A.I. Scientists Are Here. But Is Progress Accelerating? | EP 170

Summary

Kevin Roose and Casey Newton interview Sam Rodriques, the founder and CEO of FutureHouse and co-founder of Edison Scientific, about the current state of AI in scientific research. Sam’s company recently released Kosmos, an AI agent that he says can accomplish six months of doctoral or postdoctoral-level research in a single 12-hour run. The conversation separates the hype from reality around AI’s impact on science, with Sam offering a surprisingly grounded perspective despite being an optimist about the technology.

Sam explains that Kosmos works by orchestrating hundreds of AI agents that contribute to a structured “world model” built up over time, allowing them to maintain coherence on complex tasks. At $200 per prompt, the tool reads 1,500 research papers and writes 42,000 lines of code per run. The episode covers a fascinating distinction between modeling the natural world (like protein folding and organism design) and modeling the process of doing science (what FutureHouse does). Sam pushes back on claims that AI will cure all diseases within a decade, arguing that clinical trials are the real bottleneck, not research methodology. He predicts that by 2027, the majority of high-quality scientific hypotheses could be generated by AI agents.

Highlights

”A Decade Is Crazy for Curing All Diseases”

yt-dlp --download-sections "*22:30-23:30" "https://www.youtube.com/watch?v=Dn4OWufAggk" --force-keyframes-at-cuts --merge-output-format mp4 -o "Dn4OWufAggk-22m30s.mp4"

“Decade is crazy. If we had a drug right now that prevented aging, completely halted aging in humans between the ages of 25 and 65, you would not know for 10 years because you can’t detect in humans in that age range whether or not they’re aging for at least five or 10 years.” — Sam Rodriques, 22:30

”Kosmos Writes 42,000 Lines of Code Per Run”

yt-dlp --download-sections "*9:00-10:00" "https://www.youtube.com/watch?v=Dn4OWufAggk" --force-keyframes-at-cuts --merge-output-format mp4 -o "Dn4OWufAggk-9m00s.mp4"

“An individual run from Kosmos will write 42,000 lines of code and read 1,500 research papers on average. If you run Claude, it might write like a few hundred lines of code. So that gives you some sense.” — Sam Rodriques, 9:00

”You Should Not Expect to Just Ask GPT-7 How to Cure Alzheimer’s”

yt-dlp --download-sections "*25:00-26:00" "https://www.youtube.com/watch?v=Dn4OWufAggk" --force-keyframes-at-cuts --merge-output-format mp4 -o "Dn4OWufAggk-25m00s.mp4"

“You should not expect that you’re one day going to get GPT-7 and just ask it how to cure Alzheimer’s and it will just tell you. We do not have enough knowledge to solve it in principle even with infinite intelligence.” — Sam Rodriques, 25:00

”Majority of Good Hypotheses Made by Agents by 2027”

yt-dlp --download-sections "*34:00-35:00" "https://www.youtube.com/watch?v=Dn4OWufAggk" --force-keyframes-at-cuts --merge-output-format mp4 -o "Dn4OWufAggk-34m00s.mp4"

“I was telling people that I thought in 2026 or maybe 2027 that the majority of the high quality hypotheses generated by the scientific community would be generated by us or by agents like the ones we’re building. When I said it in 2024, I thought I was overhyping. At this point, it may be real.” — Sam Rodriques, 34:00

Key Points

• AI Science Hype vs. Reality (0:12) - Leaders like Dario Amodei, Sam Altman, and Demis Hassabis have been making bold claims about AI curing diseases; Kevin and Casey bring in a scientist to evaluate these claims
• Genesis Mission (1:28) - The White House announced a national effort to use AI to accelerate scientific innovation and discovery
• Kosmos AI Scientist (3:41) - Edison Scientific’s tool can accomplish six months of PhD-level research in a 12-hour run, costing $200 per prompt
• Six-Month Claim Validated (5:00) - Tested by giving Kosmos the same research objectives as academic collaborators; it found the same results overnight that took researchers 3-6 months
• Structured World Model (7:00) - Key innovation is a structured world model that coordinates hundreds of agents working in parallel toward coherent goals
• 42,000 Lines of Code (9:00) - Each Kosmos run writes 42,000 lines of code and reads 1,500 research papers
• Type 2 Diabetes Discovery (10:30) - Kosmos identified a novel mechanism linking a genetic variant to insulin secretion via the SRI1 gene in the pancreas
• Two Categories of AI Science (16:00) - Modeling the natural world (protein structures, antibodies) vs. modeling the process of doing science (what FutureHouse does)
• Clinical Trials Bottleneck (20:00) - The real limitation isn’t research speed but finding patients, manufacturing drugs, and running trials
• Decade Is Crazy (22:30) - Sam says curing all diseases in a decade is unrealistic; 30 years is plausible
• Serendipity in AI Science (26:00) - AI probably will preserve scientific serendipity; first-year grad students who do random things are an important source of progress
• Scientists Still Conservative (33:00) - Most labs around the world are still doing science the way they’ve done it before; adoption is slow
• Overhyped/Underhyped Lightning Round (28:00) - Vibe proving: overhyped; Lab robotics: appropriately hyped; AlphaFold 3: underhyped; Virtual cells: overhyped name; Quantum computing: overhyped; BCIs: overhyped

Mentions

Companies

• FutureHouse (3:41) - Sam Rodriques’ nonprofit building AI tools for scientific research
• Edison Scientific (3:41) - For-profit spinoff of FutureHouse that released Kosmos
• OpenAI (7:00) - One of the frontier model providers Kosmos builds on
• Google/DeepMind (7:00) - Provides models for Kosmos; also working on Google co-scientist
• Anthropic (7:00) - One of the model providers used by Kosmos
• Chai Discovery (30:00) - Company doing de novo antibody design
• Nabla (30:00) - Company working on de novo antibody design
• Arc Institute (30:00) - Designed a bacteria phage (virus) from scratch; working on virtual cells
• New Limit (28:00) - Working on biological models similar to Arc Institute
• Chan Zuckerberg Initiative (28:00) - Also building biological models
• Neuralink (29:00) - Making progress on BCIs but further out than people imagine

Products & Technologies

• Kosmos (3:41) - Edison Scientific’s AI scientist that runs for 12 hours per research query
• AlphaFold 3 (28:00) - Protein structure prediction model; Sam says probably underhyped
• Claude Code (33:00) - Mentioned as a huge unlock for biologists who didn’t know how to code

People

• Sam Rodriques (3:41) - CEO of FutureHouse and co-founder of Edison Scientific; PhD in physics from MIT
• Dario Amodei (0:30) - Referenced as making bold claims about AI curing diseases
• Sam Altman (0:30) - Referenced for similar claims about AI and science
• Demis Hassabis (0:30) - Referenced for claims about AI-driven scientific breakthroughs
• Patrick Collison (28:00) - Referenced in connection with Arc Institute’s virtual cell work
• Brian Hie (30:00) - Arc Institute researcher who designed organisms from scratch

Surprising Quotes

“Scientists who have used Cosmos generally come back to me and are like they can’t believe we’re only charging $200 for it. $200 right now is a promotional price. We actually have to eventually charge more.” — Sam Rodriques, 9:30

“You should not expect that you’re one day going to get GPT-7 and just ask it how to cure Alzheimer’s and it will just tell you. There is not enough knowledge. We do not have enough knowledge to solve it in principle even with infinite intelligence.” — Sam Rodriques, 25:00

“You’d be shocked to the extent that AI tools have not yet changed the life of a working scientist. Scientists in general are extremely conservative people.” — Sam Rodriques, 33:00

“First year graduate students have no idea what to do. And that is a huge source of scientific progress because they just do the most random kooky stuff that no one who knows anything would ever think to do.” — Sam Rodriques, 27:00

Transcript

0:00 I’m Kevin Roose, a tech columnist at the New York Times. I’m Casey Newton from Platformer and this is Hard Fork. This week, FutureHouse CEO Sam Rodriques joins us in the studio to separate the hype from the reality of AI science.

0:12 I have been obsessed with this question of what AI is and isn’t doing for science and scientific discovery. People like Dario Amodei, Sam Altman, Demis Hassabis have all been saying things about how close they believe we are to solving new scientific problems and curing diseases. Some of that is obviously hype. But there’s actually a lot of real stuff going on in AI and science that I just do not feel personally qualified to evaluate.

1:03 Science has become one of the main ways that the leaders of these tech companies want us to evaluate them. Whenever one of their models does something horrible, the message we basically get back is, “Don’t worry, we’re about to cure cancer. Just hang on tight.”

1:28 The Genesis mission was announced by the White House just before Thanksgiving — a dedicated, coordinated national effort to unleash a new age of AI, accelerate innovation and discovery.

3:41 Sam Rodriques is the co-founder and CEO of FutureHouse and Edison Scientific, based in San Francisco. FutureHouse is the nonprofit; Edison Scientific is the for-profit spinoff. They’re building what Sam calls an AI scientist. Sam has a PhD in physics from MIT and previously ran an applied biotech lab.

5:00 When I got that six-month number, my reaction originally was there is no way that this is true. We had academic collaborators who had done science previously that they had not published yet. We gave the same research objective and data set to Kosmos and it found the same things the researchers had found, overnight. The researchers said it took them 3 months, 5 months, 6 months.

7:00 Kosmos is a box you type into. It runs for 12 hours before coming back with findings. It builds on models from OpenAI, Google, and Anthropic, plus their own internally trained models. The key insight is a structured world model that coordinates hundreds of agents running in parallel toward a coherent goal.

9:00 Each run costs $200 per prompt because it uses enormous compute. An individual run writes 42,000 lines of code and reads 1,500 research papers. Scientists who have used it can’t believe they’re only charging $200; it’s actually a promotional price.

10:30 Kosmos identified a novel genetic variant mechanism for type 2 diabetes. The variant was not in a gene, but Kosmos identified that a protein binds there and connected it to the SRI1 gene involved in the pancreas secreting insulin.

16:00 Sam explains two categories of AI for science: modeling the natural world (predicting protein structures, generating antibodies, creating organisms) and modeling the process of doing science. Both are fundamentally different. Generative models for biology are the most exciting trend — designing proteins and antibodies from scratch.

20:00 Even if you have a molecule to test, you have to make it, ensure it’s high enough grade, find patients, form relationships with doctors, wait for enough willing patients. With no regulation, even with no regulation, it would be slow.

22:30 A decade for curing all diseases is crazy. If we had a drug that halted aging in humans ages 25-65, you wouldn’t know for 10 years. 30 years is very plausible for a humongous leap forward. The claims from AI lab CEOs aren’t in bad faith — reasonable people could disagree, and there’s a mixture of genuine excitement and not fully understanding clinical realities.

26:00 On serendipity: first-year graduate students have no idea what to do, and that randomness is a huge source of scientific progress. You almost want your AI scientist model to hallucinate a little bit — adding noise is important for evolution and discovery.

28:00 Overhyped/Underhyped lightning round: Vibe proving — overhyped. Lab robotics — appropriately hyped. AlphaFold 3 — underhyped. Virtual cells — overhyped (the name, not the models). Quantum computing — overhyped. Brain-computer interfaces — overhyped, further out than people imagine.

30:00 Top three AI-driven scientific advancements of 2025: AI agents for science (including Kosmos and Google co-scientist); de novo antibody design by Chai Discovery and Nabla; and the Arc Institute’s generation of organisms from scratch.

33:00 You’d be shocked to the extent AI tools have not yet changed the life of working scientists. Scientists are extremely conservative. The biggest immediate unlock is coding — biologists who didn’t know how to code can now do a lot using Claude Code and similar tools. Literature search is another huge unlock.

34:00 2026 will be the year when AI agents start to infiltrate everything. Sam predicts that by 2026-2027, the majority of high-quality scientific hypotheses could be generated by AI agents.