Hebbia's George Sivulka on why token-maxing backfires: "If you want your employees to use AI, it makes sense to force them to hit some token leaderboard or overachieve on a token expectation. It's actually not a bad way to teach your organization how to use AI. But there's one of those interesting paradoxes where the minute you choose a metric to measure, it stops being a good indicator of success." "So now that we're measuring everyone's amount of tokens, people will use tokens for all the wrong reasons. They'll use AI for things they probably shouldn't, and in a really concrete way, harm the firm because they're spending a lot of the firm's resources for the wrong reason." "There's a canonical saying that 80% of the employees do 20% of the work, and vice versa. In the same way, you'll have lots of AI use that is completely non-productive, that doesn't actually improve the company or the organization itself." "And there's a very interesting implication where AI is now starting to evoke or create the use of more AI. So with looping agents or orchestrator systems, you can ask a simple question, and it can go and create more and more work, in the same way that human organizations create a lot of more work themselves." @gsivulka @hebbia