Microsoft on AI benchmark hacking
Public AI benchmarks like SWE-bench often fail to predict how coding agents will perform within specific corporate environments. While a model may achieve high scores on open-source tasks, these evaluations do not account for proprietary SDKs, internal archit…
The discrepancy between benchmark scores and actual utility is driven by Goodharts Law, where model providers optimize training specifically for popular evaluation targets. This creates a distributio… [+995 chars]