3 ms·
This is cool. I haven't seen any good/useful agents built using GPT4 yet, doesn't seem for lack of trying, but maybe some good libraries for building them will
by furyofantares 4y ago
This is cool. I haven't seen any good/useful agents built using GPT4 yet, doesn't seem for lack of trying, but maybe some good libraries for building them will help. There's some good ideas in this one.
That said, while I am very, very impressed with GPT4 for lots of uses right now, so far it's just not clear to me that feeding it back into itself is fruitful at this point.
When I use GPT4 for coding, I give it MUCH higher level instructions than I use on a search engine, much closer to the actual problem I'm solving. But I'm still breaking the problem down into smaller problems; I need to read the output, fix errors or instruct it to fix errors, and then have it build more features on. It's similar with creative processes, brainstorming, and other writing.
These agents largely strike me as an attempt to replace this whole fact-checking/editing type routine with the LLM itself; but seeing as it's the thing the LLM is not yet good at, I'm not sure how much progress can be made there, vs just waiting for GPT5 and hoping it's another big leap in capabilities.
- joshspankit 4y agoI wonder if you’ve seen ChatGPT plugins yet. If not, check out the Wolfram blog post.
- furyofantares 4y agoYes, I think there is tons of value in plugins. I also think some of the stuff going into agents is very valuable right now; prompt management, context management. (And some downstream stuff of those things, hallucination reduction, fitting it into essentially some DSL, stuff like that.) Just not sure feeding it back to itself does much at the moment.
- joshcam 4y agoAgents being used to delegate complex tasks by using an LLM to more specific models is going to be very powerful!
- furyofantares 4y agoThat doesn't really sound like an agent, so maybe "is an agent" isn't binary, in how I'm thinking about it anyway. If it 0-shots something and executes it and halts, not an agent, if it observes output and loops once, that's a minimal agent. I'm not super skeptical of a couple loops; I just don't know that GPT4 has what it takes to drive a larger number of loops.
- joshcam 4y agoYeah I worded that wrong. I was basically describing HuggingGPT. Controller is the word I meant to say, not agent. I had agents on the brain.
- mpaepper 4y agoBasically the tools the agent is using in this repository are like ChatGPT plugins.
- eggsmediumrare 4y agoI will frequently ask gpt-4 "are you sure about that" and get an improved answer, so there does seem to be some potential there.
- furyofantares 4y agoYeah it's possible that running everything more than once produces better output, at a multiple of the price and time. Presumably you usually ask that question when you have some reason to think the answer can be improved, though.