Giuseppe Paolo

🚀Proud to share our work on the training dynamics in Transformers with Wassim Bouaziz & @viviencabannes.bsky.social @Inria @MetaAI

📝Easing Optimization Paths arxiv.org/pdf/2501.02362 (accepted @ICASSP 2025 🥳)

📝Clustering Heads 🔥https://arxiv.org/pdf/2410.24050

🖥️ github.com/facebookrese...

1/🧵

Dumb and Dumber gif: Jim Carrey plugging his ears while his companion babbles at him about something. On top of Jim, it says "Me, Writing ICML Papers" On top of his babbling companion, it says "DeepSeek Takes".

The bottom-up approach, starting from simpler systems, is much more interesting imo.

Studying how intelligence emerges naturally from simple rules is much more likely to teach us how intelligence arose in the first place.
2/2

LLMs and today's AI research chase top-down intelligence through abstract reasoning. This path might get us to AGI faster but I don’t think it will help us understand how intelligence emerged in nature.

1/2

3- Ultimately, I believe we won't get to AGI until we solve the deeper challenge: giving true agency and meaning to these machines.

3/3

2- @garymarcus.bsky.social raises a valid point: solving ARC through training isn't true AGI, and such claims are mainly hype. Real AGI should tackle these puzzles through reasoning alone, like humans do - not by optimizing for the benchmark. And I agree with it.

2/3

I am usually quite critical of LLM hypey breakthroughs, but I don’t see an issue with OpenAI training on ARC.
There is an issue, but it’s more fundamental and multifaceted.

1- As it’s often the case, the benchmark here has become the goal. @fchollet.bsky.social also said solving ARC != AGI

1/3

By not doing research only on LLMs ;)

e introduce the effective horizon, a property of MDPs that controls how difficult RL is. Our analysis is mo- tivated by Greedy Over Random Policy (GORP), a simple Monte Carlo planning algorithm (left) that exhaustively ex- plores action sequences of length k and then uses m random rollouts to evaluate each leaf node. The effective horizon combines both k and m into a single measure. We prove sample complexity bounds based on the effective horizon that correlate closely with the real performance of PPO, a deep RL algorithm, on our BRIDGE dataset of 155 deterministic MDPs (right).

Kind of a broken record here but proceedings.neurips.cc/paper_files/...
is totally fascinating in that it postulates two underlying, measurable structures that you can use to assess if RL will be easy or hard in an environment

🤩

When I was 5, I loved science, which I took to be planets and magnets and chemicals and shit.

Soon I realized science was a collective human activity, and these were just the objects of its attention.

Anyway, whenever I read an OpEd on how science isn't political, I think "Are you 5 years old?"

I could note find any paper or work proposing AI frameworks like this.
I suppose cause it's hard, but maybe I am missing something?

Since discovering @drmichaellevin.bsky.social work, I have been fascinated by hierarchical multi-agent systems in which the hierarchy develops across different ontological levels.
So hierarchy that goes
engineers -> the company they work for,
rather than
engineers -> their boss.

Right on point, as I am getting more and more interested in this topic :)

If you're an RL researcher or RL adjacent, pipe up to make sure I've added you here!
go.bsky.app/3WPHcHg

Not sure another social media app will increase my productivity.

But at least I can justify this one as useful for work.