Comment by coder-3
4 days ago
Disclaimer: talking out of my arse
Get LLMs to hillclimb your wife's treatment. The target isn't exactly verifiable but you could iterate on ways to add verification signals. There's a huge difference between asking an LLM: "hey claude, my wife has cancer, I want your help with x,y,z", which mode collapses it into extreme over-caution; versus setting up the problem like an eval, which sends LLMs into a frenzy trying to solve it no matter what.
Second thing I can share is the man who created an mRNA vaccine for his dog Rosie's cancer, and it worked! https://news.unsw.edu.au/en/meet-the-man-who-designed-a-canc...
Hey if Anthropic wants to win everyone over then this is the problem they should solve
I go down various rabbit holes with lots of different models and can get past the basic safeguards and get to real info, real discussions - usually once you get over the 'help me cure cancer' mode and into the biology of how things work they can be helpful but right now everything seems to be a dead end.
The dog cancer thing got me excited and there may still be things to explore there but I get two conflicting opinions down that path - it is too late, since she is full obstruction there is no coming back from this spiral; and it is too early since there are still drugs that have a small chance of slowing or pulling things back some.
How do you set it up like an eval?
LLMs aren't genies
Indeed. Your point?
Hmmmm....how can I phrase this sensitively?
The OP is understandably desperate for something that can save his wife. I would be too!
However, they appear to be seeking something akin to a miracle via chatting to LLMs, and the likelihood of such an approach yielding anything useful is very, very low.
6 replies →