Scientists Found a Way to Stop AI Forgetting What It Already Learned

Here's a problem most people don't know AI has: when a system learns a new task, it often wipes out what it learned before. Researchers call it 'plasticity loss' — the AI becomes stiff and can't adapt anymore. It's like an employee who gets so set in one routine that learning a new software tool makes them forget the old one. For years, no standard training method has reliably solved this, which is why your AI tools don't quietly improve as you use them.

A team of researchers tried something different. Instead of nudging one AI brain with small corrections — the usual method — they used evolution-style training: create a population of AI 'brains,' let the best performers survive, and mutate them slightly to make the next generation. Think of it as breeding rather than tutoring. They tested this across many different tasks and environments, from tiny brains with a few hundred settings to large ones with a million.

The winner was a method called evolution strategies. It was the most consistent at balancing two things that usually fight each other: learning new skills quickly while keeping old ones intact. A related method, genetic algorithms, learned new things even faster but forgot more. The team also found a clue as to why: the winning method produced AI brains that were 'sturdier' — small random changes to them didn't break their performance, and those sturdy zones overlapped between old and new tasks.

One notable finding: the classic warning signs of AI forgetting didn't show up in the evolution-trained systems at all. That suggests the fix isn't just a patch — it may be a genuinely different way of building AI that keeps improving over time. Don't expect your phone to update tomorrow; this is lab work, and the tests were in simulated environments. But it points toward AI assistants, robots, and recommendation systems that get better at new things without quietly losing the old ones.