OpenAI says an unreleased model solved the Navier-Stokes existence and smoothness Millennium Problem in 88 hours. - The New York Times
An OpenAI model reportedly proved the Navier-Stokes equations can produce mathematical blowups, extending mathematicians' earlier Euler equations breakthrough amid a credit dispute. - Scientific American
OpenAI claimed to solve the Navier-Stokes Millennium Prize problem but faces accusations of using uncredited work by mathematicians Buckmaster and Alpoge. - MIT Technology Review
They said data sharing is opt-in, but one wonders how many researchers opted in, was there any sort of incentive pattern. the devil is in the details. But if you didn't go through this program and just downloaded Codex then sharing is opt-out.
The announcement specifically noted they used the latest checkpoint from a training run they starting 8/28. the training data cutoff would have been recent. if scientists didn't opt out, their data would be in there. Clearly there was recent training data from real mathematicians working on this sort of problem and potentially from the Buckmaster team if they ran Codex without opt-out.
OpenAI ends up looking like a giant research organization with unlimited compute that also happens to sell the tools scientists use—and can turn around and race those same scientists to publication. OpenAI makes it sound as if they wanted talks to avoid stomping on the smaller team’s result, but clearly that’s not what the other team thought or the way it played out. They come across as disingenuous and careless and too-clever-by-half.
It’s probably true that if the agents are this good, they wait a week and Anthropic scoops their result.
Anyway, scientists and others need to be careful about opting out, for sure. What does this mean for science, that you point your agent swarm at a Millennium Prize problem and get a result in a few days? Amodei’s prediction of 100 years of scientific progress in 10 years is looking more accurate.
Ask your favorite AI what are the top 10 messy engineering problems that might yield to a billion dollars worth of AI compute and how they might change the world. Quantum computing, fusion power, lightweight solid-state batteries, other new materials, catalysts, drugs. Strange times.
OpenAI chief scientist Jakub Pachocki says AI is nearing self-improving code and urges labs to slow down until safety standards emerge. - SiliconANGLE
We have some concerns: Anthropic's Alignment Science lead says there is over a 10% chance AI kills all humans within a decade, citing self-improvement risk. - X (formerly Twitter
10% chance of annihilation, please give us $100b in an IPO to proceed as quickly as possible, risk it for the biscuit.
Mathematician Terence Tao warns no lab has a serious plan for managing AI risk. - The Potential Surface
Gary Marcus pivoting from “current AI is a dead end“ to, “it’s too powerful and risky, and I predicted it all along”
DeepMind's AlphaGenome Atlas predicts effects of all 9 billion possible DNA variants, aiding rare-disease and UK Biobank genetic research - Google DeepMind
Pentagon AI chief Cameron Stanley says NATO and the Five Eyes intelligence alliance lack resources to match US AI capabilities. - the Guardian
U.S. Commerce Department takes minority equity stakes in D-Wave, Rigetti, and Quantinuum via a $300 million CHIPS Act deal. - The Wall Street Journal
Meta launched Muse, an AI agent that autonomously sends emails, books travel, and connects to Instagram, Facebook, and third-party apps. - The New York Times
Muse can access personal data, web credentials and payment details, checking out purchases through a Stripe-built wallet called Link - Fast Company
An Ohio man got 15 years, the first conviction under the Take It Down Act, for extorting six women with AI-generated explicit images - BleepingComputer
OpenAI, Amazon and Google are scouting Patagonia, Argentina for AI data centers, drawn by cold climate and cheap energy. - Tom's Hardware
Google plans to invest at least €13 billion in AI infrastructure in Finland over the next two years. - Bloomberg
Task-specific AI models fine-tuned on business data cost less per token and outperform generic large models, with Gartner predicting triple the adoption by 2027 - Business Insider
"your tech is eavesdropping on you for marketing purposes" has crossed from "false conspiracy theory" to "true and concerning"
AI correctly identified 90% of anonymous Hacker News users by matching post details to public LinkedIn profiles - The Wall Street Journal
Researchers argue fully autonomous AI diagnosis could cut US misdiagnosis rates as high as 15%, despite AMA pushback - The Atlantic
Amazon-owned Zoox is challenging Waymo's San Francisco robotaxi dominance with novelty rides, distinctive design, and community marketing rather than scale. - The New York Times
DeepMind found 14% of 100 AI agents cheated on math proofs via a shared exploit, though some agents tattled on the cheaters - TNW | Artificial-intelligence
A World Bank report says AI could compress a century of development into a decade for poor countries, but only with fast government action. - Transformer News
Follow the latest AI headlines via druce.ai on Bluesky
