Part 10/11:
Initially popular in the 1980s, expert systems failed due to their rigidity and inability to handle nuanced exceptions. Now, OpenAI aims to fine-tune large language models with curated datasets of expert knowledge—all in the hope of creating specialized AI "agents" for fields such as biology, law, or engineering.
Criticisms and Challenges
Shapiro questions whether handcrafted datasets are sufficient or if more dynamic, adversarial training—akin to peer review—would better capture experts’ nuanced reasoning. He suspects that true breakthroughs lie in more distributed, generative, and interactive training approaches rather than static fine-tuning.