Autoresearch & The Automation of Science
Karpathy’s Autoresearch sparked discussion around automated scientific exploration. The core mechanism is simple: an agent iteratively modifies code or model parameters, evaluates results, and loops toward improvements. Many saw it less as a breakthrough than as large-scale automated experimentation. Its value may not be in discovering fundamentally new ideas but in scaling exploration. Systems that explore these possibilities continuously could accelerate progress.
Data: The Real Bottleneck
A recurring theme was that progress in computational biology is constrained far more by data than by models. Biological datasets are often small, inconsistent, poorly labeled, or affected by batch effects, sometimes clustering by lab or protocol rather than biology. After strict filtering, usable data becomes scarce. Unlike language modeling, biology lacks massive standardized corpora. As a result, the real advantage lies less in model architecture than in strong data pipelines, experimental systems, and tight feedback loops between models and wet-lab validation.
Modeling Approaches
Two approaches are emerging in computational biology. One bets on increasingly capable LLM-style models trained on biological data to generate discoveries. The other argues biology requires fundamentally different models: structured world models that simulate biological systems and interact with experimental feedback loops. The debate mirrors broader trends in AI. Language-style models capture patterns but may struggle with causal mechanisms, while world models aim to represent systems dynamically through perturbations and feedback. Most expect the eventual solution to combine both approaches.
The TechBio Business Model & Its Failures
The discussion also examined why many AI-driven biotech companies have struggled. Early “tech bio” startups raised massive rounds to build generalized models of biology without clear therapeutic paths. Large amounts of capital went into data generation or broad platforms with few focused applications, leading to limited deliverables and investor fatigue. More successful strategies appear far more incremental: focusing on narrow therapeutic areas, clear constraints, and specific clinical problems.