New AI Trick Makes Software Run 70% Faster Without Human Help
This could mean faster apps and lower costs for everyone.
Deep Dive
Researchers introduced BOXBOX, which enables agent-triggered harness revision and kernel-managed hot reloading while preserving execution context, so agents can use task-specific experience to recursively improve their harnesses during an ongoing task. On the new PerfRace benchmark for long-running software performance optimization, BOXBOX reached a Speedup@4h of 1.70x under matched test-time budgets—higher than Frozen Harness, offline Harness Evolution, and Harness Scaling, with the strongest baseline at 1.54x—and used task time and model cost more efficiently.
Key Points
- BOXBOX lets AI improve its own problem-solving tools while working, without needing to stop and restart.
- In tests, it sped up software by 70% in 4 hours, beating other methods.
- This could lead to faster, more efficient apps and lower costs, but it's still experimental.
Why It Matters
Faster software means quicker apps, lower costs, and smarter AI that adapts on the fly.