Stanford professors clash over AI limits: data bottlenecks vs a recent convert's 'it just works'

Stanford computational biologist Anshul Kundaje and his colleague Batzoglou have clashed publicly multiple times over the limits of AI capabilities. Kundaje criticized the trend of extrapolating model capabilities without bounds, arguing that terms like Scaling, ASI, and RSI offer no concrete roadmap for achieving those capabilities and feel more like fortune-telling than science. Batzoglou previously held a similarly skeptical stance but said his view has recently shifted — 'I used to be like you, until recently, when it started looking like it really works.'

Confirmed

• Kundaje believes AI model capabilities have a clear ceiling rooted in the nature of trainable data, citing hard biology problems with scarce data as an example where frontier models have so far underperformed.
• He questions the claim that 'AI will cure all diseases in 5-10 years,' viewing it as an example of how far trend extrapolation can go without a roadmap, easily leading to absurd conclusions.
• Responding to TuXinming's proposed route of training on interaction trajectories rather than plain text, Kundaje said such proposals have been raised repeatedly, but proponents mostly underestimate the diversity and scale of data required, and physical bottlenecks stand in the way.
• Kundaje went further: the data bottleneck can only be lifted once AI-driven detection/experimental techniques improve recursively and truly massive, fully robotic labs with automated supply chains emerge.
• He was lukewarm on recent AI math proof breakthroughs (including Millennium Prize-type results): while a huge milestone for AI capability, most have no direct real-world impact as far as he knows.
• TuXinming's specific path: pretraining essentially compresses 'world → observation → written records' into the model, while tool use provides a closed loop of 'action → feedback → training on trajectories,' which labs are already practicing through heavy tool calls and agent runs.
• Batzoglou said he has shifted from long-term skepticism to acknowledging that AI has indeed 'started working' recently.

Why it matters

The debate crystallizes the core divide in today's AI community: one side demands verifiable implementation paths for capability leaps, while the other abandons skepticism based on recent real-world performance. Kundaje's 'physical bottleneck of data diversity and scale' offers a concrete criterion for assessing AI's prospects in data-scarce fields like biology, and whether 'fully automated robotic labs' can unlock that bottleneck will be a key variable to track going forward.

2026-09-09 ~ 2026-09-09 · 10 related posts

• [source] Stanford's Anshul Kundaje: models are soberingly bad at data-poor biology, stop extrapolating from scaling — anshulkundaje · 2026-09-09
• "Scaling, ASI, RSI is fortune-telling, not science": Kundaje's critique of capability prognostication — anshulkundaje · 2026-09-09
• Stanford's Batzoglou drops his AI skepticism: "It looks like it just... works" — s_batzoglou · 2026-09-09
• Stanford professor pushes back on claims AI will cure all diseases in 5-10 years — anshulkundaje · 2026-09-09
• [source] A concrete path past the data wall: train agents on interaction trajectories, not just text — TuXinming · 2026-09-09
• Stanford professor and computational biologist clash over claims AI will cure all diseases in 5-10 years — anshulkundaje · 2026-09-09
• Stanford professor unmoved by AI math proofs: no real-world impact yet, and biology sees nothing surprising either — anshulkundaje · 2026-09-09
• Stanford's Kundaje: interaction-based training proposals underestimate data scale and physical barriers — anshulkundaje · 2026-09-09
• Kundaje: the data barrier holds until AI-built assays and fully robotic labs arrive — anshulkundaje · 2026-09-09

1 near-duplicate retellings: anshulkundaje