Lightnews — Scholar-powered news

Tom Schaul @schaul.bsky.social · Aug 27

@rl-conference.bsky.social

Tom Schaul @schaul.bsky.social · Aug 27

Where do some of Reinforcement Learning's great thinkers stand today?

Find out! Keynotes of the RL Conference are online:
www.youtube.com/playlist?lis...

Wanting vs liking, Agent factories, Theoretical limit of LLMs, Pluralist value, RL teachers, Knowledge flywheels
(guess who talked about which!)

1 23 75

Reposted by Tom Schaul

Aditi Mavalankar @aditimavalankar.bsky.social · Jul 13

On my way to #ICML2025 to present our algorithm that strongly scales with inference compute, in both performance and sample diversity! 🚀

Reach out if you’d like to chat more!

2 8

Tom Schaul @schaul.bsky.social · Jun 2

Deadline to apply is this Wednesday!

Tom Schaul @schaul.bsky.social · May 22

Ever thought of joining DeepMind's RL team? We're recruiting for a research engineering role in London:
job-boards.greenhouse.io/deepmind/job...
Please spread the word!

Research Engineer, Reinforcement Learning

London, UK

job-boards.greenhouse.io

1 4

Tom Schaul @schaul.bsky.social · May 24

The RL team is a small team led by David Silver. We build RL algorithms and solve ambitious research challenges. As one of DeepMind's oldest teams, it has been instrumental in building DQN, AlphaGo, Rainbow, AlphaZero, MuZero, AlphaStar, AlphaProof, Gemini, etc. Help us build the next big thing!

1 2

Tom Schaul @schaul.bsky.social · May 22

Ever thought of joining DeepMind's RL team? We're recruiting for a research engineering role in London:
job-boards.greenhouse.io/deepmind/job...
Please spread the word!

Research Engineer, Reinforcement Learning

London, UK

job-boards.greenhouse.io

1 8 28

Tom Schaul @schaul.bsky.social · Mar 20

When faced with a challenge (like debugging) it helps to think back to examples of how you've overcome challenges in the past. Same for LLMs!

The method we introduce in this paper is efficient because examples are chosen for their complementarity, leading to much steeper inference-time scaling! 🧪

Aditi Mavalankar @aditimavalankar.bsky.social · Mar 17

Excited to share our recent work, AuPair, an inference-time technique that builds on the premise of in-context learning to improve LLM coding performance!
arxiv.org/abs/2502.18487

🧵

AuPair: Golden Example Pairs for Code Repair

Scaling up inference-time compute has proven to be a valuable strategy in improving the performance of Large Language Models (LLMs) without fine-tuning. An important task that can benefit from additio...

arxiv.org

5 18

Tom Schaul @schaul.bsky.social · Feb 24

Some extra motivation for those of you in RLC deadline mode: our line-up of keynote speakers -- as all accepted papers get a talk, they may attend yours!

@rl-conference.bsky.social

RLC Keynote speakers: Leslie Kaelbling, Peter Dayan, Rich Sutton, Dale Schuurmans, Joelle Pineau, Michael Littman

10 36

Tom Schaul @schaul.bsky.social · Jan 31

200 great visualisations: 200 facets and nuances of 1 planetary story.

Nat Bullard @nathanielbullard.com · Jan 30

My annual decarbonization presentation is here.

200 slides, covering everything from water levels in Lake Gatún to sulfur dioxide emissions to ESG fund flows to Chinese auto exports to artificial intelligence. www.nathanielbullard.com/presentations

5

Tom Schaul @schaul.bsky.social · Jan 30

The sound of two users joining per second: "tik", "tok"...

4

Reposted by Tom Schaul

Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · Jan 27

Reposting David Silver's talk about how RL is the way to intelligence. No particular reason
www.youtube.com/watch?v=pkpJ...

David Silver - Towards Superhuman Intelligence - RLC 2024

YouTube video by Reinforcement Learning Conference

www.youtube.com

7 71

Reposted by Tom Schaul

Reinforcement Learning Conference @rl-conference.bsky.social · Jan 8

Excited to announce the first RLC 2025 keynote speaker, a researcher who needs little introduction, whose textbook we've all read, and who keeps pushing the frontier on RL with human-level sample efficiency

Announcement of Richard S. Sutton as RLC 2025 keynote speaker

4 51

Tom Schaul @schaul.bsky.social · Dec 14

Could language games (and playing many of them) be the renewable energy that Ilya was hinting at yesterday? They do address two core challenges of self-improvement -- let's discuss!

My talk is today at 11:40am, West Meeting Room 220-222, #NeurIPS2024
language-gamification.github.io/schedule/

Tom Schaul @schaul.bsky.social · Nov 28

Are there limits to what you can learn in a closed system? Do we need human feedback in training? Is scale all we need? Should we play language games? What even is "recursive self-improvement"?

Thoughts about this and more here:
arxiv.org/abs/2411.16905

Boundless Socratic Learning with Language Games

An agent trained within a closed system can master any desired capability, as long as the following three conditions hold: (a) it receives sufficiently informative and aligned feedback, (b) its covera...

arxiv.org

1 27

Tom Schaul @schaul.bsky.social · Dec 10

Don't get to talk enough about RL during #neurips2024? Then join us for more, tomorrow night at The Pearl!

Reinforcement Learning Conference @rl-conference.bsky.social · Dec 10

If you're at NeurIPS, RLC is hosting an RL event from 8 till late at The Pearl on Dec. 11th. Join us, meet all the RL researchers, and spread the word!

14

Tom Schaul @schaul.bsky.social · Dec 9

Dynamic programming has a fun origin story. In 1950, Bellman wanted to coin a term that "was something not even a Congressman could object to".
See here:
pubsonline.informs.org/doi/pdf/10.1...

pubsonline.informs.org

1 5

Tom Schaul @schaul.bsky.social · Dec 2

This year's (first-ever) RL conference was a breath of fresh air! And now that it's established, the next edition is likely to be even better: Consider sending your best and most original RL work there, and then join us in Edmonton next summer!

Reinforcement Learning Conference @rl-conference.bsky.social · Dec 2

The call for papers for RLC is now up! Abstract deadline of 2/14, submission deadline of 2/21!
Please help us spread the word.
rl-conference.cc/callforpaper...

RLJ | RLC Call for Papers

rl-conference.cc

3 19

Tom Schaul @schaul.bsky.social · Dec 2

Ohh... good morning to you too!

Clearly this got off on the wrong foot: do you want to try again, maybe more constructively (in the spirit of bluesky not being the other place)? This is a preprint, so I'd be happy to hear your suggestions for making it less "ignorant"...

1 4

Tom Schaul @schaul.bsky.social · Nov 30

Either one or many players. For "improvement" to be well-defined, one agent must be special (see footnote 6), but the multi-agent setting has many benefits.

1 1

Tom Schaul @schaul.bsky.social · Nov 29

1: open-ended means that it will keep producing novel and learnable artifacts (see the definition here: arxiv.org/abs/2406.04268), on the timescale of interest for the observer.

2: I think as a thought experiment it is valid, as it could work in principle, but of course it hasn't been built?

Open-Endedness is Essential for Artificial Superhuman Intelligence

In recent years there has been a tremendous surge in the general capabilities of AI systems, mainly fuelled by training foundation models on internetscale data. Nevertheless, the creation of openended...

arxiv.org

1 1

Tom Schaul @schaul.bsky.social · Nov 28

In section 5 (second paragraph), there's about a dozen references to language games people are already using (one per paper), some with ingenious ways to provide feedback.

Also, I suspect the workshop will ultimately have the poster abstracts online with plenty of additional material!

1 1

Tom Schaul @schaul.bsky.social · Nov 28

I'll also be giving a talk about this at the @neuripsconf.bsky.social workshop on "Language Gamification" in two weeks. Pop by if you're around!

language-gamification.github.io

Language Gamification Workshop 2024

Join us for the Language Gamification Workshop 2024, featuring keynote speeches, panel discussions, and poster sessions.

language-gamification.github.io

2 1 17

Tom Schaul @schaul.bsky.social · Nov 28

Are there limits to what you can learn in a closed system? Do we need human feedback in training? Is scale all we need? Should we play language games? What even is "recursive self-improvement"?

Thoughts about this and more here:
arxiv.org/abs/2411.16905

Boundless Socratic Learning with Language Games

An agent trained within a closed system can master any desired capability, as long as the following three conditions hold: (a) it receives sufficiently informative and aligned feedback, (b) its covera...

arxiv.org

7 18 110

Tom Schaul @schaul.bsky.social · Nov 28

@colah.bsky.social: with a few years' hindsight, how do you see the Distill space now? Is there a chance for a reboot or a rebirth in another form?

2

Tom Schaul @schaul.bsky.social · Nov 27

I think the Distill journal was really valuable in this space, but unfortunately is no longer around to help...

distill.pub

Distill — Latest articles about machine learning

Articles about Machine Learning

distill.pub

2 12

Tom Schaul @schaul.bsky.social · Nov 25

If you're happy with a book-length answer (to the broader question on which technologies empower whom, why, and when), Acemoglu and Johnson have some excellent analysis:
shapingwork.mit.edu/power-and-pr...

POWER and PROGRESS - Massachusetts Institute of Technology

Discover a bold reinterpretation of economics and history in Power and Progress, the latest book by bestselling authors Daron Acemoglu and Simon Johnson. Explore their 1000-year struggle over technolo...

shapingwork.mit.edu

1