The Bitter Lesson versus The Garbage Can

One Useful Thing · 11 months ago

The article compares two organizational approaches to AI adoption: traditional careful process mapping versus allowing AI agents to learn from outcomes alone, using the 'Bitter Lesson' from AI research (which shows that brute-force computation often outperforms encoded human expertise) as the framework. ChatGPT's new reinforcement learning-based agents, trained on desired outputs rather than process steps, produced better results than hand-crafted competitors like Manus in preliminary comparisons. The implication is that companies struggling with chaotic 'Garbage Can' organizations might bypass process mapping entirely and let AI agents find their own paths through organizational mess by simply defining desired outputs and providing training examples.