Model WatchBenchmark disagreement is a product signalWhen frontier-model evaluations disagree, the answer is not to pick a favorite score. It is to find the work your product actually needs measured.September 3 2026//2 min read#benchmarks#model evaluation#ai products#frontier modelsOpen post→
Developer ToolsSWE-bench stopped being enoughWhen a benchmark stops separating frontier coding systems, the repo needs local evals more, not less.April 8 2026//3 min read#benchmarks#coding agents#evalsOpen post→