Software, automations, and whatever experiment I'm running this week.
Grokking has mostly been shown on clean, synthetic tasks. I trained a tiny transformer on real planetary orbits, diagnosed two dataset confounds that produced a fake negative result, got a genuine 24,000-step delayed generalization once they were fixed, then ran 100 seeds to check how often that actually happens.
A proof-of-concept multi-agent pipeline where 22 specialist classifiers pre-process every query before it reaches an LLM, and what happened when I actually built and benchmarked it.
I bet a biology-inspired importance signal could beat the standard method at stopping a neural network from forgetting. On the easy benchmark it did, then it failed on the hard one. Here is the honest arc, and the mechanism that explains both.