Meta’s AI agents ran over 1,000 experiments in a month but improved search by just 0.23%

Meta’s AI agents ran over 1,000 experiments in a month but improved search by just 0.23%

Summary

Meta researchers used AI agents to run 1,097 experiments over 31 days, consuming about 360,000 GPU-hours to improve its search-suggestion model. They accepted 325 changes, a 29.6% rate, and live-traffic testing found search-session volume rose 0.23% (95% confidence interval: plus or minus 0.15 percentage points). The updated model also slowed search-interface rendering; the researchers did not disclose the size of the slowdown or the campaign’s financial cost. Meta has not said whether it will adopt the model more broadly.

Latest News!