Juan Pavez (@juanpavez) 's Twitter Profile
Juan Pavez

@juanpavez

PhD (c) in Computer Science. Working in AI for Science. Coding for food and fun.

ID: 64958584

calendar_today12-08-2009 06:17:52

301 Tweet

644 Followers

1,1K Following

Juan Pavez (@juanpavez) 's Twitter Profile Photo

Prompting is weird - you can spend an entire day working on a prompt, only to realize that you just had to change one word that the model misinterpreted... Yet, when I think about it, it's not much different from optimizing a large network ๐Ÿ˜… 1e-4 can make a big difference

Juan Pavez (@juanpavez) 's Twitter Profile Photo

Any repetitive task that takes longer than a minute should be replaced with a ChatGPT prompt... I have to remind myself

Juan Pavez (@juanpavez) 's Twitter Profile Photo

Nobody could create anything interesting if all their code files are only 10 lines and their commits are just a few lines of code .... if you've never become obsessed with a code and ended up writing 100 lines full of inspiration, then perhaps copilot could do it

Juan Pavez (@juanpavez) 's Twitter Profile Photo

Holy dios, that is a killer app! People who say that AI products are just a feature of other products sure used to say that internet was just a feature for blockbuster

Juan Pavez (@juanpavez) 's Twitter Profile Photo

There are simple and deterministic tasks, like processing data from an Excel, and there are very complex and non-deterministic tasks, like managing a company... I believe the sweet spot for AI agents currently lies right in the middle

Juan Pavez (@juanpavez) 's Twitter Profile Photo

Friday night ... watching how one of my bots provided therapy to another for 20 minutes. Is this the future? It's like season 10 of 'In Treatment' hahaha

Juan Pavez (@juanpavez) 's Twitter Profile Photo

There has been much discussion about AI providing therapy, but did you know that bots are already undergoing therapy among themselves? ๐Ÿ˜œ This demo, inspired by Kevin Fischer great demo, shows how agents can follow evidence-based therapy protocols while still being natural and

Juan Pavez (@juanpavez) 's Twitter Profile Photo

I spent the weekend completely refactoring my multi-agent framework, and I think I'll have to give "The Society of Mind" another read, possibly some good ideas for AGI can be found there ๐Ÿค”

I spent the weekend completely refactoring my multi-agent framework, and  I think I'll have to give "The Society of Mind" another read, possibly  some good ideas for AGI can be found  there ๐Ÿค”
Juan Pavez (@juanpavez) 's Twitter Profile Photo

I'm still figuring out the best way to include tables in prompts. Right now, I think turning them into a list or JSON is the way to go. Basically, you squish the column and row names into one line with the value, something like 'Row ID, Column Name, Value'. In this way, it is

Juan Pavez (@juanpavez) 's Twitter Profile Photo

Best programming opinion I've found in the internet so far: Facebook wrote GraphQL to solve THEIR problems. If you have decided to just copy it you probably are stupid

Juan Pavez (@juanpavez) 's Twitter Profile Photo

I was skeptical, but honestly, o3 feels like a qualitative step into the future .. especially with deep research... oh my god. It's hard to accept that things you've worked so hard to become good at can be surpassed so easily by AI ... But if you don't adapt, you die .. Haha