Yonatan Bisk's Avatar

Yonatan Bisk

@ybisk.me.bsky.social

Assistant Professor confused by the concept of consciousness but talkingtorobots.com in the meantime

2,381 Followers  |  229 Following  |  9 Posts  |  Joined: 23.02.2024  |  1.9113

Latest posts by ybisk.me on Bluesky

Video thumbnail

We are getting closer to have agents operating in the real physical world. However, can we trust frontier models to make embodied decisions ๐ŸŽฎ aligned with human norms ๐Ÿ‘ฉโ€โš–๏ธ ?

With EgoNormia, a 1.8k ego-centric video ๐Ÿฅฝ QA benchmark, we show that this is surprisingly challenging!

04.03.2025 04:32 โ€” ๐Ÿ‘ 23    ๐Ÿ” 9    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 1

How many (checks calendar) decades do people keep around backups of data from their thesis? Am I a digital hoarder?

05.01.2025 05:00 โ€” ๐Ÿ‘ 10    ๐Ÿ” 0    ๐Ÿ’ฌ 4    ๐Ÿ“Œ 0

Recently, papers have been published in prestigious journals (Nature Human Behaviour, PNAS) claiming that large language models (e.g., ChatGPT) solve the "false belief" task (a task requiring Theory of Mind abilities).

What is the false belief task? ->

17.12.2024 08:36 โ€” ๐Ÿ‘ 6    ๐Ÿ” 2    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

I think Bisky :)

25.11.2024 20:12 โ€” ๐Ÿ‘ 2    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0

When I first started reading papers in ~2007 we would wait for someone who attended the conference to bring back the conference booklet and tell us what to read. Weโ€™d read them. And then spend the next three months reading old papers or working cuz ๐Ÿคท It was a great way to grow up.

24.11.2024 14:52 โ€” ๐Ÿ‘ 9    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0

๐Ÿ‘‹ Time is a weird thing :) maybe I should try and convince the department to force you to give a talk sometime โ€” though ideally with @spandanagella.bsky.social too ;)

24.11.2024 02:30 โ€” ๐Ÿ‘ 1    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

Just use your own domain name?

24.11.2024 02:13 โ€” ๐Ÿ‘ 0    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

This article really spoke to me; all the science I've enjoyed and that I thought came out well has been done with a colleague that I was talking to every day and almost every couple of hours

17.11.2024 14:32 โ€” ๐Ÿ‘ 42    ๐Ÿ” 3    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

Hello, Computational linguistics/NLP world in Bluesky! We're creating the same accounts on other social media platforms in Bluesky! #NLProc

14.11.2024 00:17 โ€” ๐Ÿ‘ 133    ๐Ÿ” 31    ๐Ÿ’ฌ 4    ๐Ÿ“Œ 5

I am trying to create a robotics and ai starter pack on bluesky: go.bsky.app/DfAoaJ1

Very incomplete please comment with suggestions (or just if you're missing and want to be added!)

11.11.2024 15:01 โ€” ๐Ÿ‘ 111    ๐Ÿ” 38    ๐Ÿ’ฌ 77    ๐Ÿ“Œ 4
Post image

3. How to Train Your Fact Verifier: Knowledge Transfer with Multimodal Open Models
^^ includes @skgabrie.bsky.social who is just starting up her lab at UCLA!

10.11.2024 18:34 โ€” ๐Ÿ‘ 4    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0

2. Gradient Localization Improves Lifelong Pretraining of Language Models
TL;DR - Gradient norms tell you where your knowledge is stored and if it conflicts with what you already know.

10.11.2024 18:34 โ€” ๐Ÿ‘ 2    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0
Post image

#EMNLP2024

1. Tools Fail: Detecting Silent Errors in Faulty Tools

Are you using tools with your LLMs? Are you assuming your tools are perfect? Assuming the LLM can just handle any errors for you? ๐Ÿ˜ฌ
Dangerโ€ฆ ๐Ÿšจ Models trust tools over their own โ€œknowledgeโ€ even for simple and well trained cases.

10.11.2024 18:34 โ€” ๐Ÿ‘ 19    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0
TRI Women and Allies breakfast

TRI Women and Allies breakfast

A robot standoff. One with wheels and one with legs.

A robot standoff. One with wheels and one with legs.

Su debugging her robot teleop system

Su debugging her robot teleop system

Vidhi presenting her work on robot audio

Vidhi presenting her work on robot audio

Hi from CoRL ๐Ÿ‘‹

08.11.2024 10:27 โ€” ๐Ÿ‘ 15    ๐Ÿ” 2    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

@ybisk.me is following 20 prominent accounts