Guys, I get it that Anthropic also scrapes the internet to train the model. But I feel like scraping open web and distilling from a frontier model is different?
Distillation also directly inherits a frontier model’s alignment and behavior, without paying the underlying R&D or safety costs. That may be a different incentive problem than web scraping.
This feels similar (even if not identical) to a pharmaceutical company reverse-engineering a drug developed through years of costly R&D. It surely can lower prices and expand access to more people, but it’s not obvious that this is a long-term win-win situation. I don't know.
I’ve always felt it’s unfair that people attribute the “System 1/System 2” dual-reasoning idea to KT. Their research was mostly behavioral, and many of their classic psychology experiments (Linda problem, law of small numbers, representativeness bias, etc.) never mentioned two systems. The dual-process framework only emerged around the 2000s in cognitive psychology and neuroscience (e.g., Jonathan Evans, Keith Stanovich), which later provided brain-based evidence for it.
When the book came out, KT basically retrofitted their earlier behavioral work into this newer two-system framework. The book made the distinction famous, but that wasn’t really KT’s original contribution. Their biggest impact was bringing psychology into economics, i.e., prospect theory, alternative utility functions, and ultimately the creation of behavioral economics. I think people often don’t give enough credit to what they actually pioneered, and instead celebrate them for concepts they didn’t really originate.