(This blog post is derived from this video response on my YouTube channel.)

Dwarkesh Patel has been making the case for continual learning for a while, both with guests on his podcast and, most recently, in a blog post and an accompanying video.

I agree with his basic point. A frozen model that cannot learn from experience or adapt to its surroundings is limited. If we want AI systems that can carry out long-running work and do open-ended problem solving in the real world, some form of continual learning is necessary.

What I think he misses is classifying tasks and matching kinds of tasks to the kinds of continual learning they might need. He also seems to imply that the only useful or “real” way to do continual learning is through some sort of fluid or continuous updating of underlying model weights.

In the video, Dwarkesh says:

“I don’t think you can have AIs that perform whole jobs as competently as humans if they’re forced to just write Markdown files…”

He then gives an illustrative example: imagine an endless line of students trying to learn the saxophone. Each student has never played before. One enters the room, fails, writes down what went wrong, and hands those notes to the next student. No matter how detailed the notes become, none of them will learn to play the saxophone this way.

That is a good argument for continual learning. But it is an argument about a particular kind of task.

Skills, Rules, and Knowledge

In an earlier video, I looked at Jens Rasmussen’s classic Skills, Rules, and Knowledge framework. Rasmussen divides human performance into three broad categories.

Skills are fluid, integrated sensory-motor performances. Playing a saxophone is a skill. So is riding a bicycle or driving a car. Once you are proficient, you do not consciously execute a written procedure. Your perception and action are tightly coupled, and your whole body participates.

Rule-based work begins with a familiar situation. You recognize a pattern, then carry out a reasonably well-understood procedure: if these conditions hold, do these things. A lot of what we call knowledge work is actually rule-based work. It has known inputs, bounded variation, and a repeatable process that produces an output.

Knowledge-based work is open-ended problem solving. The situation is novel. There is no stored procedure that directly applies, so you have to reason abstractly. Proving a new theorem, debugging a failure you have never seen before, or diagnosing a genuinely unusual medical case all fit here.

Dwarkesh’s saxophone example is squarely in the first category. Of course prose is a poor substitute for the sensory feedback and motor adaptation involved in learning an instrument. No amount of Markdown can tell you the sound of a badly played note.

But it does not follow that written knowledge is an inadequate learning mechanism for the other categories.

Civilization Is a Giant Counterexample

There is a massive counterexample to the idea that important learning cannot happen by writing things down and passing them to the next learner: all of civilization.

We know physics, chemistry, mathematics, politics, and philosophy because generations of people externalized what they learned. Many of the people whose ideas we rely on have been dead for centuries. We cannot interact with them, watch them work, or ask them for another training run. We read what they wrote, combine it with what others wrote, test it against the world, and write down something new for the next generation.

In modern agent terminology, civilization is an enormous collection of Markdown files passed from one context window to the next (people’s brains).

This distinction matters for AI systems. Learning does not have to mean that every useful experience is dissolved into the base model’s weights. It can also mean improving explicit procedures, accumulating durable external knowledge, and retrieving the right material when a future task needs it.

For skill-like tasks, weight updates or some equivalent form of low-level adaptation may be essential. For rule-based and knowledge-based tasks, external memory, written procedures, tools, and structured knowledge may do much more of the work than the saxophone analogy suggests.

Who Owns the Learning?

There is also a practical enterprise objection to treating continuous weight updates as the only serious form of learning.

In the Alex Karp interview that some people described as a “crash-out”, he made a reasonable point: companies care about controlling their compute, models, data stack, and “alpha”—the proprietary operating knowledge that gives them an advantage.

Suppose the large model labs solve continual learning. Every interaction with a company’s systems, workflows, decisions, corrections, and exceptions can now improve the model. Who owns that improvement? Does it remain exclusive to the company that produced it, or does it become part of a provider’s weights?

This is not the same as running your databases on AWS. When my code operates on my data in someone else’s data center, the cloud provider is supplying infrastructure. It does not follow that the provider is absorbing my database into its own institutional knowledge.

A continually learning model potentially crosses that boundary. Even if providers contractually promise not to train on enterprise data today, the architecture raises the question directly: where does the learned state live, who controls it, and who else benefits from it?

Many companies will not want their most valuable accumulated judgment hosted in a globally shared model. They will want a scoped learner: an instance that learns only for them, that they control, and perhaps that runs inside their own boundary.

Frozen Weights Can Be a Feature

This is why the separation between frozen weights and learning through skills, rules, and external knowledge can be a feature rather than a bug.

The frozen model provides general capability. Company-specific procedures, tools, corrections, and knowledge remain in a layer the company owns. That layer can change continuously without forcing the organization to surrender its accumulated learning to the model provider or to retrain an enormous base model after every useful experience.

None of this means weight updates are unimportant. They may be indispensable for some kinds of learning, especially tasks that resemble sensorimotor skills. But “continual learning” is not a single mechanism, and the right mechanism depends on both the task and the ownership boundary.