'AI Finds a Way' AI routinely surprises & outwits us, cleverly hacking reward functions, finding loopholes, and doing unexpected, undesired, & unsafe things. Because they're fascinating, & to inform AI safety discussions, we're collecting/verifying such anecdotes.Please submit!👇

Thought Cloning covered in Science. Congrats Shengran Hu! I'm extremely excited about the possibilities of Thought Cloning, so it's nice to see coverage of it in the co-top science journal worldwide. Thanks Matthew Hutson Matthew Hutson for a great article!
science.org/content/articl…

Artificial intelligence may benefit from talking to itself.

My latest for Science Magazine News from Science, on work at NeurIPS Conference #NeurIPS2023 , by Shengran Hu and Jeff Clune at UBC Computer Science. Thanks also Shunyu Yao and annaborghi. science.org/content/articl…

Here is a YouTube link to my #NeurIPS2023 invited talk 'Open-ended and AI-generating Algorithms in the Era of Foundation Models.' Includes exciting new OMNI results in an infinite task space. Lmkwyt! youtu.be/TO_xQzWXTcI NeurIPS Conference ALOE Workshop Goal-Conditioned RL Workshop Joseph Suarez (e/🐡)

I will be at #NeurIPS next week to present our work Thought Cloning at Poster Session 4 on Wed 5-7pm (#1421). Welcome to check it out!

Also happy to chat about any topic on open-endedness, language models, deep RL, etc. Let me know if you wanna meet 😃

Introducing Thought Cloning

Thought Cloning could enable a revolutionary leap in AI capabilities. For the first time, agents would not just blindly mimic human behaviors, but gain insight into the underlying thought processes behind those behaviors. Just as language transformed

I am delighted to report that Thought Cloning was not only accepted to NeurIPS NeurIPS Conference , but awarded a spotlight. Thanks to the AC, reviewers, and community. A huge congrats to Shengran Hu! #NeurIPS #neurips2023

Autonomous driving with Chain of Thought - autopilot thinking out loud in text!

LINGO-1 is the most interesting work I've read in autodriving for a while.

Before: perception -> driving action
After: perception -> textual reasoning -> action

LINGO-1 trains a video-language

I spoke with the Globe & Mail about my concerns regarding AI existential risk, along with others like Geoffrey Hinton David Duvenaud Gillian Hadfield Sheila McIlraith (and skeptical takes by others). I think it is helpful to be having this conversation publicly
theglobeandmail.com/business/artic…

Very cool work!

Just saw this. Amazing.
Feels like it will be one of the keystone concepts in AI. I love how it supports safety (w/o requiring decisions about *what* is safe a priori during training/tuning/RLHF), interpretability, and steerability, all while improving performance.

Clone your thoughts with AI

Introducing Thought Cloning AI!

This groundbreaking framework empowers agents to think like humans.

It does not this by replicating the behaviour of humans.

It however does not stop at that; it captures the very thoughts humans have while

Introducing Thought Cloning: AI agents learn to *think* & act like humans by imitating the thoughts & actions of humans thinking out loud while acting, enhancing performance, efficiency, generalization, AI Safety & Interpretability. Led by Shengran Hu
arxiv.org/abs/2306.00323 1/5