· 07:00 · 7 min + 3 min Explained
OpenAI and Anthropic say their AI disproved a famous computing rule
AI models at two labs claim to have broken a basic assumption about hard problems, and experts are now checking. Also: Claude in drug labs and city halls, Google's Genie 4, the hunt for chips beyond Nvidia, and fixes for coding agents.
Written by AI and checked against our editorial standards. Every claim links to its primary source. Not reviewed by a human editor.
Get it in your inbox at 7am every day ↓
Today in 60 seconds
- AI models at OpenAI and Anthropic claim to disprove a key rule about hard problems. Anyone curious whether AI can do original research.
- Anthropic shows Claude's early drug work and gives US government 40% off. Anyone following where AI firms look for big customers.
- Google's Genie 4 makes a world you can walk through, now with sound. Anyone into games, video or robots.
- OpenAI reportedly plans to buy Google's chips, and an Nvidia-free cloud raised $100M. Anyone wondering what will happen to AI prices.
- Claude Code stops overwriting long rule files, and Copilot agents get their own login. Developers who use coding agents.
1. OpenAI and Anthropic say their AI disproved a famous computing rule
The short versionAI models at OpenAI and Anthropic claim to have broken a basic assumption about how hard some puzzles are. Experts are checking, and nothing is confirmed yet.
Some puzzles are easy to check but seem very hard to solve. Given an answer, you can check it fast. But finding one seems to take exponentially longer as the puzzle grows.
Computer scientists turned that hunch into a rule called the Exponential Time Hypothesis, or ETH. Russell Impagliazzo and Ramamohan Paturi first published it in 1999, and hundreds of papers build on it.
On Monday, OpenAI said an internal model had found a method whose running time grows more slowly than that, which would make ETH false. The model found it on 16 October, and Lean, a program that checks proofs, confirmed it the next day, The Next Web reports. Hours after one of its mathematicians said OpenAI’s work built on earlier research, Anthropic posted its own disproof, found by an internal Claude model.
The catch is that nobody has fully reviewed the claims yet. Impagliazzo, Paturi and Ryan Williams of MIT lead the review of OpenAI’s claim. Their role “should not be taken as certification,” they write. Anthropic adds that its result says nothing about solving real puzzles faster.
OpenAI is “reasonably confident” it breaks no encryption, Williams told TechCrunch.
For decades, "this cannot be solved fast" was part of the ground computer science stands on. If an AI cracked that, AI is doing original research, not just helping. Williams says computer scientists should probably take P = NP, the field's biggest open question, a little more seriously.
Source: thenextweb.com · Reported · techcrunch.com · Reported · github.com/anthropics · Official
2. Anthropic shows Claude’s drug work and cuts prices for US government
The short versionAnthropic says Claude sped up early drug research at six partner firms. It is also giving every level of US government 40% off Claude.
Drug labs and governments are slow, careful buyers, and both drown in paperwork.
1 of 2 Claude in the drug lab
On Tuesday, Anthropic described its work with six partners in drug development. At Enveda, Claude read a large stack of manufacturing papers. It caught every problem the human reviewers had found, plus seven more, and cut review time by 89%. Anthropic also bought Mithril Medicine, a one-person firm working on two rare muscle diseases.
The catch: almost all of it is at an early stage, so no drug has been tested in people. If results hold, Mithril’s first drug could reach human trials in late 2028 at the earliest.
2 of 2 40% off for government
US federal, state, local and tribal governments now get 40% off Claude Enterprise, Claude for Government and the API. There is no minimum spend. Anthropic says the deal will save the government $1 billion.
Clever demos no longer win the big contracts. AI firms now have to prove their worth in slow, careful places like a drug lab or a city hall. A low price gets them in the door.
Source: anthropic.com · Official · anthropic.com · Official
3. Google’s Genie 4 builds a world you can walk through, with sound
The short versionGoogle DeepMind's Genie 4 makes a world you can move through in real time, and it now creates its own sound, music and speech.
A video model makes a clip you watch. A world model makes a scene you can walk around in, like a video game that nobody coded.
Google DeepMind showed Genie 4 on Tuesday. You move with the keyboard and change the scene with text, for example the weather. It is Google’s first real-time world model that makes its own audio.
The catch: only trusted testers can use it today, and Google says only that subscribers get it “soon”. Details can also drift over long sessions.
Games are what people notice first, but Google also pitches these worlds as training grounds. A robot or an agent can practice there over and over without breaking anything real.
Source: deepmind.google · Official
4. OpenAI reportedly plans to buy Google’s chips, as Nvidia rivals draw money
The short versionOpenAI plans to buy Google's AI chips, The Information reports. And a cloud firm built on chips from Nvidia's rivals has raised $100M.
Most AI runs on Nvidia’s chips. Google’s TPUs are a cheaper alternative, Reuters writes, so big buyers want them too.
1 of 2 OpenAI and Google’s TPUs
OpenAI plans to buy Google’s chips, called TPUs, The Information reported on Tuesday. It cites one unnamed person, Reuters writes, and gives no terms. Reuters reported in June 2025 that OpenAI had begun renting TPUs, and deals with Anthropic have already driven demand for them.
2 of 2 A cloud without Nvidia
General Compute supplies other cloud firms with servers that use chips from Cerebras, SambaNova and Positron. It raised $100M, and its value rose from $60M in May to $800M, TechCrunch reports.
It is like a town with one petrol station getting a second one. The first can no longer charge whatever it likes, and in the long run that shapes what you pay for AI.
Source: investing.com (Reuters) · Reported · techcrunch.com · Reported
5. Claude Code stops overwriting long rule files, and Copilot agents get their own login
The short versionClaude Code fixes a bug that could overwrite large CLAUDE.md files. And GitHub lets Copilot agents work under an app's own identity.
1 of 2 Claude Code 2.1.312
Claude Code reads a file called CLAUDE.md for your project’s rules. Version 2.1.312 fixes a bug where Claude could overwrite a very large CLAUDE.md or rules file in a subfolder after seeing only its start. So it could replace rules it had never read.
Do this: Update Claude Code to version 2.1.312 if your project has long rule files.
2 of 2 Copilot agents with their own login
Until now, a bot that used Copilot borrowed a person’s login. In a public preview, a GitHub App can now use the Copilot SDK under its own identity. The company pays, and its spending limits apply.
It is like giving a contractor a visitor's pass instead of your key card. An agent with its own badge can be tracked and switched off on its own.
Source: githubusercontent.com · Official · github.blog · Official
Open source & local AI
Liquid AI’s small LFM3 models hear, see and speak on a laptop
The short versionLiquid AI released six small open models for text, images and speech. They are built to run on your own devices, such as laptops, not only in the cloud.
Most voice assistants send your words to the cloud. Liquid AI’s LFM3 family is small enough to stay on the device.
The text model fits in about 1 GB at 4-bit. The Omni model takes text, images and sound, and can answer in speech. Liquid says the text model beats the much larger Qwen3.5-9B on long browser tasks, though Qwen still wins on knowledge tests.
The catch: the scores are Liquid’s own, and the model is not meant for coding agents. Firms with more than $10 million in yearly revenue need a separate deal, VentureBeat reports. Ollama runs the models only on Macs for now.
Small models that hear and speak mean a voice assistant can work offline. What you say never leaves the room, which matters to anyone bound by EU privacy rules.
Try it: On a Mac with Ollama, run ollama run lfm3.
Source: huggingface.co · Official · huggingface.co · Official · github.com/ollama · Official · venturebeat.com · Reported
Paper of the day
Maximum effort often costs more without solving more
The short versionResearchers at Together AI and two US universities found that an AI model's highest effort setting can cost far more without solving more. Starting low and stepping up works better.
Many AI tools let you pick an effort level. More effort means the model thinks longer, and you pay for every extra word. So most people assume more is safer.
The researchers tested eight models, including Claude Sonnet 5 and GPT-5.6, on math, science and coding tasks. On tasks that every level solved, the highest effort cost up to 42 times as much as the lowest. Higher effort also failed some tasks that lower effort had solved.
Their fix: start at the lowest level, and step up only when a check shows the answer is wrong. That matched or beat fixed top effort for every model.
For Claude Sonnet 5 on a terminal coding test, success rose from 76.1% to 83.0%, and the cost fell.
The catch: the paper is not peer reviewed, and stepping up needs a way to check answers.
Top effort works less like a quality dial and more like a taxi meter that keeps running. For most tasks, a short ride and a check at the end is enough.
Source: arxiv.org · Official
Explained: if you have a little more time
3 terms from today's news · 3 minWhat is the P vs NP problem?
P vs NP is the biggest open question in computer science. It asks: if it is easy to check that an answer is right, is it also easy to find that answer? "P" stands for problems a computer can solve fast. "NP" stands for problems where a computer can check a proposed answer fast. Every problem in P is also in NP. The question is whether the two groups are really the same. > **Think of it like this:** a jigsaw puzzle is slow to put together, but a glance tells you if a finished one is right. P vs NP asks if there is always a clever shortcut, so building the puzzle is never hopelessly slower than checking it. The Clay Mathematics Institute gives an example. You must pick 100 of 400 students for a dorm, and some pairs must never live together. If a coworker hands you a list, you can check it quickly. But finding a good list yourself seems hopeless, because there are more possible lists than atoms in the known universe. Stephen Cook and Leonid Levin first stated the problem in 1971, each on their own. It is one of the Millennium Prize Problems, and it is still unsolved. Researchers also use stricter guesses. One is the Exponential Time Hypothesis, from 1999. It says that for a well-known logic puzzle called 3-SAT, the time needed grows exponentially with the puzzle's size, whatever trick you use. If that guess is true, P and NP are different. But proving it false would not mean P equals NP. **Why it matters to you:** much of your online safety relies on some problems being very hard to solve. The encryption behind your bank login is one example. That is why claims that AI has shaken one of these beliefs get so much attention.Two more from today
What is formal verification?How a math-style proof can show that code behaves correctly for every input, not just the ones you tested.
Formal verification means proving, like in math, that a program always does what a written rule says. A test checks some inputs. A proof covers every input.
A computer program called a proof checker reads each step, so a wrong proof is caught. The catch is cost: proofs are slow to write, so they have mostly been used where bugs are very dangerous.
What is "compute" in AI?Why the chips and data centers behind AI are the biggest bill an AI company pays.
In AI, “compute” means the raw computing power behind a model: special chips, mostly GPUs, in large data centers, plus the power to run them.
Labs need compute twice. First to train a model, then every time someone uses it. That second part, called inference, grows with every new customer.
How sure are we?Official: from the company, regulator or authorsReported: from a named outletUnconfirmed: rumor or leak