Future work idea: build my own tool — maybe a poker-solver API.

ChatGPT can follow hand logic well, but sometimes drops real clunkers. A ReAct-style loop with solver calls mid-conversation could hang in the weeds of EV calcs.

Another observation: I’ve noticed “expert AI trainer” jobs in bioinformatics, data science, chemistry, etc. The work is to solve problems and explain your thinking step by step.

That’s basically ReAct training data. Important work, but undervalued when treated as annotation.

There’s also the eyebrow-raising effect of seeing “Acknowledge user progress” in a reasoning trace right before the AI tells you “Good job!”

The illusion cracks a little.

Yoav Farbey ( @yoavf.bsky.social ) has a hilarious repo on this theme:
github.com/yoavf/absolu...

For an agent–LLM team, though, ReAct’s trick of marking up explicit Thought/Action/Observation tuples is a whole different story.

The output space grows combinatorially, but because nonsense actions are harder to justify in language, effective entropy may shrink.

The more I study AI, the more I watch myself think, and the weirder both human and computer brains seem.

For me, reasoning with vs. without verbalizing feels about the same — like the same neurons fire either way.

If you’re curious, the paper is here: arxiv.org/abs/2507.19849

Would love to hear what others think about ARPO — or if you’ve seen clever crossover ideas like using bioinformatics for GenAI training. 👀

For me, ARPO was a chance to stretch math & stats muscles and think harder about how GenAI works under the hood. The more we understand the mechanics, the more effective we’ll be at putting these systems to work.

Bioinformatics déjà vu.

ARPO’s “soft advantage attribution” compares rollouts at the token level, but once sequences diverge it gets brittle. Reminds me of sequence alignment in genomics — maybe Smith–Waterman could help here.

Tool calls are magic.

They massively extend LLM capabilities, cut down on hallucinations, and give researchers a natural breakpoint mid-calculation. You can actually peek into the stream of tokens and see what’s happening.

Reflections that stuck with me:

What does it mean to be “agentic”? The line between an algorithm and an agent feels blurry. ARPO suggests it’s about adaptability: agents curate their own tools and reasoning paths.