@karpathy Yes, I do this a lot (probably too much 😅). I tend to think this as the human CoT , which could be more important context than some polished prompts for the LLM to condition its response on.
“I like this direction, but it’s optimizing the wrong notion of good.”
#ICML2026: we are presenting LILO: Bayesian optimization w/ language feedback on Thur (7/9), 10:30–12:15 Hall A #3612
https://t.co/dVrbEi6xzM
Grounded route to language-guided agent improvement. Let's chat.
Preference optimization often reduces feedback to comparisons or rankings
But decision makers often know more than “A >> B”
They can explicitly state their preferences and priors in words
LILO is a new framework turning such feedback into signal for Bayesian Optimization
🧵
Paper here: https://t.co/y4IyGQHns9
More details on how LILO turns language feedback and prior knowledge into BO-driven, uncertainty-aware optimization.
Excited to present LILO at #ICML2026.
One lens: a principled auto-research-style harness built on Bayesian optimization. LLMs surface deep prior knowledge and capture language feedback; BO turns it into uncertainty-aware search/exploration.
I’ll be at the poster. Let’s chat!
Preference optimization often reduces feedback to comparisons or rankings
But decision makers often know more than “A >> B”
They can explicitly state their preferences and priors in words
LILO is a new framework turning such feedback into signal for Bayesian Optimization
🧵
Check out our work on LILO led by @kasia_kobalczyk and @ItsMrLin. LILO allows users to perform black box optimization with natural language feedback.
Hope to see you at the poster at #ICML2026
Paper: https://t.co/d94kfL4AIB
2024 PhD Internship opportunity! Join me and my amazing colleagues on Meta's Adaptive Experimentation team to work on Bayesian optimization, probabilistic modeling / Gaussian Processes, and sample-efficient decision making: https://t.co/5uR0NNXLQU
You knew that was coming: Code LLama !!!
- Llama-2 tuned for code generation, debugging, etc
- base models, Python-specific, and instruction-tuned.
- 7B, 13B and 33B params models
- Available with the same license as Llama-2
- Blog: https://t.co/SBU5go1wcY
- Paper: https://t.co/mDws7g6Ibf
- Code: https://t.co/HR9D0iDnEk
- Models: https://t.co/dIuk96kdne
It just occurred to me that it's been a while since I used a non-dark-themed IDE. Is this the industry standard now? Miss my old Eclipse days (not really).