Constantin Ruhdorfer

📸👨🏽‍💻 -PhD Student at the intersection between Theory of Mind, MARL and Human-AI cooperation at IMPRS-IS and the HCI-CS department at Stuttgart University

[Artificial Intelligence Papers @SciFi]

Explicit Modelling of Theory of Mind for Belief Prediction in Nonverbal Social Interactions. https://t.co/CvhU3YxB12

[Constantin Ruhdorfer @CRuhdorfer]

And another one from Matteo Bortoletto! In our work we test different cognitively inspired neural network architectures in multi-modal social interactions to reason about the mental states of humans. He will present the work at @ecai2024! ArXiv: https://t.co/eAu6VpRvv1

[Neel Nanda @NeelNanda5]

I wrote a mech interp reading list two years ago, but a LOT has changed - announcing v2!

This is a highly opinionated list of favourite papers, not a lit review - I try to summarise key takeaways, critique, advise on which parts to prioritise vs skip, and explain WHY I like it!

[Artificial Intelligence Papers @SciFi]

Benchmarking Mental State Representations in Language Models. https://t.co/p5yNfSfhon

[Constantin Ruhdorfer @CRuhdorfer]

Shout out to Matteo for leading this delightful study on mental state representations in language models! Talk to him at the Mechanistic Interpretability workshop at ICML if you are there about language models, mechanistic interpretability and Theory of Mind!

[Micah Carroll @MicahCarroll]

Important new work!

I somewhat disagree with the conclusions though: imo persuasion capabilities will very likely surpass this curve once key players start performing "persuasion RL training/unhobbling" (beyond RLHF). Unfortunately, there will be many incentives to do this!

[Constantin Ruhdorfer @CRuhdorfer]

@jsuarez Interesting catch! I never accumulate but do you have any insight into whether one should? Is the pseudocode or implementation „correct“?

[Constantin Ruhdorfer @CRuhdorfer]

@_chris_lu_ @JonnyCoook Ah thats great, thanks!

[Shangtong Zhang @ShangtongZhang]

We prove that transformers can implement temporal difference learning in the forward pass. It's not only that TFs behave like TD, it's that the forward pass over layers is exactly equivalent to iterations of TD! https://t.co/blLpDdPPG3 w/ @wangjiuqi, Ethan, @HadiDaneshmand

[Constantin Ruhdorfer @CRuhdorfer]

Very excited about Matteo presenting our paper at #ACL2024 this year and congratulations for making it into the main conference!

[Artificial Intelligence Papers @SciFi]

Limits of Theory of Mind Modelling in Dialogue-Based Collaborative Plan Acquisition. https://t.co/bS1YBorVSu

[Constantin Ruhdorfer @CRuhdorfer]

@sai_prasanna From which paper is that?

[Constantin Ruhdorfer @CRuhdorfer]

@aaron_defazio Okay, thanks! I think I saw someone post an optax implementation here recently so might give it a shot over the weekend

[heiner @HeinrichKuttler]

OK fine I'll tap the sign

https://t.co/rnp0bMP9KN

[Constantin Ruhdorfer @CRuhdorfer]

@LukasSchaefer96 Started learning it today as a PyTorch person, love the functional nature but not trivial at all! But this work absolutely delivers

[Constantin Ruhdorfer @CRuhdorfer]

@_chris_lu_ @alexrutherford0 @benjamin_ellis3 @MatteoGallici Awesome work! This might be the convincing I needed to learn JAX after all

[Constantin Ruhdorfer @CRuhdorfer]

@MariusHobbhahn @80000Hours Could be a nice addition to the after-hours feed!

[Nathan 🔎 @NathanpmYoung]

Effective altruism, longtermism, catastrophic risk and billionaires

- why reducing catastrophic risk matters
- fitting it with global poverty and animal welfare
- why billionaires are funding it

Many people helped with this, errors my own