By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
AI’s Next Trillion-Dollar Market Isn’t Software | Max Jaderberg, President of Isomorphic Labs
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
Great conversations run long. Their best ideas don't have to.
uListen is free on iPhone, iPad, Android, Chrome, and Claude.
See howCHAPTERS
Jump to publisher timestampsChapter summaries
From core AI at DeepMind to the conviction that AI can generalize in biology
Max Jaderberg traces his background in core deep learning and high-profile challenge problems, then explains how internal AlphaFold progress changed what he believed AI could do. He describes AlphaFold as the proof point that models can generalize in biology and chemistry, motivating the leap to founding Isomorphic Labs around an audacious health mission.
- •Max describes his history in computer vision, generative models, language models, and reinforcement learning work at DeepMind
- •He explains how seeing AlphaFold results internally built conviction that AI can generalize in biology/chemistry
- •He frames AlphaFold as a proof of concept rather than a direct solution to making medicines
- •He connects that proof point to launching Isomorphic Labs focused on using AI for human health
Why making a medicine is hard: many coupled problems, long timelines, high failure rates
Max explains that AlphaFold alone doesn’t create medicines, because drug development requires solving many difficult problems at once. He outlines the chain of scientific and practical constraints—from target selection to safety and delivery—that make the process long and failure-prone.
- •Max says medicine creation requires many tricky problems to come together, not one breakthrough
- •He states it takes about 10 years from conception to a medicine for patients
- •He notes a high failure rate in development, citing that about 90% of medicines fail
- •He lists challenges including choosing the right protein target, designing effective molecules, reaching the right body location, and avoiding toxicity
Building “half a dozen plus” AlphaFold-level breakthroughs into a holistic drug design engine
Max argues that success in drug discovery needs multiple AlphaFold-like leaps that combine into a single platform. He describes identifying the “superpower” capabilities that would radically improve chemists’ and biologists’ effectiveness, then pursuing them systematically.
- •Max says the AI solution is a collection of breakthroughs that must work together
- •He describes identifying specific “superpowers” that would transform drug design
- •He mentions areas such as understanding cells and disease biology, molecule design, and human toxicity
- •He says the team has spent years systematically pursuing these breakthroughs toward a holistic engine
Two fronts where AI helps: understanding disease biology and designing interventions
Max splits Isomorphic’s work into understanding and design. He describes foundational biology models that help interpret proteins, DNA, RNA, and pathways in healthy and diseased states, and then models that help design medicines like small molecules, antibodies, and peptides.
- •Max frames AI’s role as both understanding biological systems and designing interventions
- •He emphasizes the complexity of biology and limited current human understanding
- •He describes models that map interactions among proteins, DNA, RNA, and cellular pathways
- •He outlines design targets including small-molecule pills and injectable modalities like antibodies/peptides
Commercializing without waiting a decade: partnerships, internal programs, and two business modes
Marina presses on how long-term research becomes a business, and Max explains their approach: build the platform while working on real drug design from day one. He describes partnering with top pharma to avoid “marking your own homework,” alongside internally funded programs where the company owns the IP.
- •Max says they avoided building in isolation for a decade before commercialization
- •He explains early partnerships brought “the industry’s hardest problems” and external validation
- •He describes two approaches: internally initiated programs (fully funded, owned IP) and partner programs
- •He frames real-world commercial constraints as essential for shaping the platform and models
When AI is “ready for real work”: predictive value over perfection, and closing the benchmark–reality gap
Max contrasts deployment requirements in other industries with drug discovery, where even imperfect prediction can create meaningful value. He explains how internal benchmarks and evaluations guide model development, but must translate into results on live drug programs—or else the benchmarks need redesign.
- •Max contrasts high-precision requirements in domains like self-driving with drug discovery’s empirical baseline
- •He says value comes from achieving a meaningful predictive delta over the industry, not perfection
- •He describes using benchmarks/evals as proxies and translating gains onto live programs aggressively
- •He warns that benchmark progress can fail to translate, requiring redesigned evals to match reality
Managing extreme ambition with a mission tree, milestones, and quantitative “leaderboards”
Max explains why a huge mission matters and how he operationalizes it so teams can execute day to day. He describes breaking the mission into a “mission tree,” then setting measurable milestones—sometimes intentionally “gamified”—to create signals of progress during long timelines.
- •Max argues big ambition changes how you build a company and its goals
- •He describes a “mission tree” from “solve all disease” down to individual-level missions
- •He emphasizes practical milestones as litmus tests for whether progress is real
- •He gives an example of setting a quantitative benchmark target far above the world’s baseline and tracking incremental gains via leaderboards
Where the next AlphaFold-like moments may appear: simulating parts of the physical world
Max discusses domains beyond biology where AI can provide step-change value by simulating outcomes that currently require real-world experiments. He distinguishes reasoning models that generate hypotheses from models that can predict experimental outcomes, arguing many applied fields remain bottlenecked by physical testing.
- •Max highlights the complementarity between reasoning models and world-simulation models
- •He describes a bottleneck: needing to “poke the real world” via physical experiments to see outcomes
- •He suggests opportunities for AlphaFold-like moments in domains like material science and engineering design
- •He frames the breakthrough as replacing some physical experimentation with model-based simulation
How to tell if you have the right data: signal, noise, and new ways to generate unlocking datasets
Max explains that readiness often comes down to data availability and quality. He describes assessing whether existing datasets contain enough signal to support the desired predictive capability, and notes that creating new data-generation methods can unlock previously infeasible problems.
- •Max says evaluating a modeling idea starts with whether the necessary predictive capability has supporting data
- •He warns that sparse, noisy, or low-signal data can limit impact regardless of ambition
- •He reframes data scarcity as an opportunity: invent new ways to collect or generate better data
- •He states his view that algorithms and compute are strong, making data the key constraint in many applied domains
Hiring AI engineers without biology backgrounds: building interdisciplinary “intersection” talent
Max describes why Isomorphic hires top people by discipline rather than expecting pre-existing experts in “AI drug design.” He emphasizes curiosity and learning ability, and argues that newcomers’ naivety can drive first-principles thinking that’s valuable in emerging fields.
- •Max says there is no long-established world expert in making drugs with AI because the field is new
- •He outlines hiring across AI, software engineering, chemistry, biology, and clinical expertise
- •He values candidates willing to learn and work across disciplines, building intersectional skill internally
- •He notes a benefit of outsiders: fresh perspectives and first-principles analysis
The interview questions that reveal how candidates learn and recover from failure
Marina asks how he tests for curiosity and learning speed, and Max explains the prompts he uses. He looks for stories about entering unfamiliar domains and technical failures, focusing on emotional response, approach, and what the person did to learn their way through.
- •Max asks candidates to describe learning a new field and what it felt like
- •He probes for examples of working in uncomfortable, unfamiliar situations
- •He asks about technical work that failed and how candidates figured out solutions
- •He uses these stories to understand how people think and learn under uncertainty
From millions of hypotheses to a handful of molecules: generative design plus “virtual lab” filtering
Max explains how AI-generated hypotheses become actionable by pairing design agents with predictive “world model” simulations. The goal is to reduce expensive lab work by narrowing down vast candidate sets to a small number of high-confidence molecules for real-world synthesis and testing.
- •Max describes AI generating huge numbers of candidate molecule designs/hypotheses
- •He explains using complementary predictive models as virtual labs to forecast experimental responses
- •He frames the pipeline as design agents proposing candidates and simulation models refining them
- •He says this can reduce experimental screening from hundreds of thousands of designs to around 20 or 50 before lab testing
How researchers’ roles change when agents propose designs and run experiments
Max argues AI shifts both AI research and chemistry from hand-authoring individual ideas to defining search spaces and objectives. He describes a move toward “auto research,” where people set goals and constraints while agents generate and test variations autonomously.
- •Max says individual AI researchers increasingly define a space of model changes and goals rather than single hypotheses
- •He describes agents proposing ideas and running experiments within that defined space
- •He draws an analogy in chemistry: drug designers define objectives and design space, and agents generate molecule iterations
- •He frames the shift as humans focusing on objective-setting while agents handle large-scale exploration
Max’s personal AI stack, long-running drug design agents, and how to evaluate agent value
Max shares how he uses multiple models and agents for productivity, especially for data analysis and visualization. He then explains that drug design agents can run for days, with success defined by multi-objective goal spaces, and argues agent development must be evaluation-driven to avoid self-deception about usefulness.
- •Max says he uses many models internally, including Gemini and open-source models, with agents connected across his workflow
- •He highlights data analysis/visualization as a “mind-blowing” personal use case when company data is plugged into an agent stack
- •He explains agents can run for multiple days, with success defined by multi-property objectives and tradeoffs rather than a single metric
- •He emphasizes evals/leaderboards to determine whether agents save time or create scientific/economic value, especially in domains that aren’t fully verifiable on a computer
What stays human: people leadership, attention as a bottleneck, and why clinical trials still fail
Max identifies human management and conversation as a core part of his job he doesn’t see as automatable, and argues human attention and physical constraints remain limiting resources. He also explains why regulation and safety evidence are necessary, and why many drugs still fail in clinical trials due to toxicity and limited predictiveness of preclinical models.
- •Max says he spends much of his time managing and talking with people, and doesn’t see that as something to automate now
- •He describes human attention, time, and physical space as enduring bottlenecks even as AI improves
- •He explains the regulatory process is evidence-heavy for patient safety before entering clinical trials
- •He describes why failure remains high in the clinic, including unintended toxicity and limits of animal models in predicting human toxicity
When AI-designed drugs might reach patients, plus guidance for working in AI for science
Max is cautious about predicting when AI-designed drugs will appear in pharmacies, but describes ongoing work to push candidates into the clinic and mentions areas of focus like oncology, immunology, and inflammation. He closes with advice for aspiring AI-for-science practitioners: learn the underlying science, focus on choosing good questions, and develop “research taste” through experience and cross-domain thinking.
- •Max says it’s hard to predict pharmacy timelines, noting the process takes time, while pointing to ML-touched programs entering clinics
- •He describes Isomorphic working to advance its own pipeline of fully AI-designed drugs toward clinical trials
- •He mentions excitement about programs in oncology as well as immunology and inflammation, while saying it’s too early to know which patients benefit first
- •He advises people to learn science fundamentals, focus compute and effort on the right questions, and build research taste through experience and generalist crossover ideas
Publisher timestamps
- 0:00Intro
- 3:02Why Developing a Drug Can Take 10 YearsRead this chapter
- 5:38How AI Helps Scientists Design New MedicinesRead this chapter
- 10:22How to Tell When AI Is Ready for Real Work
- 13:12How to Turn a Big Mission Into Measurable Goals
- 16:58Where the Next AlphaFold Moment Could HappenRead this chapter
- 19:00How to Tell if You Have the Right Data for AI
- 20:55Why Max Hires AI Engineers Without a Biology BackgroundRead this chapter
- 23:09The Interview Question Max Uses to See How You LearnRead this chapter
- 24:45How AI Narrows the Search to 20 or 50 MoleculesRead this chapter
- 26:58What Researchers Do When AI Comes Up With the IdeasRead this chapter
- 28:29Max’s Personal AI StackRead this chapter
- 29:30Inside a Drug Design Agent That Runs for DaysRead this chapter
- 30:46How to Tell if an AI Agent Is Worth UsingRead this chapter
- 32:22The Part of His Job Max Keeps Human
- 33:17Why Promising Drugs Still Fail in Clinical TrialsRead this chapter
- 35:58When AI-Designed Drugs Could Reach PharmaciesRead this chapter
- 40:14What to Learn if You Want to Work in AI for ScienceRead this chapter