The 'puzzle that gets harder the closer you look' framing is exactly the part most AI optimism skips. If ~98% of the genome is context-dependent regulation rather than discrete genes, that's not a parsing problem a bigger model solves — it's combinatorial. Makes me wonder if the real bottleneck is model capability or just that we don't have clean enough measurement of cell-state transitions to train on in the first place.
The 'puzzle that gets harder the closer you look' framing is exactly the part most AI optimism skips. If ~98% of the genome is context-dependent regulation rather than discrete genes, that's not a parsing problem a bigger model solves — it's combinatorial. Makes me wonder if the real bottleneck is model capability or just that we don't have clean enough measurement of cell-state transitions to train on in the first place.