Share this page!

Enter URL or ID to Unroll

You can paste full URL like: https://x.com/threadreaderapp/status/1644127596119195649
or just the ID like: 1644127596119195649

How to get URL link on X (Twitter) App

  1. On the Twitter thread, click on or icon on the bottom
  2. Click again on or Share Via icon
  3. Click on Copy Link to Tweet
  4. Paste it above and click "Unroll Thread"!
  5. More info at Twitter Help
Tomer Ullman Profile picture
Tomer Ullman
@TomerUllman
Assistant Professor, Department of Psychology, Harvard University. Computation, cognition, development.
Dec 7, 2022 • 7 tweets • 3 min read
Trying out some intuitive psychology / theory-of-mind with openAI's chatGPT.

Starting with a classic Sally-Ann task.

For each one of the vignettes, I'm showing chatGPT's answers from 10 tries.

Again, one might initially think: Wow! This thing gets basic theory-of-mind! ImageImage But maybe this ToM vignette is so used chatGPT has simply seen it?

Let's try less frequent terms. Other LLMs I've tried have failed this variation in the past.

Still works! ImageImage
Oct 18, 2022 • 11 tweets • 4 min read
החדשות בישראל מוצפות בסקרים לקראת הבחירות.

הסקרים קפואים ודומים אחד לשני. על פניו, סימן שהם אמינים, והמציאות פשוט קפואה.

אבל האם הסקרים קפואים ומסכימים *מדי*?

בדקנו (אני ו - @alonyakter), ונראה שהתשובה היא: כן, הם דומים מדי.

הסבר בשרשור. Image לכולנו ברור שאם סוקרים 800 איש באקראי, ואז 800 איש אחרים, יכול להיות הבדל בין שני הסקרים.

ההבדל הזה *צפוי*, והוא *אמור* להיות בגודל מסוים.
Aug 2, 2022 • 9 tweets • 3 min read
Do models like DALL-E 2 get basic relations (in/on/etc)?

Colin (Coco) Conwell and I set out to investigate. The result is now on arXiv:

“Testing Relational Understanding in Text-Guided Image Generation”

arxiv.org/abs/2208.00005 DALL-E 2 and the like are mighty impressive, but people have also noted their limitations with common sense, negation, composition, and more.

We decided to focus specifically on relations.
Jul 17, 2020 • 11 tweets • 4 min read
Stock photos of busy parents on the computer, rated for realism:

1: Mom has both hands on computer, child seems to need no supervision at all. Work is getting done. Mom is even smiling! This sucks, 2 out of 10. Computer is open but not being looked at, only one hand available. Triple tasking: baby, phone, writing. Baby seemed horrified by mom's grant pitch. Pretty good, 8/10.
May 11, 2020 • 4 tweets • 3 min read
Facebook AI released Blender, "largest-ever open-domain chatbot".

I tried some things.
====

FACEBOOK AI: "truly intelligent, human-level AI...must effortlessly understand the broader context of the conversation and how specific topics relate to each other."

FACEBOOK BOT: FACEBOOK AI: "This is the first time a chatbot has learned to blend several conversational skills — including the ability to assume a persona, discuss nearly any topic, and show empathy — in natural, 14-turn conversation flows."

FACEBOOK BOT: