Blog

Disagreement between a model and a weakened copy of itself does not mark the steps that decide the answer. It marks local fragility instead, and the difference matters for any attribution method.

Sparse autoencoders recovered transcription factor structure from a DNA model without labels, and a shuffled-modality ablation exposed cross-modal signal that the headline metric could not see.

Changing the prompt reordered which model looked best, and swapping the judge model shifted scores far more than rewording the judge instruction. The number describes a model, a prompt, and an evaluator together.

Sparse networks collapse with depth if their weights train, and stay stable past 20 layers if they do not. Running the configuration nobody publishes is what isolates optimization as the cause.

Inside one fixed pipeline, weaker models ran over 180 seconds with error rates above 80 percent while a stronger one answered in 45 seconds with none. Good orchestration made failures safe, not rare.

Optimizing past convergence made faces worse, because the solver began fitting the landmark detector's errors instead of facial geometry. The same failure shape appears whenever you optimize an imperfect target.