Roy Fox (@[email protected])

Excited to share our work, "Skill Set Optimization", a continual learning method for LLM actors that: - Automatically extracts modular subgoals to use as skills - Reinforces skills using environment reward - Facilitates skill retrieval based on state allenai.github.io/sso 🧵

it's like how politicians “apologize if anyone misunderstood what I meant”; ChatGPT is a little weasel

I absolutely love how ChatGPT never actually admits any of its mistakes, instead “apologizing for the confusion”. It's like: “look, one of us seems very confused, but let's not get hung up on who. here's the plan: I won't deny partial responsibility, and you still tip me, k?”

Are you interested in causality, machine learning and healthcare? Come work with Mihaela van der Schaar (@MihaelaVDS) and me in a joint PhD or postdoc at Cambridge University, UK and the Technion, Israel Contact via email: shalit-lab AT technion ac il

@yoavgo @beenwrekt I think of expertise as having better processed info than most. But in both senses of the word, you need non-experts to convey context, and chat can enable that. Not current magic-chat-systems, for sure – those mostly enable the “imagining” part that makes demos impressive.

@roydfox @beenwrekt "retrieving info in the context of the situation" is called expertise. (and the current magic-chat-systems are not retrieving..)

In which ways are the large language model demos of 2022 more impressive than the self-driving car demos of 2012?

@yoavgo @beenwrekt the value is not in the expertise, it could even be widely known info. it's about retrieving that info in the context of my personal current situation. chat is the best interface to convey this context.

@roydfox @beenwrekt these are not really about "chat" though, are they? they are about medical expertise, financial expertise, and legal expertise.

@yoavgo @beenwrekt medical advice, financial advice, legal advice, to name a few

@roydfox @beenwrekt what are these killer chat facilitated applications you are imagining?

@yoavgo @beenwrekt @beenwrekt doesn't think the problem with driving is sensing, but I can't image what sensing looks like when it has nothing to do with remaining issues. Does it embed perceptions informatively and usably? Then the entire problem is solved, no? 3/3

@yoavgo @beenwrekt Another answer to the original question: 2022 demos feel potentially more disruptive. It's easier to imagine an infinite list of killer chat-facilitated applications than driving-facilitated ones. Unless the “killer” part is literal. 2/3

@royf/109497774626167323" target="_blank" rel="nofollow noopener">sigmoid.social/@royf/10949777…

The deadline for our #NeurIPS2022 InfoCog workshop is approaching soon [updated 🗓️: Sep 22]. We expect to have some funding to support a few selected presenters of accepted papers, and a special issue of Open Mind associated with the workshop! More info 👇 sites.google.com/view/infocog-n…

Very excited for the 2022 edition of the #neurips Deep RL workshop! A few fun changes, see below. Also, submission deadline: Sep 22

We (@BeEngelhardt, @NailaMurray and I) are proud to announce the creation of a Journal-to-Conference track, in collaboration with JMLR and conferences NeurIPS 2022, ICLR 2023 and ICML 2023! neurips.cc/public/Journal… iclr.cc/public/Journal… icml.cc/public/Journal…

📣 Very excited to announce our in-person #NeurIPS2022 workshop on Information-Theoretic Principles in Cognitive Systems! Check out our lineup of invited speakers and CFP, submit short papers by September 19 sites.google.com/view/infocog-n… #InfoCog2022 @NeurIPSConf

המעבדה שלי בטכניון מחפשת עוזרי.ות מחקר! לפרטים נוספים: drive.google.com/file/d/16jetx0… כמו כן, מגייסים סטודנטים.ות לתארים מתקדמים. אל תהססו ליצור קשר! lacclab.github.io

Policy Space Response Oracles (PSRO) mixes over a population of deep RL policies to approximate a Nash equilibrium, but exploitability can increase from one iteration to the next. We introduce Anytime PSRO which does not increase exploitability. Arxiv: arxiv.org/abs/2201.07700

Come join us tomorrow (Mon) for the #NeurIPS2021 Meaning in Context workshop! We aim to advance human-machine communication by understanding how pragmatic reasoning works in humans and how it can inform AI website: mic-workshop.github.io NeurIPS link: neurips.cc/virtual/2021/w…

This. Our top conferences are "captive regulators" of noteworthy research, and Big ML has subtly shifted their focus to experiments that have low scientific value, high environmental footprint, and — significantly — only they can run. #demeritNeurIPS

@jackclarkSF Agree that more gov investment in academic AI research would be great, but it won't address root issue. Gov money is free of incentive to profit, but cannot compete w/ industry money while there's huge incentive to profit off AI => must have regulation to stop profitable AI abuse

But this is a non-issue, because for a variety of reasons, governments have got their wallets out and built open-to-society testbeds for frontier physics, like CERN, then plugged them into academia. Why can't we make same investments for AI sector? Troubling and befuddling.

AI policy is industrial policy. Here's Putin discussing a variety of Russian AI initiatives in a lengthy speech given earlier this month: Russia wants to pass laws to ease access to data for its businesses and to spur domestic AI development. en.kremlin.ru/events/preside…