The closest I came to the future in 2021 was a waiting list I never joined. I was new to the field then, and a colleague told me about a model you could reach through an API that would, he said, answer almost any question you put to it. There was a queue for access; he may have signed up, but I told myself it would take a while and never got around to it. I did skim the announcement post, and I remember the thought that crossed my mind before I filed it away: once the masses find out about this, it is going to be huge. Then I lost interest and went looking at other things. That model was GPT-3, and what stays with me now is not the model but the shrug: I saw a corner of what was coming, labelled it correctly, and still felt nothing move. It would be nearly two years before I understood that the ground had shifted anyway, and that the thing which shifted first was not the machine but what a small number of people had let themselves believe was possible.

In 1988 the roboticist Hans Moravec noticed something embarrassing about artificial intelligence: it is “comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility.”¹ Nearly four decades later that paradox still defines robotics, but the way the field attacks it has been completely rebuilt, twice. This post traces that rebuild: how hand-engineered control gave way to reinforcement learning, why reinforcement learning then lost the manipulation battle to a much older idea (copying people), and how the last four years fused everything into robot foundation models that look strikingly like their language-model cousins. The goal is a self-contained tour: enough math to see why each transition happened, real numbers from the papers that drove it, and a running experiment you can reproduce.

Table of Contents

Table of Contents

There is a version of the AI pitch that every engineering leader has now heard. A controlled study hands developers an AI coding assistant, they finish the task 55.8% faster, a vendor slide rounds it to “twice as fast”, and the budget conversation ends with a quiet expectation: the same roadmap in half the time, starting next quarter. This is a short essay about why that expectation fails, and what the people who run teams should steer by instead.

Every morning, your coding agent shows up brilliant and amnesiac. It can refactor a module, chase a bug through five files, and quote the standard library from memory, and tomorrow it will remember none of it: not the bug, not the fix, not the thing you told it three times about how deploys work here. Andrej Karpathy put the complaint bluntly in October 2025: “They don’t have continual learning. You can’t just tell them something and they’ll remember it.” This post is a field guide to the two connected problems hiding in that sentence, remembering (keeping state across sessions) and learning (getting better with experience), and to the machinery being built for both: retrieval scores and memory pipelines, RL-trained librarians, a catastrophic-forgetting experiment you can run in numpy, the elastic weight consolidation derivation in full, and the new architectures that let weights learn at inference time.