AI Accelerates Scientific Discovery but Clinical Trials Remain Bottleneck
TL;DR
- AI agents can accelerate scientific discovery by completing six months of a PhD or postdoctoral scientist's work in a single run, significantly compressing research timelines.
- Generative models for designing novel proteins and antibodies from scratch represent a transformative capability, drastically reducing the time and effort required for drug development.
- AI's ability to process vast datasets and identify novel mechanisms, such as connecting genetic variants to specific gene expression and disease pathways, unlocks new scientific insights.
- While AI can accelerate hypothesis generation and data analysis, the fundamental bottleneck for medical breakthroughs remains the lengthy and complex process of clinical trials.
- The adoption of AI tools by scientists is slow due to the conservative nature of scientific methodology, with coding and literature search being early areas of significant impact.
- AI's potential for serendipitous discovery, akin to accidental breakthroughs, is likely to be preserved, as errors and "hallucinations" can lead to unexpected and valuable findings.
Deep Dive
AI is poised to accelerate scientific discovery by modeling both the natural world and the process of scientific inquiry itself, but widespread adoption will be gradual due to the conservative nature of scientific practice. While AI agents are rapidly advancing, the ultimate realization of AI-driven cures and scientific breakthroughs will be constrained by the inherent time scales of experimental validation, clinical trials, and the fundamental limits of current human knowledge, rather than purely technological limitations.
The core of AI's impact on science lies in its ability to model complex systems, whether predicting protein structures or simulating scientific workflows. Sam Rodriguez, CEO of Future House and Edison Scientific, highlights the development of "AI scientists" like Cosmos, which can perform research tasks equivalent to months of human work in a single run. This is achieved through sophisticated agent orchestration and a "structured world model" that maintains context and coherence over long, complex tasks. While these tools can replicate existing research rapidly and even generate novel hypotheses, the process of scientific discovery and validation remains inherently human-intensive. The "six-month" equivalence is measured against the time human researchers previously took to reach similar conclusions, not the absolute time required for discovery.
However, the path to AI-driven scientific revolution is not solely a technical one. The inherent conservatism of scientific fields, particularly biology, means that new methods are adopted slowly. While coding and literature search tools are seeing rapid adoption due to their ability to overcome existing bottlenecks, more frontier applications like AI scientists will require demonstrated success and broader uptake before becoming standard practice. Furthermore, the most significant bottlenecks in areas like medicine are not in discovery but in the lengthy and costly processes of experimental validation, clinical trials, and regulatory approval. AI can optimize the planning and execution of these later stages by ensuring experiments are based on the most comprehensive understanding of existing knowledge, but it cannot shortcut the fundamental need for empirical testing in humans.
The promise of AI curing all diseases within a decade, as suggested by some AI leaders, is viewed with skepticism. The primary limitation is not the lack of AI capability but the time required for human trials and the sheer complexity of many diseases, some of which may not be solvable with current knowledge. While AI can accelerate the identification of potential solutions and optimize experimental design, it cannot bypass the need for rigorous, time-consuming validation. The true impact will be a significant leap forward over the next 30 years, driven by AI's ability to unlock insights from existing data and to plan more effective experiments, rather than through a sudden, complete solution to all diseases.
The current landscape of AI in science is bifurcated: one branch focuses on modeling the natural world (e.g., protein folding, generative design for antibodies, organism design), while another, exemplified by Rodriguez's work, models the process of doing science itself. The former has seen significant breakthroughs, particularly in generative models capable of creating novel proteins and antibodies. The latter, AI agents designed to conduct research, is rapidly emerging as a key area for future development. Despite the impressive capabilities of tools like Cosmos, the pace of adoption for these more advanced AI scientists will be slower, contingent on scientists' inherent caution and the need for clear, measurable benchmarks to assess progress. Ultimately, while AI will undoubtedly revolutionize scientific workflows, the timeline for its most profound impacts, such as curing major diseases, will be dictated by factors beyond computational power, including experimental realities and the pace of human validation.
Action Items
- Audit AI research process: Identify 3-5 areas where AI agents can automate hypothesis generation and data analysis to accelerate scientific discovery.
- Implement structured world model: Develop or integrate a system for AI agents to maintain a coherent state of knowledge for complex, multi-step scientific tasks.
- Track AI-generated scientific hypotheses: Measure the rate of novel contributions from AI agents compared to human researchers over a 6-month period.
- Evaluate AI-driven experimental design: Assess the impact of AI-generated hypotheses on the efficiency and success rate of subsequent validation experiments.
- Develop AI-specific scientific benchmarks: Create measurable metrics for AI performance in scientific discovery, analogous to math olympiads, to guide progress.
Key Quotes
"the concept here is there's so much more science that we can do than we have scientists right and so how do we scale up science and the thing that is um that happened with cosmos that is pretty cool is cosmos is like the first thing that i think that we've made that actually really feels like an ai scientist when you're working with it right"
Sam Rodriguez explains that the core idea behind their AI scientist, Cosmos, is to address the limitation of human capacity in scientific research. Rodriguez highlights that Cosmos is the first AI they have developed that genuinely emulates the experience of working with a human scientist, suggesting a significant step forward in AI's role in scientific discovery.
"the six month number specifically um the way that we measured this was we had a bunch of academic um you know scientists who had done a bunch of science previously that they had not published yet and we basically gave the same research directive and the same um uh data set to the ai to cosmos and we asked it you know to go away and um just make new discoveries and it would come back and it had found the same things that the researchers had found overnight"
Rodriguez details the methodology for quantifying Cosmos's efficiency, comparing its output to human researchers' timelines. This comparison, where Cosmos achieved results overnight that took human scientists months, demonstrates the AI's potential to dramatically accelerate the pace of scientific discovery.
"one of the main limitations with ai um systems today um is that they're just limited in the length of the task and the sophistication of the task that they can carry out before they kind of go off the rails they like you know forget what they're doing they no longer are on task and what we figured out was a way to have them contributing to this world model that gets built up over time that basically describes like the full state of knowledge about the task that they're working on"
Rodriguez identifies a key limitation in current AI systems: their inability to maintain focus and sophistication over extended or complex tasks. He explains that their breakthrough with Cosmos involves a "structured world model" that accumulates knowledge, enabling the AI to manage intricate, multi-agent processes coherently towards a defined goal.
"and so we asked cosmos we gave it a bunch of raw data um about a huge number of different genetic factors so like what the variants are what proteins bind near the variants right like all these kinds of things and just asked it for type two diabetes to go and um you know identify a mechanism associated with one of these variants um and it came back and it identified this was a variant that was not in a gene and cosmos identified that this is actually somewhere where a different protein binds it was able to identify what protein binds and what gene is being expressed and connected that to the actual mechanism of that gene ssr1 -- which is involved in the pancreas in secreting insulin right"
Rodriguez provides a specific example of Cosmos making a novel scientific discovery related to type two diabetes. The AI identified a mechanism associated with a genetic variant that was not previously understood, pinpointing a specific protein binding site and its connection to insulin secretion, demonstrating its capacity for generating new scientific insights.
"the other side of the ai for science world is building models that can for example predict the structure of proteins that can generate a new antibody that can create a new organism from scratch which are all things that have kind of like happened in 2025 where there's just a huge amount of momentum"
Rodriguez categorizes AI's role in science into two main areas: modeling the natural world and modeling the scientific process itself. He highlights the significant momentum in modeling the natural world, citing advancements in predicting protein structures, generating antibodies, and creating new organisms from scratch as key developments.
"i think that honestly like this year is the year of agents this was the year when people discovered agents and so i i do like you know in good faith have to put myself have to put us on that list and also with google co scientists i mean we're not the only people who are working on this and you know google has been doing a great job there are a bunch of other people so ai agents for science definitely and then like generative design is just having a huge moment right"
Rodriguez identifies "agents" as the defining advancement in AI for science in the current year, noting that this was the year the concept gained widespread recognition. He includes his own work and Google's Co-Scientists as examples, alongside the significant progress in generative design, as key areas of AI-driven scientific breakthroughs.
Resources
External Resources
Books
- "The Genesis Mission" - Mentioned in relation to a White House initiative for AI-accelerated innovation.
Articles & Papers
- "Cosmos" paper - Mentioned as the publication detailing the AI scientist tool and its findings.
People
- Sam Rodriguez - Co-founder and CEO of Future House and Edison Scientific, guest on the podcast.
- Dario Amodei - CEO of Anthropic, mentioned as a leader of AI companies discussing AI's role in scientific progress.
- Sam Altman - CEO of OpenAI, mentioned as a leader of AI companies discussing AI's role in scientific progress.
- Demis Hassabis - CEO of Google DeepMind, mentioned as a leader of AI companies discussing AI's role in scientific progress.
- Cynthia Erivo - Singer, mentioned as an example of incredible vocal talent.
- Kevin Roose - Co-host of Hard Fork, reporter at The New York Times.
- Casey Newton - Co-host of Hard Fork, from Platformer.
- Patrick Collison - Mentioned in relation to the Arc Institute's work on a virtual cell.
- Patrick Shu - Mentioned in relation to the Arc Institute's work on generating organisms de novo.
- Brian He - Mentioned in relation to the Arc Institute's work on generating organisms de novo.
Organizations & Institutions
- Future House - Nonprofit organization co-founded by Sam Rodriguez.
- Edison Scientific - For-profit arm spun out of Future House, developing an AI scientist.
- The New York Times - Publication where Kevin Roose works.
- Platformer - Publication associated with Casey Newton.
- OpenAI - AI research company whose language models are used by Cosmos.
- Google - Company whose language models are used by Cosmos.
- Anthropic - AI company whose language models are used by Cosmos.
- White House - Mentioned in relation to the Genesis Mission announcement.
- MIT - University where Sam Rodriguez earned his PhD in physics.
- The Arc Institute - Organization mentioned for its work on virtual cells and generating organisms.
- Newlimit - Organization mentioned for its work on virtual cells.
- Chan Zuckerberg Initiative - Organization mentioned for its work on virtual cells.
- Google Co-Scientists - Mentioned as an example of AI agents for science.
- Chai - Company mentioned for its work on de novo antibody design.
- Nobla - Company mentioned for its work on de novo antibody design.
Tools & Software
- Cosmos - AI scientist tool developed by Edison Scientific.
- Claude - AI model mentioned for comparison of code generation capabilities.
- GPT-7 - Hypothetical future AI model.
- Gemini - AI model mentioned for its coding capabilities.
- Rovo - AI tool by Atlassian for streamlining workflows.
Websites & Online Resources
- Betterment.com - Website for investment and savings services.
- LinkedIn.com/hardfork - Website for posting jobs on LinkedIn.
- Fidelity.com/tradermore - Website for Fidelity's trading platform.
- Scad.edu - Website for Savannah College of Art and Design.
- Fidelity.com/baskets - Website for Fidelity's basket portfolios.
- Columbia.com - Website for Columbia Sportswear.
- Rovo.com - Website for Rovo by Atlassian.
- Fidelity.com/commissions - Website for Fidelity's commission information.
Other Resources
- AI Scientist - Concept of an AI agent capable of performing scientific research tasks.
- Structured World Model - Key insight in Cosmos for orchestrating AI agents.
- Genesis Mission - Federal effort to accelerate AI-driven innovation and discovery.
- Protein Folding - Area of AI modeling the natural world.
- Generative Models - AI models that can produce examples with desired characteristics.
- De Novo Antibody Design - AI capability to generate antibodies from scratch.
- De Novo Organism Design - AI capability to design new organisms from scratch.
- AI Agents - AI systems designed to perform tasks autonomously.
- International Math Olympiad (IMO) - Competition used as a benchmark for AI progress.
- Vibe Proving - AI systems writing mathematical proofs.
- Robotics for AI Lab Automation - Use of robotics to automate scientific labs.
- Alpha Fold 3 - Protein structure prediction model.
- Virtual Cell - Concept of simulating a cell in a computer.
- Quantum Computing - Technology mentioned in the overhyped/underhyped segment.
- Brain-Computer Interfaces (BCIs) - Technology mentioned in the overhyped/underhyped segment.