Amazon's New Tool Lets AI Teach Itself — Maze Scores Jump to 95%

Amazon has shared a detailed playbook for something called reinforcement learning, which is how AI learns by trial and error rather than by memorizing answers. Think of teaching a dog to find a treat in a maze: you can't explain every turn, you just reward the runs that work. Amazon used its cloud service, SageMaker HyperPod, plus an open-source training tool called SkyRL, to teach a model that can see images to navigate picture mazes. Its success rate jumped from about 44% to over 95%.

The trick is a technique called GRPO, short for Group Relative Policy Optimization. Plain version: the AI runs the same maze several times, then grades those attempts against each other. The runs that beat the group average get reinforced; the losers get dialed back. No separate "judge" model is needed, which keeps things simpler and cheaper. The model only earns a reward when it actually reaches the goal, so it has to figure out good moves on its own.

The part most businesses will care about is the plumbing. Training this kind of AI can take hundreds of hours of expensive graphics chips across several machines. If one chip dies mid-run, a normal setup loses all that progress. Amazon's system constantly checks the health of every machine, swaps out broken ones automatically, and saves its place so training resumes from the last saved step instead of starting over. Amazon says a single failure would otherwise cost hours of work.

The catch is that this is not a plug-and-play product. Following Amazon's walkthrough requires a fairly serious cloud setup — multiple GPU servers, specialized storage, and several technical installs — and it's aimed at engineers, not casual users. It's also a demonstration on a narrow task: solving mazes. Still, the pattern matters. As AI companies race to build models that can act over many steps, reliable, failure-proof training infrastructure becomes the competitive edge — and that ultimately shapes what tools, prices, and capabilities reach the rest of us.