Wednesday, September 9, 2026HotTea archive editionVerified 2:39 AM PDT

The lead

OpenAI says 10,000 agents solved a $1 million math problem

An unreleased model produced a proof that a smooth fluid can develop a singularity. Lean accepted the formal version, but mathematicians are still reviewing the claim and disputing who deserves credit.

Read the full story →Listen to the edition

Quanta Magazine ↗Tristan Buckmaster, New York University ↗

Listen to this edition

Prefer audio? The briefing has chapters and a full transcript.

Follow on Apple Podcasts Follow on Spotify Transcript All episodes →

The briefing

The rest of the morning

4 more stories

Models, products and companiesHot

Meta gave Muse a browser, a way to pay and a place inside WhatsApp

Meta launched Muse in the United States for adults on September 8. People can use a separate app, a website or WhatsApp to give it work. Meta says Muse can browse, fill forms, send email, book travel, negotiate and keep working after the app closes. It can also make purchases through Stripe's Link system, which creates a one-time card number. Meta says the agent stops for approval before it sends an email or makes a purchase. The agent and user data live inside a dedicated cloud computer that Meta calls Muse Secure VM.

Meta's safety case comes from the company, not an independent security review. Meta says a separate system called Sentinel controls every outside connection, and the main agent cannot override it. Meta has not published a full outside audit or a rate for successful bypasses. The company has also not shown how Muse handles conflicting instructions from websites and connected apps. Muse becomes more useful as people give it more private data. A security failure would then expose more of that data.

Meta ↗Associated Press via PBS News ↗

Read article →

Infrastructure, energy and macroeconomicsHot

Washington offered up to $1.9 billion to restart a nuclear plant as AI power demand grows

NextEra Energy says it closed a U.S. Department of Energy loan of up to one point nine billion dollars to restart Iowa's Duane Arnold nuclear plant. The reactor stopped operating in 2020. NextEra plans to return the 615-megawatt plant to service by the first quarter of 2029, subject to regulatory approval. TechCrunch reports that Google has a power agreement tied to the restart and is considering data centers nearby. The loan follows federal support for another reactor restart connected to Microsoft.

The loan is worth up to one point nine billion dollars. That is not a final construction bill or proof that the plant will reopen on schedule. The Nuclear Regulatory Commission still has to approve the restart. NextEra's estimates for jobs, taxes and economic benefits come from a company-backed study. The Energy Department says the plant will lower electricity costs, but its announcement does not explain the rate design. It also does not say who bears the risk if costs rise or the work runs late.

NextEra Energy ↗TechCrunch ↗Reuters ↗

Read article →

Companies, markets and AI geopoliticsHot

Samsung led Mistral's €3 billion round to build a European AI system

Mistral announced a three billion euro Series D at a post-money valuation above twenty-one billion euros. Samsung Electronics led the round. The European Union-backed Scaleup Europe Fund and existing investor PSG Equity joined as co-leads. Mistral says the money will fund model research, compute capacity, infrastructure and international sales. Its chief financial officer told Reuters that Mistral now has more than 125 enterprise customers. The executive said Mistral is on track for one billion dollars in annual recurring revenue by year-end.

Mistral supplied the revenue target, customer count and claim that this is Europe's largest private technology equity round. Raising a large round does not prove that its models can close the performance, distribution or compute gap with larger U.S. rivals. Mistral sells European control, but still relies on chipmakers and investors around the world. Samsung's lead role adds another foreign dependency.

Mistral ↗Reuters via MarketScreener ↗

Read article →

Science, medicine and societyHot

DeepMind precomputed every possible one-letter mutation in the human genome

Google DeepMind released AlphaGenome Atlas on September 8. The database contains model predictions for about nine billion possible single-letter changes across the human genome. DeepMind says the data occupies about one petabyte and covers both protein-coding DNA and the much larger noncoding portion that helps regulate genes. Researchers can search the atlas without running the model themselves. DeepMind also released one score that ranks each variant by its predicted biological effect. Access is free for noncommercial research.

The atlas contains predictions, not nine billion lab results. DeepMind says collaborators validated selected examples, but the company also states that AlphaGenome is not approved for clinical use. Nature quoted outside researchers who called the resource useful while warning that experiments and patient details still matter. A single impact score can help rank variants, but it can also hide which model assumption drove the ranking.

Google DeepMind ↗Nature ↗Scientific American ↗

Read article →

Analysis

A machine-checked proof can still leave the hardest questions open

OpenAI's Navier-Stokes claim shows how quickly models can produce formal work. It also shows that reviewing the proof, tracing where it came from and settling credit must keep pace with the compute.

Lean checks the statement it receives

A formal checker can verify each encoded step. Mathematicians still have to confirm that the encoded theorem matches the problem and that the assumptions are acceptable.

Parallel compute changes who can compete

A private lab can run thousands of research paths at once. Most academic teams cannot match that budget, even when their earlier ideas supply the route.

Provenance becomes part of the result

Dated prompts, run records and public proof files can show when an idea entered a system. Without them, credit disputes become harder to settle after a fast release.

The watchlist

Signals that could change the read

Independent mathematicians rebuild the Lean artifact, compare it with the written proof and state whether it resolves the official problem.

Outside testers publish prompt-injection and approval-bypass results, while Meta reports any incidents tied to messages, payments or connected apps.

The Nuclear Regulatory Commission posts a restart schedule and conditions, while utility filings disclose cost and customer-rate protections.

Mistral reports audited revenue and new compute capacity, while AlphaGenome users publish lab checks across rare disease and population studies.

Across the desks Proof gets expensive

AI systems can now produce claims and artifacts faster than people can review them. The scarce work is checking the result, tracing the credit and deciding what deserves trust.

Navier-Stokes run10,000 agentsQuanta reports that OpenAI's internal model used about 10,000 agents for 88 hours before another model formalized the proof in Lean.

Muse approvalsEmail and payment checksMeta says Muse pauses for approval before sending an email or making a purchase, while a separate system screens outside connections.

Iowa nuclear loanUp to $1.9BNextEra says the federal loan will support a planned restart of the 615-megawatt Duane Arnold plant, subject to regulatory approval.

Mistral round€3B raisedMistral says Samsung led a Series D that values the French company above €21 billion after the investment.

Editorial direction, not a financial index. Each signal is tied to this edition’s reporting.

Edition validated · 5 stories · 12 unique sources

About HotTea & our sources →