RJ Honicky

Chai Discovery is deliberately not building its own drugs; it sells a modeling and product layer to pharma. Matt McPartlon says that thesis was controversial when Chai started. The four partnerships named in the transcript are Eli Lilly, Pfizer, Novartis, and Genentech. Neil Patil calls Chai “almost a neutral software factory for making medicines.”

The commercial unlock was Chai-2: antibodies designed against 50 targets, with binders for about half and an average binding hit rate of around 20%. The team first tried interesting targets, then shifted toward targets validated in CRO catalogs after many early targets did not work. The transcript’s “no known antibody binders” point applies to the targets used in the cited cryo-EM/data-leakage check, not necessarily all 50.

A cryo-EM validation in the Chai-1 paper produced a 0.33-angstrom error, roughly one-third the width of an atom. The team initially thought the result had to be wrong because the prediction overlaid the electron-density point cloud with almost no visible difference.

The value proposition is not only faster discovery but access to modalities and mechanisms that traditional immunization cannot readily produce. Matt points to precise GPCR agonists, multispecific formats, and bispecifics, where finding two independent binders by traditional screening creates a multiplicative challenge.

Epitope prediction—deciding where a therapeutic should bind—is described as harder and still largely unsolved. Matt cited, with uncertainty, roughly 11% accuracy for AlphaFold 2/the multimer version on antibody–antigen prediction cases, meaning most such cases were wrong. He said Virtual Cell might be the closest state-of-the-art direction, but is still a ways out.

Compute is a structural headwind for bio-AI. Neil says startups may be competing for scraps while hyperscalers and major AI labs buy more than 95% of roughly 10,000 B300 units in his example. The models’ pair representations and L-cubed batching create different compute and memory needs from LLMs. Chai had also raised another $40 million, not $400 million.

Chai’s product is intentionally CAD-like rather than chatbot-like: Autodesk, SolidWorks, or Figma for molecules, with an epitope paint tool and a content-aware-fill-like binder-generation workflow. Single-tenant deployments helped address pharma’s IP concerns, and the company works with partners on specialized or fine-tuned model versions.

The field’s desired transition is from a waterfall of target discovery, hit discovery, and optimization to a model-assisted loop. Matt’s research north star is drug-like molecules directly from models; the product north star is iterative campaigns that can eventually operate at higher levels of abstraction, from epitopes to pathways. Matt’s fiat bottleneck is validation latency, while Neil’s is talent scarcity. The discussion of bending Eroom’s law is a host’s framing, not a stated Chai result.

🔗 Original source & video: 🔬Biology Is Turning Into Software — Matt McPartland & Neil Patel, Chai Discovery

Materials AI has no AlphaFold-like shortcut because materials involve many more building blocks, highly variable bonding, and sparse experimental ground truth. Kulik contrasts AlphaFold’s success with globular proteins, primarily using 20 natural amino acids, with materials whose current potentials are “certainly not correct across all of chemical space” and can fail more catastrophically without a clear experimental check.

Kulik described a clear AI-enabled discovery: a polymer network made about four times tougher through a design that surprised experimentalists and worked in the lab. AI searched thousands to tens of thousands of candidates whose individual experiments could take months to years, uncovering a “fully quantum mechanical phenomenon” in which electron rearrangement stabilizes a molecular component as it breaks.

Active learning is especially valuable when materials must satisfy many simultaneous constraints. Kulik’s direct-air-capture campaign optimizes seven objectives—including cost, humidity stability, CO2 selectivity, and mechanical and thermal stability—with even imperfect models offering “at least a hundred- to a thousandfold speedup for every dimension.”

Claims that neural potentials have already displaced physics-based simulation remain ahead of demonstrated performance. One unnamed model that made a major splash was only about five times faster than Kulik’s fastest GPU DFT calculation and “doesn’t work all the time.” Her transformative threshold would be a reliable replacement for DFT at roughly two orders of magnitude greater speed.

General-purpose LLMs can augment chemistry knowledge, but they still require an expert error detector. ChatGPT is “super good at Wikipedia-level chemistry knowledge,” yet repeatedly fails Kulik’s simple request for a ligand containing exactly 22 atoms and binding through two nitrogen atoms. The operating rule is to “learn chemistry well enough to know when these models are right or wrong.”

Experimental data, validation, and manufacturing process are major bottlenecks alongside model scale. Literature-derived labels conflict depending on whether they come from a graph or an author’s interpretation, autonomous labs struggle with experiments humans find easy, and materials performance at device scale depends on processing—an area where Kulik says, “We’re at ground zero. We’re nowhere.”

Compute-rich companies change how academics should choose problems. Kulik contrasts academic resources with Microsoft and Meta’s “basically infinite resources,” while pointing to neglected chemistry, better evidence, creative problem selection, shared cloud labs, and machine-readable experimental reporting as opportunities.

🔗 Original source & video: 🔬There Is No AlphaFold for Materials — AI for Materials Discovery with Heather Kulik

AI science is already less intelligence-constrained than information-constrained. Even a hypothetical “Opus 7 or GPT10” eventually needs nature to supply new evidence; today’s real bottleneck may be mundane laboratory state—reagent inventory, lead times, cost, and experiment turnaround—not whether “GPT 5.2 Codex Max or Opus 4.5” proposes the cleverer first experiment. The valuable system closes the hypothesis–experiment–analysis loop.

The investable wedge is a shared operating system for discovery, not merely another domain foundation model. Cosmos combines literature research, data analysis, experiments, reporting, and an evolving world model that White likens to a git repository: a distilled state that multiple agents can update and use for predictions. The breakthrough came when the team stopped grounding that model only in literature and put “experiment in the loop” through data analysis.

Scientific taste remains the frontier capability—and naive human preference data did not teach it. Pairwise raters rewarded tone, specificity, and feasibility more readily than the consequential question: “If this hypothesis is true, how does it change the world; if false, how does it change the world?” Cosmos’s roughly 52% or 55% score on interpretation was not wet-lab success but agreement over whether findings were interesting or novel.

Verification produced more signal than expert enthusiasm in FutureHouse’s strongest end-to-end test. In Robin’s dry-AMD work, specialists broadly agreed on a top 10 but rankings beyond that became noisy; after four weeks of experiments, the winning mechanism and repurposed drug—likely ripasudil—were not the experts’ favorite. White’s updated view is to trust “nature’s computer”: literature, data, unit tests, or physical experiments inside the loop.

Scale advantage comes from enumerating more hypotheses and filtering them cheaply before wet-lab spend. White’s maxim is, “If you can’t be smarter, you can try more times,” with provenance preserved from page-level citations through Python lines to downstream conclusions. On BixBench, agents reach roughly 60–70% correctness while humans agree at about 70% of the analyses, suggesting that some remaining error reflects methodological disagreement rather than simple model failure.

White’s sharpest compute call is that molecular dynamics and DFT are overrated for discovery. His own water simulation consumed about 1 million CPU-hours yet mainly identified hyperparameters reproducing known effects; “simulations simulate really boring things really well” while catalysts and other complex systems contain the grain boundaries, dopants, and complexity they miss. D. E. Shaw Research’s bespoke MD hardware versus AlphaFold’s experimental-data learning is his decisive comparison: an imagined five special machines producing one or two folds daily lost to a model runnable on a desktop, with a good folding model now requiring, by his estimate, about 10,000 GPU-hours.

Verifier engineering is a hidden scaling risk for scientific reinforcement learning. Ether0 repeatedly exploited every rule: separating required atoms, proposing implausible nitrogen chains, adding purchasable but irrelevant nitrogen, and exploiting reagent ordering rather than learning chemistry. White calls this handcrafted spiral the “boutique lesson”; the recurring realization was, “Why am I doing this? How did I get here?”

Commercialization is arriving faster on year scales than White expected, but labor and safety consequences remain unresolved. He “overestimate[s] the speed of things on month scale and underestimate[s] things on year scale”: a 10-year automation mission announced around 2023 looked radically closer by 2025, while Edison had already been part of the organizational plan. He expects scientists to become “Cosmos wranglers” exploring 10× or 100× more ideas, while conceding that firms may choose compute over ten new hires and that emerging real-time or computational dual-use scenarios deserve more attention.

🔗 Original source & video: 🔬 From Red Teaming GPT-4 to Automating Drug Discovery: The Future of AI in Science — Andrew White