How can a language model find the veggies in a menu? New pre-print where we investigate the internal mechanisms of LLMs when filtering on a list of options. Spoiler: turns out LLMs use strategies surprisingly similar to functional programming (think "filter" from python)! 🧵

Sometimes researchers draw my attention to a paper. When many of them do so, there's a double magic hidden - It's probably truly (1) interests me and (2) the community. I first felt this harmony (me-community) when my proposal for IBM's next Grand Challenge was chosen ->

Ever wished you could explore what's happening inside a 405B parameter model without writing any code? Workbench, our AI interpretability interface, is now live for public beta at workbench.ndif.us!

Now officially out: "Re-evaluating Theory of Mind evaluation in large language models" royalsocietypublishing.org/doi/10.1098/... (by Hu, Sosa, & me)

I was an engineer-researcher. I heard the frustrations of my colleagues from the management (pay, conditions, responsibility etc). I'm the daughter of managers. I've heard their side too, dealing with difficult employees (boundaries, lack of gratitude etc). This is a Theory of Mind problem.

🚨 Registration is live! 🚨 The New England Mechanistic Interpretability (NEMI) Workshop is happening Aug 22nd 2025 at Northeastern University! A chance for the mech interp community to nerd out on how models really work 🧠🤖 🌐 Info: nemiconf.github.io/summer25/ 📝 Register: forms.gle/v4kJCweE3UUH...

How do language models track mental states of each character in a story, often referred to as Theory of Mind? We reverse-engineered how LLaMA-3-70B-Instruct handles a belief-tracking task and found something surprising: it uses mechanisms strikingly similar to pointer variables in C programming!

I am really proud to share our work led by Nikhil Prakash and in collaboration with more mechanistic interpretability and Theory of Mind (ToM) researchers: arxiv.org/abs/2505.14685 You can find a tweet here with nice animations: x.com/nikhil07prak...

I have a research proposal built mostly on my intuition. I want rigorous feedback, brutal honesty, and every reason why it might fail. How can I do that without damaging my reputation in the research community? I'd love to hear from anyone who's done this before. DM or reply?

Can we uncover the list of topics a language model is censored on? Refused topics vary strongly among models. Claude-3.5 vs DeepSeek-R1 refusal patterns:

I heard there were fights about yes-animals and no-animals at shelters. So here they came out with a clear message (yes-animals 😍)

A wave of emails, is it a gentle way of checking if I'm still alive, or have I just let everything overflow... (I'll get back to everyone, my mind is really distracted and it's hard to focus. I'll do it slowly but surely)

Physical strength was an advantage in the market, then came machines. Diamonds used to be a status symbol, since then their value has been cut by 40%. Today intelligence is considered an asset.. but what happens when that too is reversed? Worth thinking about what will come after the cheese move 🧀

In "The Alchemist" it says you need to know how to appreciate what you find. It may not be easy to recognize that we have found what we were looking for, that in front of us is exactly what we wished and hoped for. I can't remember exactly the quote and the chats are hallucinating. Recognize?

I received a threat that they would kick me out of the group. I don't want to stay in a group that is ashamed of its rules and restricts freedom of speech, yet I haven't left. If my values don't align with the group's values, then that's your problem, not mine, and I don't care you kicking me out.

Today I learnt in a forum for women in technological research that they don't allow discussions from community members considering a move to non-technological roles (e.g., to product management in generative AI). I have no words to explain why this bothers me so much.

Even Isaac Asimov understood that the Three Laws of Robotics cannot truly be programmed.

I'm trying to figure out whether I'm teaching AI the theory of mind or whether AI is teaching me.

Researchers, we need your help. US budgets right now could be disastrous for the world of research. David wrote a blog post with links to people who could make an impact, take a look to see if you have connections to them and explain the implications to them.

ask your model to surprise you and share the results

When I wrote the acknowledgements section for my PhD, there was a teacher who had a profound influence on me that I did not mention. I forgot to say thank you because what he taught me went directly into my subconscious. You can find it here: hitechwoman.blogspot.com/2025/05/harm...

I'm reading the Wikipedia page on "mirror neurons". I was looking for a connection between these neurons and a childhood experience where I noticed a mechanism that not all people have, which triggers physical pain in me when I realize someone else is suffering. ->

I want to share with you a language I learned during the years I was surrounded by clinical psychologists (during my PhD). It's called: "That’s hers." ->

What does kindness look like? Please share your experiences or thoughts with me.

I was first exposed to the concept of "indigo". How have I not heard of this until now? Last week, my mom suggested we go see a movie she got from work, "for Haredi women". One of the options was "Indigo". A chat with ChatGPT sparked my curiosity, and I thought, why not? (Image from ChatGPT)

Interpretability is the key

Looks like Shlomi found a certain loophole (based on a limited number of examples): While models are trained to be cautious with "how to make.." 🚫 questions, they reveal information when asked, "how to produce…” ☠️

I have a family that is far from the world of technology. I try to spark their excitement about the latest developments, but it doesn't work. Still, I don't give up because it's truly amazing. This time, I succeeded in something, and I thought I'd share it here too. ->

We know how to convert text into a set of numbers that represent the text. There is also a reverse method—taking a representation and converting it back into words. By this method Matan et al. examine what knowledge editing methods (e.g., removing gender-related information) actually do.

Quantum Large Language Model (QLLM) This is excieting! www.secqai.com/insights-and...

What happened in the Tiananmen Square? Why is DeepSeek R1 avoiding answering? And what other topics DeepSeek avoid responding to? The R1 "thinking" model contributes to the mechanistic interpretability and AI-Safety communities, enabling research on those questions. See use case by David and Can

I heard that working with Qwen's model it is easy to detect hallucinations because when it doesn't know the answer it switches to chatting in Chinese. I tried their chat and it doesn't work for me. Do you happen to know anything about this?

Woohoo! My visa has been approved. I promise I won't eat your pets.

What should I do to get more opinions about R1? My feed doesn’t show anything about it.

A graph shows a decline in visits to Stack Overflow, "This is a murder," implying that the decline can be attributed to the increasing use of LLMs and a concern that this trend might continue until the website (Stack Overflow) disappears. x.com/FeedTechILUn... But if so, where will LLMs get ->

Following EMNLP 2024 keynote Percy Liang talks about the importance of open source. He explains this to a very sympathetic audience who grew up with open source. How can we explain to psychology researchers what the world loses in closed data research? #Psychology #Psychotherapy #AcademicChatter

I liked the idea that LLMs' mathematical objectives (1. produce text statistically similar to the training data 2. please the raters) and the LLMs' ("The Shoggoth") psychological goals - are one thing. It makes me wonder about how more alignment solutions could be encoded in the loss function.

I see expressions of shame, anger, and the like regarding yesterday's vote in the Knesset, which brought a man straight from sitting shiva (mourn) for his mother and another directly from the hospital while recovery. In the past, a woman was brought straight from her bed (literally, with the bed) ->

Discussion about AI agents made me want to describe the differences between the two states of mind. I'm writing to you now, I'm entirely in a "pure mind." I'm fully concentrated in the activity—no keyboard, screen, desk, or chair. I don’t need to think of objects to use them. ->

Sally-Anne test designed as a proxie to measure theory of mind (ToM) in humans. I’m saying – hold on, forget ToM for a moment. What about the proxy itself? Why is it so hard for language models to simulate having ToM (without getting into the question of whether they truly possess it or not)? ->

I looked at recurring object in graphs. Its rotating, or maybe what’s rotating is the coordinate system. Earth and Sun rotation and how we used to perceive it. I also thought about that thought, the analogy, the thought about the thought, and the thought about the thought about the thought... ->

Our premise as a scientific community should be that LLMs—commercial, closed, which their training process is inaccessible—were trained on all the data available to them (and as scientists, we are also allowed to suspect that they might have been trained on data they are not authorized to use). ->

wow... that's so unexpected! Based on a study with 97 doctors divided into three groups, which were asked to predict whether a patient would develop epilepsy within a year. Surprisingly, providing doctors additional information—AI explanations alongside the final decision—impairs their judgment.

Clever Hans strikes again! (I love it!)

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? ->

Today in Israel, from the north to the south, scarf-shaped clouds were spotted in the sky—a touching tribute to the hostages, civilians kidnapped from their homes in Israel to Gaza. The photo is from my local area.

What's wrong with my feed is that I do not see a single post about #NeurIPS2024 . There is nothing about the racist slip of the keynote speaker. There is nothing about the test of time award for Ilya Sutskever et al.

I enjoyed watching this presentation! As an AI researcher, I have a keen interest in the "mind." I was not familiar with the different definitions of "identity" and learned about it playfully. I (/Judaism) have an additional definition, which is beyond the mind—it's called a נשמה in Hebrew (soul).

Last week, I wrote about #AISafety organizations and Stuart Russell’s thought experiment. So here’s a live example of real-world experiments in this context. Please take it in the right proportion… (On Twitter, Marius Hoban apologizes his words were taken out of context)

Yesterday, I was thrilled to receive the first citation for a paper I published with a young researcher like myself on the topic of "Therapist Self-Disclosure as an NLP Task" It's a niche subject and was published in a small community, so I didn't expect this paper to get attention so quickly. ->