Hiranmay Darshane
Large language models.
"Quick, Patrick, without thinking: if you could have anything right now, what would it be?"
"Um… more time for thinking."
- SpongeBob & Patrick Star
Get in touch on twitter (@hdarshane) where I'm very active.
Alternatively: darshanehiranmay [at] gmail [dot] com
Posts
Thinking out loud: evolution and pretraining
Pretraining specific behaviors vs meta-training learning systems.
Squint enough and RLing CoT reasoners is approximable as Monte Carlo Tree Search policy learning.
Exploring why CoT reasoning is fundamentally a search process and how we may push it further.
Some intuitions for RL post-training on self-supervised base models
Intuitions for RL post-training on self-supervised base models.
My lecture on the perplexing and counter-intuitive phenomena of "Grokking" in deep neural nets
Some eclectic perspectives from physics and compression to better grok the idea of grokking itself. (External link to YouTube)
Theoretical Physics could do with borrowing some ideals from AlexNet and Sutton's Bitter Lesson
Moving Beyond "Beauty", "Elegance" and suchlike Abstractional Biases and dogmas, towards more utilitarian trade-offs.
Why we must seek a Science of Systems
A case for working towards a general, application-agnostic and unified theory of Complex Systems with high explanatory power. (Featured on Hacker News front page)
On AI Agents for Desktop Computing.
Malleable, ad-hoc, on-the-fly generated interfaces; autonomous agents that work as mini-you(s).