Hard Fork Podcast Summary

Episode: Where Is All the A.I.-Driven Scientific Progress?
Released: December 26, 2025
Hosts: Kevin Roose (The New York Times), Casey Newton (Platformer)
Guest: Sam Rodriguez, Co-founder and CEO, Future House and Edison Scientific

Overview

This episode dives into the real state of AI-powered scientific advancement. While AI titans and government leaders talk up the promise of a new era—curing cancer, solving climate change, redefining scientific progress—there’s a gap between hype and reality. The hosts bring in Sam Rodriguez, a scientist and AI entrepreneur at the center of the field, to break down what AI can and can’t do in science today, to separate the wild projections from sober reality, and to forecast what’s coming next.

Key Discussion Points

Why Now? The AI-for-Science Hype Cycle

• AI leaders (Sam Altman, Demis Hassabis, Dario Amodei) claim AI is on the cusp of solving major scientific problems.
• Political initiatives like the White House’s Genesis Mission position AI as a national tool for accelerating discovery.
• Science has become a "main way" for tech leaders to justify risk and negative side effects:

  “Whenever one of their models does something horrible, the message…is, don't worry, we're about to cure cancer.” — Casey Newton (02:53)

What Is Cosmos? An “AI Scientist” in Practice

Cosmos is the latest "AI scientist" tool developed by Sam's Edison Scientific.

• Claim: Can accomplish 6 months’ worth of a human scientist’s work in a single run (08:29)

• How it works:

  • Input: A research objective and data set
  • Output: New scientific findings, often matching or exceeding what academics take months to find
  • Built on top of several LLMs (OpenAI, Google, Anthropic) and custom internal models
  • Uses a "structured world model" to manage long, complex goals

  “Cosmos is like the first thing…that actually really feels like an AI scientist…It comes back with insights that are actually really deep and interesting and sometimes wrong. But about 80% of the time, right.” — Sam Rodriguez (08:29)

• Cost: $200 per prompt—expensive, but justified by computational needs and the value created (11:56)

  “It uses a lot of compute…One run will write 42,000 lines of code and read 1,500 research papers on average.” — Sam Rodriguez (12:05)

• Novel Discoveries: Cosmos has already found net new scientific results (not just duplications), such as discovering mechanisms for genetic variants linked to Type 2 diabetes (14:04)

The Scientific Pipeline: How AI Changes Work

• Where does AI help? Primarily in data analysis and hypothesis generation (“drawing conclusions from gathered data”).

  “At this point, basically it's like step number three that Cosmos is aimed at.” — Sam Rodriguez (15:24)

• Validation: Human scientists still must deeply check AI’s findings and run experiments to confirm them (15:52)

• The Real Bottleneck: It's not the discovery, it’s testing—human clinical trials, finding subjects, manufacturing compounds, regulatory approval:

  “You have to run clinical trials...If we had a drug right now that prevented aging…you would not know for 10 years.” — Sam Rodriguez (25:44)

• AI’s value: Optimizing what experiments to run, and squeezing insight from data that would otherwise be ignored, but not bypassing the need for validation.

The Landscape: Where’s the Money, What’s Hot

• Two Frontiers:

  • Modeling the natural world (e.g. protein folding, molecule/antibody design, organism creation)
  • Modeling the process of doing science (AI agents that handle the workflow/end-to-end research steps)

• Big Advances:

  • Generative models: producing new proteins, antibodies, and even organisms from scratch

  “These are models that can produce examples of proteins or antibodies…that have desired characteristics basically from scratch. This is a new capability that we have never had before." — Sam Rodriguez (21:50)

  • Literature search and code generation: Massive unlocks for scientists who couldn’t previously analyze data at scale or write code.

Memorable Quotes & Moments

• On hype and reality:

  “A decade is crazy…if I'm wrong, everyone wins. But like a decade is crazy.” — Sam Rodriguez (25:33)

• On the pace of change:

  “I think that you'd be shocked to the extent that they have not yet [changed science]. Scientists in general are extremely conservative people…” — Sam Rodriguez (35:01)

• On serendipity and AI:

  “You almost want your AI scientist model to hallucinate a little bit so that it doesn’t lose that quality.” — Kevin Roose (24:52)

• On clinical trial bottlenecks:

  “Even with no regulation, it would be slow.” — Sam Rodriguez (27:55)

• On adoption:

  “Literature search…that’s a huge unlock that’s going to get adopted very quickly.” — Sam Rodriguez (35:01)

Overhyped or Underhyped? (Lightning Round, 36:14–39:02)

Sam gives his rapid-fire take on new frontiers:

• AI math proofs (IMO): Overhyped for science, but a good driver of AI progress
• Lab robotics: Appropriately hyped — “It is going to be totally transformative, but the tech isn’t there yet.”
• AlphaFold 3 and protein folding: Probably underhyped, transformative even amid the buzz
• “Virtual cells”: Overhyped—true virtual cell simulation is still very far away
• Quantum computing: Overhyped
• Brain-computer interfaces: Overhyped; much further out than the media suggests

2025 in Review and What’s Next

Top Three 2025 AI-Driven Advances: (39:14–40:39)

1. AI agents for science (Cosmos, Google Co-Scientists, etc.): Automation of complex reasoning and workflow
2. Generative design of antibodies: e.g., CHAI, NABLA—potential for instant design of new drugs
3. De novo design of organisms (ARK Institute): Creating new life forms in silico, though the practical use is uncertain

Looking at 2026:

• AI agents will experience “an explosion”—moving from pilot to pervasive in labs
• Sam predicts by 2027, most high-quality research hypotheses may originate from AI agents

  “2026 is going to be the year when we just see these agents start to infiltrate everything…It’s already happening.” — Sam Rodriguez (40:55)

Timestamps for Key Segments

• [07:39] — Sam Rodriguez introduction & background
• [08:29] — Cosmos: what it is, what it does
• [11:56] — The cost and computational demand of Cosmos
• [14:04] — Cosmos’ novel discoveries in genetics
• [17:19] — The real bottleneck: clinical trials, not discovery
• [20:16] — The lay of the land: two kinds of AI/science
• [21:50] — Generative models in science
• [22:46] — Reliability checks, AI “homework” for humans
• [23:25] — AI and the value of serendipity in science
• [25:33] — Reality check: AI won’t cure all disease in 10 years
• [27:55] — The slowness of scientific testing
• [35:01] — How much science has actually changed
• [36:14] — Lightning Round: Overhyped or Underhyped?
• [39:14] — Top 2025 breakthroughs and what’s coming next

Final Takeaways

• The breakthrough potential for AI in science is real, especially for analysis, hypothesis generation, and design in genetics and biology.
• Misconceptions persist: much of the public and even AI leaders underestimate the continued importance of experimentation and validation, as well as the conservative pace of scientific adoption.
• The near future will see the proliferation of “AI agents” in research—dramatically increasing what’s possible behind the scenes, though headline-grabbing disease cures remain a longer-term project.

Notable Quote:

“What AI will allow us to do is it will allow us to discover a lot of things where we already have the information…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.” — Sam Rodriguez (27:58)

If you’re a scientist, policymaker, or just a tech enthusiast, this episode delivers a grounded assessment of AI’s real-world impact on discovery—and a forecast for the “S-curve” yet to come.