First Draft Club

Agents are reflecting and learning

Feels just like yesterday that OpenAI launched o1 to crickets.. only to be out-headlined by DeepSeek and their new reasoning model. I think it's easy for us to dismiss what reasoning models are capable of.. "just prompt it a little different". We still compare them against their language model brethren. They're different

Dig into the training method and you realize company after company hit ceilings on their ability to throw parameters and more data to infinitely get their models.. smarter..

I've said it before and I'll say it again. I'm less impressed the by the models and their ability to sound vaguely human, but really impressed by their ability to understand us and what we say. Now they can see and hear. But now.. they're reflecting (?)

Next frontier that we're just getting a glimpse into is having our agents be able to think ahead and learn from past mistakes. We saw the primitives of it with Yohei Nakajima trying to push the boundaries of the Pippin framework..

Building Higher apparel at High Line 🧢 
Host at /clickcast /cpg || Husband to @shiwen

I just tried it out last night with an earlier o1 conversation where we were trying to decrypt the final two zodiac letters, and it kinda did it after a couple tries. It matches some other person’s decryption. Kinda wild tbh

https://www.sfgate.com/crime/article/zodiac-unsolved-ciphers-my-name-is-solutions-16265489.php

Gpt o3 confirms

Minimally I usually have time to just let o1 rip with deep research. If I need some back and forth with mixed use 4o is still pretty good