5/5
@ema-ridopoco.bsky.social and @andreadittadi.bsky.social will be presenting a poster in San Diego and Luigi and I will be presenting at Eurips (eurips.cc) in Copenhagen so come on by! 😄
Really appreciated Joshua Gans' postmortem on an experiment in vibe researching
joshuagans.substack.com/p/reflection...
Taste is still paramount and the models are instruction-tuned to sycophancy to all hell
🏹 Job alert: Postdoc in Human-Computer Interaction with Explainable AI at University of Copenhagen
📍 Copenhagen 🇩🇰
📅 Apply by Feb 1st
🔗 https://employment.ku.dk/faculty/?show=153139
The promise of AI chat assistants: they solve 90% of the problems users have (by looking up the docs and telling them)
My reality: need to spend 10 minutes trying to get to a human, to solve an issue I need customer support to look into
Around minute 8 I sign up to a competitor
Andrej Karpathy’s take on AI coding agents feels grounded. The industry’s chasing full autonomy when models still hallucinate too much.
Agents that churn out a thousand lines of code leave you either blindly trusting them or slogging through reviews. These tools should embrace their fallibility.
3/5
The two models agree on their prediction for the highest likelihood label. They also disagree on the ranking by likelihood of the remaining labels, and while this has a negligible effect on the KL divergence, it means the relation between their representations is non-linear.
1/5
We study when and why representations learned by different neural networks are similar from the perspective of identifiability theory, which suggests that a measure of representational similarity should be invariant to transformations that leave the model distribution unchanged.
Beatrix M. G. Nielsen
2/5
We prove that a small KL divergence between models is not enough to guarantee similar representations. Here is an example of how to construct two models with small KL divergence, but representations which are far from being linear transformations of each other.
I am happy to announce that our article "When Does Closeness in Distribution Imply Representational Similarity? An Identifiability Perspective" has been accepted at NeurIPS 2025! 🎉 arxiv.org/abs/2506.037...
Details below 👇
4/5
We also show that it is possible to define a metric between probability distributions and a measure of representational dissimilarity such that when distributions are close in this sense, we get similar representations.
