very underrated: give an LLM a list of resources you like and ask it to surface related resources like books, blog posts, etc. lots of esoteric learning material out there.
@psswordprotectd I have a “Thought Explorer“ project with Claude whose purpose is to help me connect ideas I’m curious about to other thinkers and similar ideas / sources for me to explore.
It’s been very very edifying and I’ve enjoyed a few books it suggested looking into.
The value of knowledge defined as availability of information has gone down over the last 30 years. The value of knowledge as input into judgement is steadily increasing.
@AnnaRMills Both have local modes and models that keep your data on your computer, but I'd confirm their policies to be on the safe side.
In many cases the transcription is followed by a formatting step that uses an LLM, which can sometimes use a hosted model.
@InderosD Have you found this more effective than running entire workflow directly in RepoPrompt? I've been searching for people using Fable with it, but this is the only tweet I've found so I had to ask.
You have Claude Fable for only a few days. Here's how to make the most of it.
Introducing /improve: use your most capable model to audit your codebase and write plans for cheaper models to execute later.
Studies your code, figures out bugs, perf, tech debt, missing tests, what to build and writes plans any agent can run.
@sebgoddijn The way I see it, I try and do a subscription cost’s worth of work on the 1st day of the cycle, so the rest feels like a bonus.
Not the most bang for buck or tokenmaxxing approach, but it helps keep me sane.
@rameswar08@_simonsmith I think it depends on the user. Under a certain monthly price threshold, I’d rather default to the best models on each platform and adjust the reasoning effort based on how important the task is.
Immediate feedback works best for simple tasks. Delayed feedback works better for complex ones.
Because a brief delay lets the learner self evaluate first.
The human teacher who returns thoughtful comments two days later might be providing more effective feedback than any AI.
"Every gap you leave, Claude fills with in-distribution choice." – @trq212 at @CAISconf
I keep thinking about this quote, contrasted with the market and users creating and asking for features and apps that let them provide LESS input. 🫠
@pzakin There are certain tasks that today’s LLM’s can already perform excellently, but this made me think of there might be more ambitious tasks within a domain that would benefit from faster and cheaper access to frontier-level intelligence or larger context windows.