My colleagues are morons. They're "evaluating" AI research tools and it's going about as well as you'd expect.

Once they have found a good one, they can bootstrap the process :)

What use cases are your colleagues trying to address?

@asgs novel drug pathways, can't really say more than that

@atheist "ChatGPT, how to outsmart the authorities and become a drug lord?"

@Lensflare they are at least using rag ai but it still hallucinates

@atheist that means it found a good drug, if it already hallucinates

are they using AI to evaluate which is best?

Is the AI telling em that it's the best?

It is, isn't it?

@Wisecrack they're asking how a drug effects a process in the body and it's telling them that the drug definitely effects the process in this way. I don't think they've had an answer of "it doesn't" so far...

@atheist LLM's are made to sound as agreeable as possible. I wonder if that is a side-effect of companies training them in such a way that it leads people to want to use them more, thus buying and spending more api credits?

@Wisecrack I think just a lot of their training data is stolen from the Internet or books and at best a tiny minority of it is student/teacher dialog. Even stack overflow, you get a question but the response isn't going to be "I don't know".

@Wisecrack basically LLMs are well trained in the art of improv, they "yes and" their way to the answer

@Wisecrack I think it could also be a side effect of the identity that is given to it. It is told to act like a personal assistant.

@atheist noticed the 'yes and' thing too.

@Wisecrack but like, that's literally how they work. They take a question and word by word guess which word comes next. They're not thinking ahead. Not really.

@atheist thats what my small side work will hopefully enable.

Predicting the next thought (for some definition of 'thought') as opposed to the next word.

It'd be cool to merge symbolics and graph based algorithms with sequence prediction.