AI for Science Just Had Its ChatGPT Moment (and Scientists Aren’t Extinct)

Summary

In this episode of Attention Span, Ksenia Se argues that AI for science just had its “ChatGPT moment” — and uses one extraordinary week of breakthroughs as proof. An internal OpenAI reasoning model disproved Paul Erdős’s long-standing unit distance conjecture by pulling in unexpected concepts from algebraic number theory (infinite class field towers, Golod-Shafarevich theory) and applying them to a problem in geometry. Fields Medalist Tim Gowers called it “the first really clear example of AI solving not just an unsolved math problem but a really well-known unsolved math problem.”

On May 19th, Nature published a historic trio of peer-reviewed papers. Google DeepMind’s Co-Scientist — a multi-agent system that runs a tournament where sub-agents propose, critique, and Elo-rank ideas — proposed 30 leukemia drug candidates that were validated in real wet lab cell lines. FutureHouse’s Robin autonomously found a never-before-suggested treatment (Reposidil) for dry age-related macular degeneration. Google and Harvard’s ERA wrote scientific simulation code that beat the CDC’s own COVID-19 hospitalization forecasts.

Ksenia’s argument is that this is not the end of human science but the beginning of its most creative era. AI runs as a “co-scientist” in a lab-in-the-loop framework — it cannot run physical experiments and, crucially, cannot decide which questions are worth asking. Drawing on her own experience editing The Question, she insists curiosity is infinite: every good answer ignites ten new questions. By collapsing decade-long research cycles into hours or days, AI lets researchers take on the impossible — reversing cellular aging, stable fusion, microplastics cleanup — instead of burning a 40-year career on a single bottleneck.

Highlights

”The first really clear example of AI solving a really well-known unsolved math problem”

“the first really clear example of AI solving not just an unsolved math problem but a really well-known unsolved math problem.” — Ksenia Se (quoting Tim Gowers), 1:54

yt-dlp --download-sections "*1:39-2:30" "https://www.youtube.com/watch?v=WzoPLV3pYUs" --force-keyframes-at-cuts --merge-output-format mp4 -o "erdos-conjecture-cracked.mp4"

”30 drug candidates… actually validated in real wet lab cell lines”

“It proposed 30 drug candidates for acute myeloid leukemia, which were actually validated in real wet lab cell lines. At the exact same time, a nonprofit lab called Future House published Robin.” — Ksenia Se, 2:45

yt-dlp --download-sections "*2:45-3:30" "https://www.youtube.com/watch?v=WzoPLV3pYUs" --force-keyframes-at-cuts --merge-output-format mp4 -o "co-scientist-leukemia-drugs.mp4"

”They cannot decide what questions are actually worth asking”

“These systems are built as co-scientists. They operate in a lab-in-the-loop framework. They cannot run the physical experiments without us, and more importantly, they cannot decide what questions are actually worth asking.” — Ksenia Se, 3:48

yt-dlp --download-sections "*3:48-4:30" "https://www.youtube.com/watch?v=WzoPLV3pYUs" --force-keyframes-at-cuts --merge-output-format mp4 -o "lab-in-the-loop.mp4"

”A good answer does not satisfy curiosity. It ignites it.”

“A good answer does not satisfy curiosity. It ignites it. It gives you the vocabulary, the context, and the imagination to ask something even deeper, something you could not have even conceived of the day before.” — Ksenia Se, 5:03

yt-dlp --download-sections "*5:03-5:40" "https://www.youtube.com/watch?v=WzoPLV3pYUs" --force-keyframes-at-cuts --merge-output-format mp4 -o "curiosity-ignites.mp4"

”A deeper bucket to reach the water we’ve never been able to touch”

“But science is not a box. It’s an infinite unfolding landscape. AI is not draining the well of discovery. Quite the opposite. It’s giving us a deeper bucket to reach the water we’ve never been able to touch.” — Ksenia Se, 5:38

yt-dlp --download-sections "*5:38-6:20" "https://www.youtube.com/watch?v=WzoPLV3pYUs" --force-keyframes-at-cuts --merge-output-format mp4 -o "deeper-bucket.mp4"

Key Points

• A massive scientific revolution is underway (0:39) — Not because AI is replacing scientists but because it enables radical, mind-bending acceleration of discovery
• OpenAI cracks the Erdős unit distance problem (0:58) — An internal reasoning model disproved the long-standing belief that a simple square grid was optimal
• Unexpected cross-domain math (1:39) — The model brought infinite class field towers and Golod-Shafarevich theory from algebraic number theory into geometry
• Nature’s historic trio of peer-reviewed papers (2:04) — Published May 19th, showing AI breaking boundaries in biology and coding
• DeepMind Co-Scientist uses Elo tournaments (2:22) — Multi-agent system where sub-agents propose ideas, critique each other, and rank hypotheses using chess-style Elo ratings
• FutureHouse Robin finds a new blindness treatment (2:45) — Autonomously identified Reposidil for dry age-related macular degeneration, a drug never previously suggested for that condition
• Google + Harvard ERA beats the CDC (3:14) — Empirical Research Assistance designed COVID-19 forecasting models that outperformed the CDC’s own predictions
• Lab-in-the-loop framework (3:48) — AI handles algorithmic logistics; humans decide which questions are worth asking and run the physical experiments
• The Question editor’s fear of running out of questions (4:14) — A personal anecdote: she once worried people would exhaust curiosity, but every answer opened ten more questions
• Science is not a box, it’s an infinite landscape (5:19) — Critics assume science has a finish line; AI proves the opposite
• The real tragedy of science is time (5:54) — Researchers spend 40-year careers on one or two narrow puzzles, waiting on slow simulations and manual screening
• ERA collapses months into hours (6:38) — Writes the high-performance code that biologists and physicists usually spend months debugging
• Materials science bottleneck for climate (6:49) — Carbon capture and solid-state battery electrolytes traditionally take a decade of trial and error; Orbital Materials and CuspAI compress that
• Axiom Math attacks hard math problems (7:11) — Part of a wave of smart people leaving slow academic cycles to build acceleration-first systems
• No question will be structurally unsolvable (8:01) — Young researchers can take on impossible problems without fear they’ll consume a decade
• The dream list (8:09) — Reversing cellular aging, beating cancer, stopping climate change, stable nuclear fusion, cleaning ocean microplastics
• Thank you, Demis Hassabis (9:00) — Ksenia credits Demis Hassabis as the inspiration for the episode

Mentions

Companies & Labs

• OpenAI (0:58) — Internal general-purpose reasoning model that cracked the Erdős unit distance problem
• Google DeepMind (2:22) — Built Co-Scientist, the multi-agent leukemia drug system
• FutureHouse (2:45) — Nonprofit lab that published Robin, found blindness treatment
• Harvard (3:14) — Co-published ERA with Google for scientific code generation
• Orbital Materials (7:03) — Designs new materials at the atomic level using AI
• CuspAI / Cosmos AI (7:03) — Atomic-level material design startup mentioned alongside Orbital
• Axiom Math (7:11) — Company working on solving hard math problems with AI
• The Question (4:14) — The Q&A platform where Ksenia was editor-in-chief

Products & Systems

• Co-Scientist (2:22) — DeepMind’s multi-agent brainstorming system using Elo tournaments
• Robin (2:45) — FutureHouse autonomous agent that scans literature and proposes drugs
• Reposidil (3:00) — Drug Robin proposed for dry age-related macular degeneration
• ERA (Empirical Research Assistance) (3:14) — Google/Harvard AI for writing high-performance scientific simulation code

Concepts & Theories

• Unit distance problem (0:58) — Erdős’s long-standing question about maximizing unit-distance pairs in a plane
• Infinite class field towers (1:39) — Algebraic number theory concept the OpenAI model imported into the proof
• Golod-Shafarevich theory (1:39) — Another deep number-theoretic tool used in the proof
• Elo rating system (2:22) — Chess ranking method Co-Scientist uses to rank competing hypotheses
• Lab-in-the-loop (3:48) — Framework where AI handles compute and humans run physical experiments

People

• Paul Erdős (0:58) — Legendary mathematician who proposed the unit distance problem
• Tim Gowers (1:39) — Fields Medalist who called the OpenAI proof a watershed moment
• Demis Hassabis (9:00) — Credited as the inspiration for the episode; “he has never lost his curiosity”

Surprising Quotes

“An internal general purpose reasoning model from OpenAI just completely disproved that long-standing conjecture. Do you think it ran numbers? Yes, it did. But it also did much more.” — Ksenia Se, 1:25

“It designed forecasting models for COVID-19 hospitalizations that actually outperformed the CDC’s own prediction models.” — Ksenia Se, 3:28

“I used to wake up in the middle of the night staring at the ceiling, gripped by this bizarre, irrational fear. I was terrified that one day, people would just run out of questions.” — Ksenia Se, 4:29

“Biologists and physicists often have to spend months acting as mediocre software developers, debugging code just to test a single hypothesis.” — Ksenia Se, 6:23

“AI isn’t the end of science. It is the beginning of its most creative, open-ended era. We are not building machines to think for us. We are building machines to help us think further than we ever thought possible.” — Ksenia Se, 8:37

Transcript

Ksenia Se: 0:00 I want to ask you a few questions. Why is it dark at night? Why do we dream? What is consciousness? What existed before the universe? Do you know answers to all these questions? I don’t. And that’s why AI and science is such a big deal right now. And we just got a few incredible proofs of this. Curious?

Ksenia Se: 0:22 Hi, my name is Ksenia, welcome to Attention Span. For almost all of human history, these were the kind of questions we thought are unsolvable. The mysteries so vast that we just accepted we might never know the answers. But here’s the thing, we are currently living through a massive revolution in science that is about to change the timeline of discovery forever. And it’s not because AI is going to replace human scientists and make us obsolete, it’s about radical, mind-bending acceleration. And we just had a week of breakthroughs that proved it.

Ksenia Se: 0:58 Let’s start with the absolute bombshell that dropped from OpenAI. For decades, mathematicians have been stumped by a famous problem called the unit distance problem, proposed by the legendary Paul Erdos. The prevailing belief among the world’s smartest minds was that a simple square grid pattern was the absolute best way to maximize the number of unit distance pairs in a plane. Well, an internal general purpose reasoning model from OpenAI just completely disproved that long-standing conjecture. Do you think it ran numbers? Yes, it did. But it also did much more.

Ksenia Se: 1:39 It brought incredibly sophisticated, unexpected concepts from algebraic number theory, things like infinite class field towers and Golod-Shafarevich theory, and applied them to a completely different area of geometry. Fields medalist Tim Gowers called the proof… the first really clear example of AI solving not just an unsolved math problem but a really well-known unsolved math problem.

Ksenia Se: 2:04 And that wasn’t even the only massive news. On May 19th, the very prestigious magazine Nature published a historic trio of peer-reviewed papers showing that AI is breaking boundaries in biology and coding too. Google DeepMind unveiled Co-scientist, a multi-agent system that acts as a supercharged brainstorming partner. It uses a tournament-style of thinking where different virtual sub-agents propose ideas, critique each other, and rank their hypotheses using the same Elo rating system we use for chess players. It proposed 30 drug candidates for acute myeloid leukemia, which were actually validated in real wet lab cell lines. At the exact same time, a nonprofit lab called Future House published Robin.

Ksenia Se: 2:59 Robin was tasked with finding a treatment for dry age-related macular degeneration, the leading cause of blindness. It autonomously scanned the literature, identified a biological pathway, and proposed a drug called Reposidil. The crazy part, Reposidil had never been suggested for this condition before. And to tie it all together, Google and Harvard published ERA, Empirical Research Assistance. ERA is an AI designed to write the highly specialized, high-performance code scientists use to run simulations. It designed forecasting models for COVID-19 hospitalizations that actually outperformed the CDC’s own prediction models.

Ksenia Se: 3:40 I know one of the first thoughts might be, “Oh no, AI is building AI researchers that will eliminate the need to have human researchers and scientists.” But if you look closely, these systems are built as co-scientists. They operate in a lab-in-the-loop framework. They cannot run the physical experiments without us, and more importantly, they cannot decide what questions are actually worth asking. That part is still entirely, beautifully human. This whole worry about AI running out of things for humans to do reminds me of a personal memory from my own career.

Ksenia Se: 4:14 Years ago, I was a editor in chief of a popular platform called The Question. It was a site where anyone could ask anything, and our job was to find the absolute best, most qualified experts to answer them. When we first launched, I was incredibly stressed. I used to wake up in the middle of the night staring at the ceiling, gripped by this bizarre, irrational fear. I was terrified that one day, people would just run out of questions. I thought, “What if we answer everything? What if there is nothing left to ask?”

Ksenia Se: 4:49 But guess what? We never ran out of questions. In fact, the opposite happened. Every time an expert wrote a brilliant, definitive answer, it didn’t close the conversation. It opened ten new ones. A good answer does not satisfy curiosity. It ignites it. It gives you the vocabulary, the context, and the imagination to ask something even deeper, something you could not have even conceived of the day before.

Ksenia Se: 5:19 And that is the fundamental truth about human curiosity. We will never run out of questions. When people worry that AI will automate science and make human researchers obsolete, they are assuming that science has a finish line. They think that there is a box of mysteries, and once the AI ticks them all off, the scientist can go home. But science is not a box. It’s an infinite unfolding landscape. AI is not draining the well of discovery. Quite the opposite. It’s giving us a deeper bucket to reach the water we’ve never been able to touch.

Ksenia Se: 5:54 Right now, the real tragedy of science is time. Think about a brilliant researcher… They spend their entire 40-year career chasing the answer to one, maybe two incredibly specific puzzles. They spend decades writing code, waiting for slow physical simulations to run, hand-screening thousands of chemical compounds, let’s say. That is heroic work, but it’s also agonizingly slow.

Ksenia Se: 6:21 This is where a system like Era is a game changer. Writing scientific software is a massive bottleneck. Biologists and physicists often have to spend months acting as mediocre software developers, debugging code just to test a single hypothesis. Era collapses that timeline from months or years into mere hours or days. The AI handles the algorithmic logistics, allowing the scientist to focus on the actual science.

Ksenia Se: 6:49 It is the same story in materials science, which is a massive bottleneck for solving climate change. Finding a new material for carbon capture or a solid-state battery electrolyte traditionally takes a decade of trial and error. Startups like Orbital Materials and Cosmos AI are now using AI to design materials from the atomic level in a fraction of the time. There are companies like Axiom Math working on solving hard math problems. And more incredibly smart people are starting to leave the world of slow academic cycles and closed labs to build systems designed for acceleration. Thanks to AI, they now can do that.

Ksenia Se: 7:26 AI does not replace the scientist’s judgment. It amplifies it. It is our chance to stop spending a lifetime on a single bottleneck and start solving humanity’s most pressing crises in weeks.

Ksenia Se: 7:38 So what does the future actually look like? It looks like a world where we actively encourage people to ask more, bolder, and more ambitious questions. If you are a young researcher today, you don’t have to shy away from an incredibly complex, high-risk problem because you are afraid it will eat up ten years of your life with no guarantee of success. You can take on the impossible questions because with AI supporting the infrastructure of science, we have entered a time when no question will be structurally unsolvable. We can dream bigger. We can ask how to reverse cellular aging. We can fight cancer. We can learn how to stop climate change. How to build perfectly stable nuclear fusion reactors. How to clean microplastics from our oceans, knowing we finally have the computational horsepower to actually find the answers.

Ksenia Se: 8:32 That is the absolute beauty of this transition. AI isn’t the end of science. It is the beginning of its most creative, open-ended era. We are not building machines to think for us. We are building machines to help us think further than we ever thought possible.

Ksenia Se: 9:00 Father. Thank you, Demis Hassabis, for inspiring me to record this video. That is all for today. Thank you for watching Attention Spen. Let’s keep this conversation going in the comments. If you had an AI co-scientist by your side today, if you dream about having, doing more science projects, what unsolvable questions would you want to ask first? I’ll see you next week.