The AI That Got Smarter without Practice
A new approach lets AI agents sharpen what they know by imagining experience instead of having it — and filtering that imagined experience more patiently than any real trial could.
Ask anyone how they got good at something and the story has the same shape. They did the thing, over and over. The cook who ruined the first dozen omelets. The driver who stalled at every junction until, one morning, she didn't. We rarely poke at the assumption underneath, because it feels too obvious to question: you improve by doing. By failing, adjusting, trying again. Thinking about it isn't enough. Repetition is.
Some artificial intelligence systems are now built to do things rather than just answer questions about them — to take multi-step jobs off your hands, like booking a trip, digging through research, or writing working code. To do that well, such a system keeps something like a memory: a working bank of strategies and shortcuts rather than a store of facts, little learned moves it can reach for when a new task rhymes with an old one. The interesting question is how that bank improves over time. The leading answer looks a lot like old-fashioned practice. You send the system out on real tasks, thousands of them, watch what works, and keep the lessons that paid off. One strong method runs close to thirteen thousand real trials. It learns the way a new hire learns when you turn them loose on live customers: every lesson is genuine, and every lesson is expensive.
A method called Marginal Advantage Accumulation skips the trials. Instead of making a change and asking "did that work?", it asks "what would happen if I did?" and hands the question to a second AI to judge. But the design turns on one point. A single verdict is noisy. One good outcome might be skill, or it might be luck, and from inside that one result you cannot tell the two apart. The real-trial methods commit anyway, on a single round of evidence. This one refuses to. Before a shortcut earns a permanent place in memory, it has to prove itself not once but across many different settings — many different kitchens, you might say. A move that's merely lucky won't keep winning. A move that's genuinely good will. Wait for the signal to repeat, round after round, and the noise washes out.
The results are the part that should stop you. The system that never practiced, that ran zero real trials, came out on top in fourteen of sixteen tests, against every method it was measured against, the thirteen-thousand-trial one included. It used roughly three-quarters less computation to learn, and finished in about a fifth of the time. Faster and cheaper you might expect. More accurate you would not.
But the whole structure rests on one hidden bet: that an AI can reliably judge the worth of its own would-be improvements. The authors say so plainly. Machine judges have known weaknesses. They tend to favor longer answers, or whichever option they happen to see first. Lean on a flawed judge and the signal driving all that learning rots. You haven't escaped the cost of testing in the real world so much as swapped it for a cheaper stand-in, and the stand-in only pays off while the judge stays reliable. Whether a machine can be trusted to grade itself is one of the field's genuinely open questions. The method's brilliance and its fragility live in exactly the same place.
The most pressing unanswered question is how reliably an LLM judge can evaluate the true benefit of a memory change across diverse tasks compared to direct feedback from the environment.