Before psychedelic therapy for wartime trauma, there was narcosynthesis

Re:Before psychedelic therapy for wartime trauma, there was ... by photochemsyn on Tuesday October 10, 2023 @11:57AM (#37833826)

A LLM trained on the complete corpus of 19th century medical knowledge would be pretty useless (they had no infectious disease model). It's a fair argument that the state of modern psychological knowledge is comparable to early 20th century infectious disease knowledge (rather poor).

In terms of training corpus, I wonder how that is managed? Training LLMs on out-of-date scientific papers doesn't seem like a good idea. It might make sense to exclude primary research reports entirely and instead stick with current reviews and textbooks?

[ Parent | on HN ]

Re:Before psychedelic therapy for wartime trauma, there was ... by ipnon on Tuesday October 10, 2023 @12:26PM (#37834226)

LLMs are great at producing “free undergrads.” By that I mean it’s now easy to train a model that can produce textbook answers to textbook questions, that is well-defined solutions to well-defined problems. Modern LLMs will not be able to replace or augment physicians much because so much of medicine comes down to understanding the patient’s context.

[ Parent | on HN ]

Re:Before psychedelic therapy for wartime trauma, there was ... by ljlolel on Tuesday October 10, 2023 @01:07PM (#37834684)

LLMs understand context pretty well, that's their magic, and one thing i've noticed is that they are much more thorough than a person: they won't forget the context in the next moment or the next month. A human doctor can do better but they have to really care a lot to do better. Also, they'll only be able to do some things without attending to all necessary tasks (for example one usually overlooked is communicating with caring words).

[ Parent | on HN ]

Re:Before psychedelic therapy for wartime trauma, there was ... by j16sdiz on Wednesday October 11, 2023 @12:57AM (#37841037)

> they won't forget the context in the next moment or the next month.

GPT-4 have a context window of 32K tokens.

> (for example one usually overlooked is communicating with caring words)

If you meant keep saying "Sorry" like GPT-4 do. No, thanks, that's not caring.

[ Parent | on HN ]