Executives didn't accidentally waste AI tokens - they deliberately pushed employees to burn through them as a blunt force tactic to overcome resistance to AI adoption. At Meta, performance reviews tied to token counts led workers to run pointless agent conversations just to boost numbers. This wasn't stupidity; it was a calculated push to get skeptical teams actually using AI tools.

The strategy worked. Teams that once rejected AI now use Cursor and other tools, but Meta eliminated its unlimited token policies as subscription subsidies vanished and API costs rose. However, the token spending shift reflects deeper technical changes. We've moved from 'compounding error' - where more tokens meant worse outcomes - to 'compounding correctness,' where increased token investment actually improves results.

This creates new strategic opportunities, particularly in cybersecurity. Anthropic's Mythos exemplifies this - an LLM so capable at finding exploits that the company restricted access to critical software makers. The economics are stark: $12,500 per attempt, $125k for ten runs, with no diminishing returns observed. Security becomes a proof-of-work problem where victory goes to whoever spends more tokens.

Tools like 'loops' - continuously running agents without human oversight - enable this new paradigm. Boris Cherny's demonstration of loop-based AI workflows sparked excitement because they allow sustained, unsupervised problem solving. The era of wasteful tokenmaxxing may be over, but strategic token investment for genuine productivity gains has just begun.