stochi

World Models, RL & AI for Science.
Hotpot enjoyer.
Prev: Residency @PrimeIntellect @Tsinghua_Uni @ycombinator F25 @iitbombay

[stochi @stochi0]

now that I finally have a blue checkmark, maybe it's time to put my recent work back on the timeline lol

does finding an action that works imply an agent can predict the consequences of other actions?

two of my papers got into #NeurIPS 2026 workshops, one as a spotlight!

(1/3)🧵

[stochi @stochi0]

* can take a value of [‘A’,’Z’], [‘a’,’z’], ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9', '!', '@', '#', '$', '%', '^', '&', '*', '(', ')', '-', '_', '=', '+', '[', ']', '{', '}', '\\', '|', ';', ':', "'", '"', '`', '~', ',', '.', '<', '>', '/', '?']

[stochi @stochi0]

that new RL technique is:

RL*** (or) ***RL

[stochi @stochi0]

that new RL technique is:

RL*** (or) ***RL

[stochi @stochi0]

@willcb pace the frontier

[stochi @stochi0]

gui(t)=R(state(t), human preference)

[stochi @stochi0]

CUA (Computer-Using Agents) is a compatibility mechanism for legacy software

[stochi @stochi0]

anyone here who can endorse for arxiv https://t.co/6dKdet9Vc2? trying to get 2 papers up lol, would appreciate the help 😭

[stochi @stochi0]

medium rare

[Eric W. Tramel @fujikanaeda]

@leothecurious @arcee_ai exactly how rl’ed we talking, like mostly dead or gigafried?

[Yacine Mahdid @yacinelearning]

if you are looking to get into modern RLM harnesses I have a 2h30 min theoretical, practical and spiritual walkthrough of prime agent with @sethkarten

in this we’ll cover:
- RLM + continual harness + agent communication
- demo with lots of agents
- how to best use this thing

[stochi @stochi0]

@Kimi_Moonshot @NousResearch @PrimeIntellect got the aesthetic game figured out

[Yulun Du @Yulun_Du]

Aesthetics😍

[Nathan Lambert @natolambert]

TLDR on the kimi k3.1 pro benchmarks is that it looks like they figured out their RL, where the initial release was a very impressive base model with lighter post-training. China is very much keeping pace with capability gains.

[stochi @stochi0]

daymn boi🔥👀

[Kimi.ai @Kimi_Moonshot]

Meet Kimi K3.1 Pro, our most capable 3T-class model.

[stochi @stochi0]

the papers:

AgenticLS (Spotlight)
Separating Intervention Search from Response Prediction in Closed-Loop Biological Agents
https://t.co/halyMHsLKt
NEmo (Poster)
Which Predictions Are Worth Updating? Budgeted Predictive Knowledge for Long-Horizon Agents
https://t.co/wwQP3jQHFv

(3/3)

[stochi @stochi0]

now that I finally have a blue checkmark, maybe it's time to put my recent work back on the timeline lol

does finding an action that works imply an agent can predict the consequences of other actions?

two of my papers got into #NeurIPS 2026 workshops, one as a spotlight!

(1/3)🧵

[stochi @stochi0]

task success and predictive understanding aren't the same thing. why aren't we evaluating both?

i explored this distinction in biological discovery, but i think the question extends to materials science, knowledge work, and beyond.

(2/3)

[成龍 @graphtraversal]

BREAKING: github is currently experiencing unplanned uptime

[stochi @stochi0]

@graphtraversal lmao

[stochi @stochi0]

@xuanalogue Lmaooo

[Asuka Zheng🦭 @VoidAsuka]

everyone wants a miracle, but no one believes in magic.

maybe it’s time to follow the white rabbit. 🐰

[stochi @stochi0]

@mandylu buzzword final boss

[stochi @stochi0]

@mandylu He thinks it’s fresh too