Benjamin Fele

PhD student @unilj | Part of Young Researcher programme @JSI_SLO | Working with artists since 2019 | Open-ended learning is where it's at.

[Joscha Bach @Plinz]

@robleclerc Yes. But the secret to a healthy and thriving subculture is to prevent it from being too useful and economically valuable, so it is not overrun with people who want fame, status and money. A-Life is way more fun than most other AI or bio conferences

[Benjamin Fele @mc_dropout]

@MorePerfectUS @grok where can i get the full interview?

[Cong Lu @cong_ml]

Delighted to share our new work on automated open-ended capability discovery for frontier models!

Model eval is getting harder with more capable models and so we ask: can models self-explore and tell us their capabilities? 🔍🕵️‍♂️

Many thanks to @shengranhu and @jeffclune ❤️❤️❤️

[Sergey Levine @svlevine]

Scaling laws in deep RL? Turns out that batch size, learning rate, and UTD (update-to-data) for getting the most efficient and scalable deep RL has predictable relationships. Checkout the analysis in new work by @_oleh & collaborators: https://t.co/QnX7EKX6yy

[Clément ROMAC @ ICML 2025 @ClementRomac]

Great to see an extension of GLAM applied to larger LLMs and more complex environments ➡️ https://t.co/TPtHuqqsas 🔥

GLAM can be used to align and functionally ground LLMs in external environments' dynamics through Online RL (
https://t.co/zNjFu7XBsf)

TWOSOME is used on...
1/4

[Sergey Levine @svlevine]

Language models trained with RL should do much more than just satisfy preferences. They can achieve conversational goals, do strategic reasoning, etc. But we need the right tasks to evaluate progress on this. LMRL gym aims to provide this: https://t.co/keM7nV2nSn

A thread 👇

[Benjamin Fele @mc_dropout]

🥲

[Pedro Domingos @pmddomingos]

DeepMind’s big bet was deep reinforcement learning, but it hasn’t panned out on any real-world problems.

[Guillaume Lample @ NeurIPS 2024 @GuillaumeLample]

Mistral 7B is out. It outperforms Llama 2 13B on every benchmark we tried. It is also superior to LLaMA 1 34B in code, math, and reasoning, and is released under the Apache 2.0 licence.
https://t.co/krGs0xwbLH

[pete wolfendale @deontologistics]

To pursue something for its own sake means not subordinating one’s epistemic drive to understand that thing to the thing as defined by a fixed conception that has been imposed upon you from without. Herein lies the delicate connection between autonomy and general intelligence.

[Gary Marcus @GaryMarcus]

Unlike airplanes, cars, nuclear energy, food, and medicine, AI is magically immune from serious problems!

Anyone who thinks otherwise is a “grifter”!

[Tivadar Danka @TivadarDanka]

One major reason why mathematics is considered difficult: proofs.

Reading and writing proofs are hard, but you cannot get away without them. The best way to learn is to do.

So, let's deconstruct the proof of the most famous mathematical result: the Pythagorean theorem.

[David Pfau @pfau]

The reason AI isn't more of a rigorous science is that the kind of person drawn to the idea of automating thinking does not actually have that much interest in thinking.

[Cassidy Laidlaw @cassidy_laidlaw]

Excited to present our new paper on bridging the theory-practice gap in RL! For the first time, we give *provable* sample complexity bounds that closely align with *real deep RL algorithms'* performance in complex environments like Atari and Procgen.

[Imbue @imbue_ai]

Learn about designing simpler and more principled RL algorithms and much more with @ben_eysenbach from CMU!

Blog: https://t.co/MU39poLVJH
Video: https://t.co/EORozXGU7r

[Jascha Sohl-Dickstein @jaschasd]

The hot mess theory of AI misalignment (+ an experiment!)
https://t.co/OukfipSkIJ

There are two ways an AI could be misaligned. It could monomaniacally pursue the wrong goal (supercoherence), or it could act in ways that don't pursue any consistent goal (hot mess/incoherent).

[Benjamin Fele @mc_dropout]

https://t.co/8XgHn9xiUl

3 days later

[AK @_akhaliq]

Toolformer: Language Models Can Teach Themselves to Use Tools

introduce Toolformer, a model trained to decide which APIs to call, when to call them, what arguments to pass, and how to best incorporate the results into future token prediction

abs: https://t.co/EimSc0SBXs

[Benjamin Fele @mc_dropout]

How long until we see language models grounded in APIs, enabling them to interact with real-world and trained with RL from human feedback? This could enable applications from robot control to user interface navigation.

[Kevin Fischer @kevinafischer]

This paper is not receiving enough attention:

GPT 3.5 displays emergent theory of mind

https://t.co/wOcen3ZVTw

[Derek Thompson @DKThomp]

I'll be thinking about this paper for a while: "Theory of Durable Dominance"

https://t.co/YlfHOKrosX

Upshot: "Contrary to economic assumption that increased competition displaces [dominant firms], competition instead entrenches dominants”

TLDR: Superstar effects in everything