New AI Trick Catches Silent Computer Errors 5x Faster

Every computer chip occasionally makes a silent mistake. A stray cosmic ray or a hot spot can flip a single bit — turning a 1 into a 0 — and the program carries on as if nothing happened. Engineers call these 'soft errors.' In a video game, that's a harmless glitch. In software running a hospital's records, a bank's transactions, or an aircraft's navigation, it can mean wrong numbers with no warning at all.

One popular defense is duplication: run the same calculation twice and compare the answers. If they disagree, you know something got corrupted. The problem is that doing everything twice makes software roughly twice as slow — and the more complicated the program, the worse it gets. Nobody wants a safety feature that grinds their systems to a halt.

That's where the new work comes in. Two researchers, Yafan Huang and Guanpeng Li, created PaRID. Instead of testing the software while it runs — an expensive, time-consuming process — they use an AI language model to predict ahead of time which parts of the code genuinely need duplicating and which don't. The AI acts like a smart editor, deciding where protection is worth the cost. The code is then rewritten automatically at build time, so no extra monitoring is needed afterward.

On standard scientific benchmark programs, PaRID cut the performance penalty from about 163% down to roughly 60%, while catching the same number of errors — and in some cases ran up to five times faster. The catch: this is still a research paper, tested on benchmark code rather than real commercial products, and a 60% slowdown is far from free. Still, it points to a future where reliability and speed stop being a trade-off.