Currenltly, i know more negative than positive examples. At least, you are in position to change that

Mmmm

why do you think all the tasks aren't aggregated in one place, like prime intellect hub or kaggle arena?

Battleship arena for LLMs

another cool imperfect information task

www.gabegrand.com/battleship/

cc @sharky6000.bsky.social

bluesky is for american type of people

uncomfortable

Grokipedia (@Grokipedia) / X Grokipedia (@Grokipedia) / X

same type beat
x.com/Grokipedia

Can you share a screenshot, please. I have never seen one as well

le chat

Felicitations

everyone has recently started to write code so fast — entire projects are being done over a weekend

interesting how to become better at this

Large Action Models: From Inception to Implementation

Lu Wang, Fangkai Yang, Chaoyun Zhang et al.

Action editor: Edward Grefenstette

https://openreview.net/forum?id=bYdKtf0Q31

#agent #agents #ai

Thank you for the clarification

First time seeing someone calling Simplex algorithm fast, but OK. It's like calling a hashmap or A* fast. Let's check out the paper.

Nowadays, we say "chopped"

31k submissions to AAAI is fucking crazy

Replicable Reinforcement Learning with Linear Function Approximation Replication of experimental results has been a challenge faced by many scientific disciplines, including the field of machine learning. Recent work on the theory of machine learning has formalized rep...

I think I posted about it before but never with a thread. We recently put a new preprint on arxiv.

📖 Replicable Reinforcement Learning with Linear Function Approximation

🔗 arxiv.org/abs/2509.08660

In this paper, we study formal replicability in RL with linear function approximation. The... (1/6)

still true, i reckon

I now understand what you're up to. However, i mistakenly thought about PPO all the time, which is slightly different and may require additional assumptions for the convergence and, thus, deliberate notion of policy.

Sorry.

Not necessary, afaik

However, if you formulate the problem like more general stochastic optimisation over a sequence of tokens, then maybe the notion of policy is unne

RL problem, if it could be denoted as such. For example, if one wants to finetune their LLM with RL, the solution of the problem would be the policy, or a function of the model

When you present the results of any RL-based experiments, you need to evaluate your model to get the evaluation return. But with respect to what policy you get the return? That's why the notion of the policy, even for language generation, should be clear

Yes, but you still need to understand what the policy of the LLM is because the objective optimisation has to find a policy that maximises cumulative return...

Maybe due to the choice of words, it's upsetting, but a language model can indeed act a policy or strategy. Refer to the paper "Strings as strategies..." However, to be fair, the definition might be written better.

PPO is still more than a bag of heuristics, I would argue. It might have been proposed as a heuristic approximation of the trust region method, but with time has become a theoretically grounded procedure.

Btw have you got any opinion about "just a band" dead poets society?

67

YouTube video by Bob Vylan Bob Vylan - He's a Man

btw every msterpiece has its own cheap copy, like
youtu.be/anlghGgWuAs?...

what i won't add to the list - Ivan Dorn, lol

YouTube video by dan le sac Vs Scroobius Pip Thou Shalt Always Kill (Original Video) Higher Quality - dan le sac Vs Scroobius Pip

-- Dan Le Lac vs. Scroobius Pip "Thou shalt always kill"

youtu.be/CWrMGXwhFLk?...