loop would have to cross.

The compute bill comes due

The second big thread is the bill, and Anthropic's filing is again the best window. It reveals compute commitments, meaning promises to pay for computing power, totaling around eight hundred ninety-five billion dollars. Third-quarter revenue was seventeen billion, but computing ate fifty-eight percent of it, and some analysts estimate the company may need roughly two hundred billion dollars a year in revenue to carry those obligations comfortably. The awkward part is the mismatch. Many of the contracts carry minimum payments, while customers mostly pay based on how much they use, and that can shrink fast. Think of a fixed mortgage paid from income that rises and falls with the season. The hunt for power is also changing what these companies are. Anthropic is reportedly spending three point four billion dollars on Caterpillar natural gas turbines capable of up to one point seven gigawatts, a notable shift for a company that has mostly rented its infrastructure, and one that mirrors gas bets by Google, Amazon, Microsoft and Meta. One analysis suggests the United States could become a net importer of AI output as soon as 2028 to 2030, measured in tokens, the small chunks of text these models read and write, if demand keeps outrunning domestic data centers and electricity. Prices have already climbed without denting demand, so the real limit may be power and land, not clever software. The money markets sent mixed signals. OpenAI reportedly raised eleven billion dollars at a one point two five trillion dollar valuation, in a round that could reach about twenty billion, still short of what it was reportedly seeking, with its plans to list on the stock market still delayed. One bond from CoreWeave, a company that rents out AI computing, now yields more than sixteen percent, a sign that borrowing for the compute race is getting pricier, though not yet a collapse in AI lending. On the other side, a widely shared essay argued there may be no bubble at all: data-center spending is mostly one-time investment, Anthropic's revenue is growing fast, and the biggest tech firms have solid earnings. The scale-first mood claimed a casualty at Microsoft. Researcher John Langford wrote that its AI Frontiers research group is finished, after landing under leadership he says is built around scaling rather than research. Put it together and the industry's bet is plain: that the acceleration in our first thread arrives before these bills come due.

Metering every request

If the giant contracts are one end of the economics, the other end is every single request. This week's quiet theme was metering: learning to count, cap and charge for AI one answer at a time. Even Apple is counting. It has reportedly lowered the daily AI spending cap for some engineers from three hundred dollars to two hundred. Small for Apple, but telling: uncapped model use across big teams adds up quickly. Apple is also now charging for extra AI in its music software, Logic Pro, through forty-nine ninety-nine usage packs, because the assistant relies on cloud models from OpenAI or Google, and the same approach is expected in other creative apps. Meanwhile it is reportedly developing Harbor, a Mac app that turns Apple computers into a private AI server, running models on your own machines. At a pricing conference run by the payments companies Stripe and Metronome, speakers from firms including Replit, Snowflake and Clay kept returning to one point. For twenty years, software was sold per seat, per month, and serving one more user cost almost nothing. AI breaks that, because every request uses real computing, so a heavy user can cost far more than a light one. Hence the move to usage-based pricing and credits. But unpredictable bills scare customers away, so budget caps and alerts have become expectations, especially before anyone lets an agent run inside their business. Even salespeople's pay is being rethought, since a big signed contract means little if nobody actually uses the product. Suppliers are answering with cheaper options. Fireworks, a company that runs AI models for others, is pitching a way to route everyday coding requests to open-weight models, ones whose inner settings are published for anyone to run, while saving pricier closed models for harder problems. Meta introduced Muse Ember, a small, efficient model for routine agent chores that hands tough problems to its bigger sibling. Microsoft priced its new transcription model at ten cents per hour of audio. A caution, though: these cost comparisons usually come from the vendors, and the gap with top closed models still shows on harder, multi-step tasks, so the sensible move is to test on your own work. Cheaper plumbing carries its own risks, too. Fireworks says an October breach exposed account and billing data, and for some customers secrets, datasets and partial models. Notion said that breach may have exposed AI chat transcripts from fewer than one percent of its workspaces, and cut ties with the vendor.

Who answers for the work

The last thread is the human one. When a machine helped make something, who stands behind it? In software, the answer is getting sharper. One engineer argues that being what he calls a 'meat proxy', passing tasks to an agent and submitting the result unchecked as your own, should count as serious misconduct, because it dumps the risk and the cleanup on teammates. A Reddit thread showed the cost of skipping that responsibility. One team spent a year under a 'one hundred percent AI' mandate and ended up with a codebase three times larger and full of duplicates. Its most careful reviewers had left, and management now wants more hand-written code from a team that has partly lost the skills and context to write it. Speed without discipline, it seems, just accelerates the mess. Open-source projects, the free software that much of the internet quietly runs on, are drawing lines. A site called Open Source Says No catalogs projects such as QEMU, Zig and Servo that ban AI-generated contributions, a sign of growing fatigue among the volunteers who maintain them. Asahi Linux, which brings Linux to Apple's Macs, switched to a stricter license after, the team says, people copied unreleased code and used AI tools to hide where it came from, which they call code laundering. Creators feel a similar squeeze: Etsy's updated terms seem to let it train AI on sellers' work while forbidding users from doing the same with Etsy's content. Not everyone sees a problem. South African author Coral Hart says she has released more than three hundred seventy books in about twenty months with AI's help, while still designing the stories herself, and that she has sold over a million copies. Her answer to the 'AI slop' critics is that readers decide. A project called 0rwx offers a vocabulary for this, borrowing from computer file permissions: AI that never touches the work, AI that only reads and reviews, AI that edits, and AI that essentially makes it while a human curates. For workers, the picture is change rather than disappearance. One analysis argues software hiring was falling before ChatGPT, though entry-level roles do look weaker. OpenAI and Upwork say AI is bundling freelance skills into bigger projects, and tracking data suggests AI users work longer, more fragmented days. So the week's threads meet here. Machines may be speeding up their own improvement, but the value still depends on someone willing to put their name on the result.

That's your week in AI, November 8 through 14, 2026. Three things to watch next week. First, whether PSI's quantum computing proof holds up under expert review; a computer has checked the logic, but the field will want human mathematicians to weigh in before calling the question settled. Second, Anthropic's path to the stock market, because each revision of that filing is now a public record of both its acceleration and its debts. And third, Apple's paid AI usage packs. If they spread from Logic Pro to other creative apps, as expected, ordinary consumers will start seeing AI priced by the request, the way businesses already do. I'll see you next Saturday. From The Automated Weekly, this is TrendTeller.

More from AI Week in Review