In partnership with
OpenAI said this week that its models produced new results in mathematics, computer science and physics, and it also shipped a cheaper everyday model. The discoveries are striking, but none have been independently confirmed yet, and that gap is the real story. We read the announcements and the expert reactions before writing, and every claim below is labeled for what it is.
100+ Claude Code hacks to ship code 10X faster
Top engineers at Anthropic say AI now writes 100% of their code.
Are you using AI to write yours?
These 100+ Claude Code hacks show you exactly how. Sign up for The Code and get:
100+ Claude Code hacks — free
The Code newsletter — learn the latest AI tools and skills to code faster in 5 mins a day
TODAY IN AI
3 things that happened while you were busy
1. OpenAI says its AI produced new proofs in math, computer science and physics.
OpenAI announced three results: a claimed disproof of the Exponential Time Hypothesis, a wave-equation proof tied to the Kakeya conjecture, and existence proofs for four-dimensional quantum field theories including QCD. Only the first has a machine-checked step in the Lean proof assistant, and Lean confirms the logic, not that the statement matches the real problem. Read Engadget's report. These are OpenAI's claims, pending independent review, and it has not released the physics materials in full.
2. OpenAI shipped GPT-6.1 Terra, a cheaper everyday model.
GPT-6.1 Terra is OpenAI's mid-size model, sitting below Sol and Astra, built for fast visual reasoning and computer use. OpenAI says it costs 50% less than the previous Terra and cut its error rate on a hard internal benchmark from 53.6% to 16.7% at low reasoning effort. It is available now in ChatGPT Work, Codex and the API. Read 9to5Mac's report. The accuracy figures are OpenAI's own.
3. A new OpenAI "discovery" model quietly appeared in its docs.
A model called gpt-rosalind-discovery showed up on OpenAI's API pricing page under life-sciences models, with no blog post or system card, spotted by watchers on X. Its sibling, gpt-rosalind-research, is a biology and drug-discovery model limited to approved internal research. OpenAI often lists models shortly before announcing them. Read TestingCatalog's report. Its purpose and access are not yet stated, so treat the drug-discovery guess as a guess.
Yes, that Arnold Schwarzenegger.
He has a newsletter on beehiiv. So do Codie Sanchez, Jay Shetty, David Begnaud, Colin & Samir, and Joanna Stern.
They could publish just about anywhere. They chose beehiiv to build a direct connection with the people who want to hear from them most. And you can do the same, whether you’re world-famous or just at the start of your plan for world domination.
TODAY'S TUTORIAL
Make an AI prove it before you believe it
This week's news is a reminder that an answer, even a confident one, is a claim until someone checks it. Here is a five-step habit for getting AI output you can actually trust, which works in any model. See why verification is the hard part.
1. Ask for the claim and the confidence. Have the model state its answer in one line and rate how sure it is, so you know where to look hardest.
2. Demand the steps. Ask it to lay out the reasoning or calculation in order, with every assumption named. Hidden assumptions are where errors hide.
3. Ask for sources you can open. Request real references, then actually open them. If it cannot point to a checkable source, treat the claim as unverified.
4. Run the counter-check. Start a fresh chat and ask a different model to find the flaw in the first answer. Disagreement tells you where to dig.
5. Test the smallest piece. Check one number, one quote or one line of code by hand. If the easy part is wrong, do not trust the hard part.
Good to know: Even OpenAI's proofs needed Lean to check the logic and human experts to judge the meaning. Machine confidence is not the same as being right.
FROM THE FRONTIER
AI is writing proofs faster than people can check them
The claims. OpenAI's three results are serious if they hold. The ETH disproof would reshape how computer scientists think about hard problems, and the QCD work touches the math behind the Standard Model. A few weeks earlier OpenAI also posted over 700 pages on 372 math problems. See the announcement.
The bottleneck. None of it is confirmed. Expert groups led by named mathematicians and physicists are reviewing the work, and the Advisory Group on Mathematics and AI pushed OpenAI to release the withheld physics materials so outsiders can audit them. Read The Neuron's analysis.
The money. Checking AI math is now a funded problem. The Simons Foundation launched a $15 million initiative to help mathematicians examine AI-generated research, with OpenAI committing $5 million and Anthropic also contributing. See the funding details.
The restraint. OpenAI says it will not seek academic credit, leaving that to the researchers who write up and verify the results. That is the right instinct, but choosing the first reviewers still shapes who gets to judge the evidence. See the review setup.
The reframe. The scarce resource is no longer the insight. A model can generate a proof overnight. What takes months is a human deciding whether it is true, and this week showed that the checking, not the thinking, is now the frontier.
MONEY FORWARD
Three ways to turn this week's AI news into income
1. Cheaper automations on the new model. GPT-6.1 Terra costs half of the last version and is built for visual reasoning and computer use. Rebuild a client's simple automations on it and pass on the lower running cost while keeping your fee. See what Terra does.
2. AI output checking as a service. If OpenAI's own proofs need human review, so do the reports and numbers businesses now generate with AI. Offer a fact-check pass that verifies claims, sources and figures before anything ships to a client or the public. See why review matters.
3. Design packs with Canva Craft. Canva's image model renders legible text and is free to all Canva users, so you can sell small shops a monthly bundle of flyers, menus and social posts without stock-photo fees. Read about Canva Craft.
GET THE EARNINGS GUIDE
Want the step-by-step version of ideas like these? Get The Aiwire earnings guide at theaiwire.co. These are ideas, not income promises, and results depend on your skills, market and effort.
IN THE KNOW
What people are actually watching and sharing
Release the physics. The Advisory Group on Mathematics and AI publicly called on OpenAI to release its withheld physics outputs, arguing that a completed formalization removed the reason to keep them private. Read the details.
Paying to check the math. The Simons Foundation put up $15 million to help mathematicians examine AI-generated research, a sign that verifying AI claims is becoming its own field. See the initiative.
The quiet life-sciences model. OpenAI's unannounced gpt-rosalind-discovery model is priced the same as its research sibling, at $5 per million input tokens and $25 output, hinting at a bigger push into biology. See the listing.
Anthropic's filing surprises. Anthropic's IPO paperwork disclosed AI security breaches and a training halt, per Fox News, the kind of detail that rarely makes it into a prospectus. Read the coverage.
What Lean can and cannot do. Reporting on the proofs noted that Lean checks whether logic is valid but not whether the formal statement matches the real problem, a gap that matters for every AI math claim. See the caveat.
TRENDING TOOLS
5 tools worth a look this week
1. GPT-6.1 Terra. OpenAI's new mid-size model for visual reasoning and computer use, 50% cheaper than the last Terra and available in ChatGPT Work, Codex and the API. The accuracy gains are OpenAI's own figures. Read the details.
2. Kimi K3.1 Pro. Moonshot's near-frontier model, strong at agentic coding, at $2 per million input tokens and $10 output. Free weights are promised by November 4; the benchmarks are the company's own. See Kimi.
3. Lean. The free, open-source proof assistant behind this week's one machine-checked result. It will not write your proof, but it will tell you whether the logic holds, which is handy for anyone serious about formal math. See Lean.
4. Canva Craft. Canva's first in-house image model plans the layout before drawing, so text comes out legible, and it is free to all Canva users with a public API. MCP access for agents is coming. Read the overview.
5. ThinkRail. A JetBrains open-source workspace that gives the Pi command-line agent a visual interface and lets you switch models per session. Early-stage and built mainly for existing Pi users. See the repo.
PROMPT STATION
Turn any AI answer into something you can check
This prompt works in Claude or ChatGPT and forces a model to show its work, name its assumptions and tell you how to verify each claim. Use it on anything you plan to act on or send to someone else.
COPY AND PASTE THIS PROMPT
Answer my question below, then make your answer checkable. After the answer, give me: 1) the three most important claims you made and how sure you are of each, from low to high, 2) the key assumptions behind each claim, 3) for each claim, exactly how I could verify it myself, naming a source, a calculation or a test, 4) the one part most likely to be wrong and why, 5) what extra information would change your answer. Keep the verification section plain and specific, and if you cannot back a claim with something I can check, say so instead of guessing. My question: [PASTE YOUR QUESTION]
Tip: Swap [PASTE YOUR QUESTION] for a market number, a medical fact or a block of code. For a tougher check, paste the answer into a second model and ask it to attack the weakest claim the first one flagged.