Researchers from Google, Meta, and the University of Maryland let Claude AI figure out how to make AI systems run much faster. The AI discovered a method that uses 70% less computing power while staying just as accurate.
This is like letting a race car design its own engine – and discovering it built something better than human engineers could. The AI didn’t just copy existing methods. It invented completely new approaches that human programmers probably never would have thought of.
Done in Under 3 Hours
The whole experiment cost just $40 and took only 160 minutes to complete. The AI agent, called AutoTTS, worked independently to test different approaches and find the best solution. It focused on something called “self-consistency” – basically how AI systems double-check their own reasoning.
Traditionally, making AI more accurate meant using more computing power. But this AI-designed algorithm breaks that rule. It found clever shortcuts that maintain accuracy while dramatically reducing the computational load.
The breakthrough happened because the AI could test thousands of variations quickly, something that would take human researchers months or years. It explored combinations and approaches that humans might dismiss or never consider.
What’s fascinating is that the AI essentially taught itself to be more efficient. It’s like having a student who not only learns the subject but also invents better ways to study.
What Comes Next
This could make AI much cheaper to run for everyone. If AI systems need 70% less computing power, that means lower costs for companies and faster responses for users. The researchers plan to let AI discover improvements in other areas too.
