@MichaelElabd

**Michael Elabd NeurIPS** @MichaelElabd [2025-12-20](https://x.com/MichaelElabd/status/2002281632125026730) In 2026, Continual learning will shift from research curiosity to a tractable problem. The hard parts we will need to address: Defining the environment: user interactions aren't clean MDPs. They're partially observable, scattered across sessions, noisy, and sparse. Credit assignment: which interaction caused the improvement? The regression? Rewards are delayed and adversarial. Stability: learning online without catastrophic drift, reward hacking, or silent degradation. Control: rollback, isolation, eval-in-the-loop, safe exploration. Non-negotiable. Infra: routing, replay buffers, off-policy learning, fast evals as core primitives. Just like pretraining scaled knowledge, fine-tuning scaled instruction-followingm continual learning will scale adaptation. --- **Pedro Domingos** @pmddomingos [2025-12-20](https://x.com/pmddomingos/status/2002373549219053872) Continual learning has been tractable for decades, and every one of those problems has been extensively studied. There’s much to do, of course, but approaching the problem as tabula rasa is not smart or good science. --- **Harshal Nandigramwar** @hnanacc [2025-12-20](https://x.com/hnanacc/status/2002371238325526617) naive question: how is continual learning different from just normal training? Let's say we train a model on 10 batches of data and get the base model. If we save this checkpoint, and when we have new data train it as batch 11, 12 and so on. And then just do the post train or some weight space arithmetic, we should have it solved? what's stopping this? if we raise the complexity a bit, we can also have data curriculums, mixtures, etc all through checkpointing and some weight space wrangling.