🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

More specifically, we identify three key assumptions that permeate the lit on LLM cultural alignment: stability, extrapolability, and steerability. All are false. Basically, LLMs just say random s*** all the time and can easily be manipulated into expressing all kinds of things.

Steerability: When prompted to align with certain groups of people, LLMs show fundamental failures to express humanlike preferences at all, let alone embody specific cultural perspectives, even under prompt optimization.

Our findings about the unreliability of assessing LLM preferences raise concerns that strategic experimental design and cherry-picking could be used to paint an arbitrary picture of an LLM's cultural preferences...

…For example, we show that simply choosing an even- versus odd-sized Likert scale can greatly change the outcome of a recent viral experiment from Mazeika et al. (2025).

Very hard

Today, NSF announced an add’l 500 NSF Graduate Research Fellowship Program awardees for the 2025-2026 cohort, bringing the total to approx 1,500. #NSFGRFP supports grad students as they pursue their dreams, build STEM skills, & become the next generation of innovators & leaders.

Is 10 bits/s our speed limit? "It appears likely that most of the neurons in the CNS contribute to sensorimotor processing for feedback control." Bingo. If you take a more encompassing view of 'information' to include sensorimotor, the 'paradox' goes away

NEW: George Alvarez (George Alvarez) of Harvard Psychology and the #KempnerInstitute shows how long-range feedback projections can enhance the alignment between #ML models and human visual processing. Watch the video: youtu.be/Ju_eD0Jwa8Q #NeuroAI20205 #AI #neuroscience #NeuroAI

Today seems like a real good day to relax and revisit the important issues in life. Here's a timeline cleanser to get us started:

Nice to see some good (less bad) news. NSF fellowships are critical support for our brilliant students.

Nature Neuroscience The brain works at more than 10 bits per second nature.com/articles/s4159…

Excited to share our recent work on how we can flexible combine visual generative models, VLMs and simulators for visual synthesis! This enables physics engine controlled video generation, graphics engine controlled image generation, and compositional image synthesis!

it’s actually sorta funny that the cogsci people who think the brain isn’t a computer are the ones that seem to think genAI is onto something

That's called Brownian motion and we use it in Langevin dynamics (among other things).

Interestingly several people have asked me about this paper on neural coding. As soon as the preprint was posted, a student asked me about it. I think it's a thoughtful and clearly written paper, though most of the ideas presented are obvious to me. It was fascinating to read

Day 1 to Day 3 at the Nordic Probabilistic AI School 2025, lectures about probabilistic modeling, causal probabilistic programming, variational inference and conformal prediction so far :)

I am once again pitching my romantic comedy: - two academics start dating - discover they are each other's terrible reviewer - hijinks ensue Working title: Love is Double-Blind

I want to address one very common misconception about optimization. I often hear that (approximately) preconditioning with the Hessian diagonal is always a good thing. It's not. In fact, finding a good preconditioner is an open problem, which I think deserves more attention. 1/4

Wonderful to chat with Paul Middlebrooks on Brain Inspired, where we discuss many things. Among them: What are we brain researchers trying to do here (wrt causality)? I also stick my neck out to predict the next big breakthrough in emotion research. Thanks for having me!