AlphaFold won a benchmark, not the problem. Scaling needs problem definition first.

The headline claim. Pushmeet Kohli is blunt about AlphaFold. It matched structures that researchers had already deposited in the PDB, and that turned out to be useful. It did not solve protein folding in the deeper sense: the true ground state, the distribution of shapes, and dynamics remain open. Proteins are disordered and context-dependent. The lesson for any engineering leader is that a benchmark win is not a solved problem.

The scaling caveat. Sal Candido challenges the naive version of scaling laws. They are not automatic; finding one is the actual research. A model can only pull information that exists in the data. His concrete example: training protein language models on messy metagenomic sequences, many of which are not even full proteins, improves the ability to design real proteins. But that can tempt teams into generating more of the wrong data instead of asking what data the problem needs.

The bitter lesson. Kohli reframes Rich Sutton's bitter lesson beyond data versus modeling. The original trap is tribalism: researchers who call themselves modelers ignore data generation, and data people ignore modeling. AlphaFold invested in modeling because the PDB took decades of global investment and could not be augmented cheaply. For the virtual cell, the opposite was true: the data did not exist yet. Define the problem first, respect constraints, and fail fast.

The craft tradeoff. AlphaFold 2 was not pure scale. It encoded physics and biochemistry into the architecture, which made it data-efficient. Kohli argues that curated data and coverage are their own art. Candido adds the counterpoint: craft still matters in scaling, through algorithmic work, bespoke architectures, and faster inference, and a slightly wrong inductive bias can hold a model back once data grows. The strategy depends on where you are on that curve.

The uncertainty lesson. Reliability is not accuracy. AlphaFold 2 scored around 90 GDT on its benchmark, but its confidence measure, pLDDT, was poorly calibrated. A model that is correct on average but overconfident on failures is dangerous because humans cannot tell which answers to trust. He wants behavioral characterization before use. Interpretability, he argues, is in the eye of the beholder: a human may not read AlphaFold's reasoning, but a larger model might.

The practical call. Both researchers are cautious about AI in the clinic. AI already touches every part of drug discovery, but a 10x acceleration depends on harder biology and better models of disease, not just bigger models. Candido says basic research and understanding are still the foundation. For engineering organizations, the transferable signal is to name the real problem, check whether the benchmark measures it, and treat confidence calibration as a release criterion.