TLDR Tech / AI
90% of Expert Work Can't Be Verified by Today's AI Training Methods
Around 90% of expert work across healthcare, legal, finance, and engineering relies on subjective judgment, making it incompatible with current RLVR-style verification. To force verifiability, teams over-specify tasks and rubrics, turning real expert reasoning into shallow instruction-following and corrupting the training signal. Verification, not data scale, is the core bottleneck, and the winners will be those who can evaluate judgment-heavy, non-deterministic work without distorting it.
Read the full story