Goliath @zero_goliath

RL research engineering intern @runrl_com, @uwaterloo cs; formerly @ritserlabs SF until Jan 2026 Joined March 2022

Tweets

191
Followers

607
Following

518
Likes

3K

Goliath @zero_goliath

3 days ago

continual learning is a proxy for sample efficiency which is a proxy for long context

0 0 2 62 0

Goliath @zero_goliath

a month ago

check out my terminal-bench env and DM me if you have feedback!

Johannes Hagemann @johannes_hage

a month ago

check out my terminal-bench env and DM me if you have feedback!

7 10 129 17K 27

Download Image

2 2 37 5K 9

Goliath @zero_goliath

a month ago

if you build good evals for the similarity of an AI-generated app/website to its real counterpart, you can RL llms to be better at creating economically valuable RL environments

0 0 5 234 3

Goliath @zero_goliath

2 months ago

whos SFTing an llm on donald boat tweets

0 0 2 153 0

i think given recent LLM RL progress, it's worth looking into scaling laws for RL out-of-distribution generalization again lesswrong.com/posts/65qmEJHD… especially this > a meaningful measure of distance between tasks, which is difficult and could deserve a project on its own

0 0 2 145 1

Goliath @zero_goliath

2 months ago

here are some interesting blog posts from the past few months on scaling RL. (credit to @yong_zhengxin for linking to three others) 1. ysymyth.github.io/The-Second-Hal… 2. yidingjiang.github.io/blog/post/expl… 3. kevinlu.ai/the-only-impor… 4. yongzx.substack.com/p/rl-vs-next-t… 5. blog.jxmo.io/p/how-to-scale…

0 2 10 615 5

Goliath @zero_goliath

2 months ago

there'll be a lot more research/tooling in the next 12 months on LLMs for long-horizon/realistic tasks. along the lines of - augmented PRMs - prioritized experience replay (+ envs that support snapshotting) - world-model interpretability most will be unpublished at private labs

0 0 7 6K 1

Goliath @zero_goliath

2 months ago

its nice to read old tweets and see replies that i know for sure arent AI slop

0 0 4 123 0

Goliath @zero_goliath

3 months ago

feel free to DM me about suggestions/feedback

Ritser Labs @ritserlabs

3 months ago

feel free to DM me about suggestions/feedback

1 1 13 5K 2

Download Image

0 0 4 191 0

Goliath @zero_goliath

3 months ago

what if u spot where the LLM messes up, roll back, and resample from that point to generate a new rollout. less wasted decoding on obvious tokens. u could even train a model to identify the resampling point using execution traces as priors