Stanford's Yoonho Lee argues that text optimization methods like prompting are undervalued compared to gradient-based weight training
The methods offer superior sample efficiency in low-data settings.
Most Activity
as usual from @yoonholeee, this is extremely well-written and a great way to organize the major arguments in this space
http://x.com/i/article/2064017981982859264
Highly recommended post by @yoonholeee discussing rich new research directions!
Harness engineering will come to an end. An era of harness learning is in front of us, with massive room for empirical and theoretical research on data, architectures, and algorithms.
http://x.com/i/article/2064017981982859264
Excellent piece by @yoonholeee!
Harness should be learned, not engineered, and you'll find rich research directions on harness optimization in Yoonhoo's post.
http://x.com/i/article/2064017981982859264
Excellent piece by @yoonholeee!
Harness should be learned, not engineered. You'll find rich research directions on harness optimization in Yoonhoo's post.
http://x.com/i/article/2064017981982859264
Highly recommended post by @yoonholeee discussing rich new research directions!
Harness engineering will come to an end. An era of harness learning is in front of us, with massive room for empirical and theoretical research on data, architectures, and algorithms.
http://x.com/i/article/2064017981982859264

@yoonholeee @a1zhang I started making Tools For This, glitchlings has tools for corrupting text as well as a starter toolkit for comparative sequence analysis (and lots of integrations), all FL/OSS 🫡
https://github.com/osoleve/glitchlings

@lateinteraction @yoonholeee rare to see someone break it down without taking sides tbh
what stood out to u the most from it?

@yoonholeee ^^ I'm building alot of this. Some differences. Improvement has to be gated, versioned for compliance. But there's ways to have best of both worlds. https://chatgpt.com/share/6a2731ea-d2d8-8330-a501-6f8001b7bb43