⿻ Andrew Trask on X: "Doing advanced logic with LLMs now has a kindof... "downhill" feeling. I feel like I can position an LLM on top of a hill by feeding in the right context. And then I give it a prompt which is... kindof... a "push" in some kind of direction. And if I gave it the right context," / Twitter

To view keyboard shortcuts, press question mark View keyboard shortcuts Home Notifications Chat Bookmarks Profile More Tweet Changbai @ChangbaiLi Tweet See new posts Conversation ⿻ Andrew Trask reposted ⿻ Andrew Trask @iamtrask Doing advanced logic with LLMs now has a kindof... "downhill" feeling. I feel like I can position an LLM on top of a hill by feeding in the right context. And then I give it a prompt which is... kindof... a "push" in some kind of direction. And if I gave it the right context, then the rube-goldberg of internal logic will lead it to the right part of the bottom of the hill. It feels like the "thinking thread" is a marble which hits these little shapes and corners I setup for it. So if my LLM gets things right... and I chase it down the hill to the bottom... and see where it ended up... and it discovered something significant... I do actually kinda feel good about myself fo rsettin gup the right rube goldberg set of constraints. Because I think the hill is like

To view keyboard shortcuts, press question mark View keyboard shortcuts Home Notifications Chat Bookmarks Profile More Tweet Changbai @ChangbaiLi Tweet See new posts Conversation ⿻ Andrew Trask reposted ⿻ Andrew Trask @iamtrask Doing advanced logic with LLMs now has a kindof... "downhill" feeling. I feel like I can position an LLM on top of a hill by feeding in the right context. And then I give it a prompt which is... kindof... a "push" in some kind of direction. And if I gave it the right context, then the rube-goldberg of internal logic will lead it to the right part of the bottom of the hi