Where Is All the A.I.-Driven Scientific Progress?
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
In this episode of Hard Fork, hosts Kevin Roose and Casey Newton explore the current state of AI-powered scientific discovery with Sam Rodriguez, CEO of FutureHouse and Edison Scientific. Rodriguez discusses his company's AI scientist tool called Kosmos, which can perform the equivalent of six months of doctoral-level research in a 12-hour run for $200. (06:05) The conversation covers the realities and limitations of AI in scientific research, from protein folding breakthroughs to the practical bottlenecks that prevent AI from immediately curing all diseases. Rodriguez provides a balanced perspective on both the genuine advances and overhyped claims in AI science.
- Main themes include the distinction between modeling the natural world versus modeling the scientific process, the role of AI agents in research, and the practical limitations of translating AI discoveries into real-world medical treatments
Speakers
Sam Rodriguez
Sam Rodriguez is the cofounder and CEO of FutureHouse (a nonprofit) and Edison Scientific (its for-profit spinoff), both focused on developing AI tools for scientific research. He holds a PhD in physics from MIT and previously spent several years running an applied biotech lab before launching his AI science ventures.
Kevin Roose
Kevin Roose is a technology columnist at The New York Times and co-host of Hard Fork. He covers the intersection of technology, society, and culture with a focus on artificial intelligence and its broader implications.
Casey Newton
Casey Newton is the founder of Platformer and co-host of Hard Fork. He is a technology journalist who covers social media, content moderation, and the broader tech industry ecosystem.
Key Takeaways
Clinical Trials Are the Real Bottleneck, Not Discovery
While AI can accelerate scientific discovery and hypothesis generation, Rodriguez emphasizes that the fundamental limitation in medical breakthroughs isn't the discovery phase—it's the clinical trial process. (15:13) Even if we had perfect drugs tomorrow, we'd still need years to test them safely in humans, recruit patients, and navigate regulatory approval. This reality check counters claims that AI will cure all diseases within a decade, as even detecting whether anti-aging treatments work would require years of observation in human subjects.
AI Excels at Finding Insights Hidden in Existing Data
Rodriguez's Kosmos system demonstrates AI's power in analyzing vast amounts of existing scientific data to uncover new insights. In one example, the system identified a genetic variant mechanism for type 2 diabetes by analyzing raw genetic data and connecting it to insulin secretion pathways. (11:57) This represents AI's sweet spot: not generating entirely new knowledge, but finding patterns and connections in data that human scientists haven't had time to explore due to sheer volume.
Scientists Are Conservative Adopters of New Tools
Despite AI's potential, Rodriguez notes that most working scientists remain conservative in adopting new methodologies. (31:23) This stems from the inherent uncertainty in scientific experiments where researchers often inherit protocols that work but don't fully understand why. Scientists prioritize reliable, proven methods over cutting-edge tools that might introduce variables. However, areas like coding assistance and literature search are seeing faster adoption because they address clear bottlenecks without disrupting core experimental protocols.
Generative Design Is Transforming Biological Research
The ability of AI to generate entirely new biological entities—antibodies, proteins, even organisms—represents a fundamental shift in how biological research operates. (36:25) Companies like Chai and Nabla are working toward a future where researchers can specify a disease target, click a button, and receive a custom-designed antibody ready for human testing. This "de novo" generation capability cuts out enormous amounts of traditional discovery work and represents one of 2025's biggest scientific breakthroughs.
AI Agents Will Soon Generate Most Scientific Hypotheses
Rodriguez predicts that by 2027, the majority of high-quality scientific hypotheses will be generated by AI agents rather than human scientists. (37:26) This represents a fundamental shift in the scientific process, where AI systems will identify promising research directions by analyzing vast datasets and literature. While humans will still validate and test these hypotheses, the initial creative and analytical work of hypothesis generation will increasingly be AI-driven, dramatically accelerating the pace of scientific inquiry.
Statistics & Facts
- Kosmos AI system writes 42,000 lines of code and reads 1,500 research papers on average per run, compared to traditional AI tools that might write only a few hundred lines of code. (10:04)
- Each Kosmos run costs $200, which Rodriguez notes is currently a promotional price that will increase in the future. Scientists typically spend $5,000-$10,000 gathering experimental data, making the AI analysis cost relatively modest. (10:30)
- The human genome contains millions of genetic variants associated with disease, but for most variants, scientists know they're associated with disease but have no idea why. (12:04)
