Continual learning isn't the crux

18 Sep, 2026

Continual learning (CL) is increasingly treated as the holy grail of machine learning, the missing puzzle piece that'll usher in a new era. The reason is that models don't "learn on the job", as illustrated by the classic example of the worker who comes in every morning with no experience and only the previous days' notes. I think this diagnosis is correct, but that the lack of CL isn't the primary roadblock.

So what is CL? For a given model $\theta$ and batches of training data $D_1,\dots ,D_n$, we want to come up with an update rule $U$ that's able to approximate conventional joint training using sequential updates.

Note that the new end state doesn't unlock any new capabilities relative to training on the union natively for a fixed dataset.¹ It just allows us to continue training a model with new data without re-training on the old.

When I first saw this definition laid out, I felt pretty underwhelmed. It's just a compute optimization, and one of uncertain efficacy: there's theoretical work suggesting that a continual learner needs memory that grows linearly with the number of tasks it's seen.

So what's the problem?

I think there's some sleight of hand going on here. Learning from real world deployment is bundled together with or otherwise treated as something downstream from CL, when they’re really two orthogonal ideas.

Most of the capabilities people ascribe to CL in systems that have it, such as humans², are better explained by the capability to learn from experience, without the verifier and repeated rollouts that RLVR depends on.

If OpenAI was able to gather useful training signal from the millions of tasks GPT completes every day, it wouldn't matter much whether the data was baked in on a rolling-release schedule vs. served as a fresh checkpoint weekly.