Dimitra Maoutsa retweeted
In the case of a narcissist/psychopath:
1. Sharing with them something you are vulnerable about is like handing them a weapon.
2. Discussing something you are passionate about is a source for them of potential sadistic pleasure in terms of sabotaging your happiness around this passion.
3. Telling them about something you’ve done wrong is a source of leverage.
4. Discussing something in them that concerns you is a greenlight for them to start undermining you to others without your knowledge.
Dimitra Maoutsa retweeted
*IF* neurons are linear+activation function, then there is elegant theory by @CPehlevan @chklovskii but also @jcrwhittington @SuryaGanguli et al that shows that the non-negative activation function, together with an energy constraint (e.g., weight decay in NN training), can lead neurons to align with individual concepts (also in NMF). This is the more formal version of what MechInterp conceptualized as privileged axes coming from pointwise nonlinear activation functions.
arxiv.org/abs/1706.00382
arxiv.org/abs/2210.01768
Dimitra Maoutsa retweeted
Random idea: What if AI conferences added an *ORAL rebuttal round* after the written rebuttal on open review?
Authors, reviewers & ACs could join a short audio call (still anonymous if possible) to discuss the paper and clear up misunderstandings.
I feel like this could help ACs better judge if reviewers really read the paper, and also whether authors actually understand what they wrote (esp in the age of AI-generated papers).
Also authors would need to get better at explaining and defending their work. Reviewers would need to better justify their concerns. This could raise the bar for both sides. Ofc scalability and language barriers could be an issue that needs to be considered. Maybe only do this when the AC thinks it's necessary (esp when reviewers disagree with each other)?
Just a random thought.
Dimitra Maoutsa retweeted
Interpretability blogs are the recipes blogs of ML. Not sure, why I have to read through the story of your grandmother's first house to find out you used gradients.
Dimitra Maoutsa retweeted
75% of respondents think @iclr_conf should lower the acceptance rate to about 10%. This could not only help with conference logistics and experience, it might also have a lasting effect on the number and quality of submitted work in later years.
Given the rise in submissions, should @iclr_conf substantially lower the acceptance rate, to around 10%?
75%Yes
25%No
101 votes • Final results Dimitra Maoutsa retweeted
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
Dimitra Maoutsa retweeted
A new study finds evidence of a geometric “mental line” of neural activity in the premotor cortex of monkeys performing tasks that require transitive inference deductive reasoning. @istsupsan scim.ag/3UmuCRw
Dimitra Maoutsa retweeted
Stimulating ~20 motor cortex neurons selected by their activity induces learned population dynamics and, depending on brain state, the movement itself. This demonstrates that specific ensemble activation is sufficient to drive learned behaviors.
nature.com/articles/s41586-0…
Dimitra Maoutsa retweeted
This is also consistent with the evidence of symbolic structure emerging in visual neural networks, across both self-supervised and contrastive models (eg dino) studied in our work! We also characterize what ideal compositional structure would look like, and quantify to what extent it actually emerges in learned representations
🤖🧠NEW PAPER🧠🤖
What is the relationship between neural networks & symbolic systems?
Based on evidence from mechanistic interpretability, we argue for a quasi-symbolic view of cognition: Neural nets implement symbols but only approximately
Link: sciencedirect.com/science/ar…
1/n
ALT Top: The title, authors, and journal of the paper. Title: "Reconciling symbols and neural networks: Evidence for a quasi-symbolic view." Authors: "Tom McCoy, Abi Tenenbaum, Zhang Enyan, Zhenghao Herbert Zhou." Journal: "Current Opinion in Behavioral Sciences" Bottom: A figure displaying how the vector representations learned by neural networks often show emergent symbolic structure. For instance, they can often be decomposed into meaningful constituents (illustrated in the figure by adding constituents to make a vector) or classified in vector space along symbolic dimensions (shown by an image of linear classifiers). However, these vectors are only quasi-symbolic: they deviate from pure symbolic structure in important ways - e.g., they can display graded similarity (“movie” is closer to “novel” than to “meal”), and they may capture graded context effects such as the word “good” affecting different nouns differently (the image shows such graded similarity among points in space
Dimitra Maoutsa retweeted
As someone fairly new to the #ASD field, I've followed the Portera Lab's work closely. Interesting stuff! cell.com/neuron/fulltext/S08…
The @dgsomucla community packed the room today to hear Dr. Carlos Portera-Cailliau from @UCLANeurology describe his career-long effort to understand FMR1 & how alterations in this gene lead to fragile X syndrome. It was standing room only & included a lively post-talk discussion about the future of #ASD therapeutics. Thanks to @UCLA_BRI for organizing this really special event.
Dimitra Maoutsa retweeted
Women are rarely quoted as experts on AI. From July 2025 to June 2026, among the well over 100,000 experts cited in domestic news coverage of AI, men appeared 4.3x more often than women.
motherjones.com/politics/202…
Dimitra Maoutsa retweeted
A new nasal spray reversed signs of brain aging and restored memory performance in preclinical trials, and it worked in just weeks.
Scientists at Texas A&M University have unveiled an experimental nasal spray that appears to reverse key markers of brain aging. The therapy targets neuroinflammaging, the chronic, low-level inflammation deep in the brain’s memory center that slowly degrades cognitive function and increases the risk for conditions like Alzheimer’s disease. By delivering microscopic biological packages known as extracellular vesicles directly into the nasal passages, the treatment bypasses the blood-brain barrier without requiring surgical or invasive procedures .
In preclinical trials, just two doses spaced two weeks apart produced results within weeks, calming hyperactive immune cells, restoring neuronal cellular energy, and boosting memory performance for several months. While human clinical trials are still needed to confirm these findings in patients, the observations indicate that age-related mental decline may not be an irreversible fate. Researchers hope this accessible approach could one day offer a simple, non-invasive alternative to complex neurological treatments and prolonged medications.
[Madhu, L. N., Kodali, M., Rao, S., et al. (2026). Intranasal human NSC-derived EVs therapy can restrain inflammatory microglial transcriptome, and NLRP3 and cGAS-STING signalling, in aged hippocampus. Journal of Extracellular Vesicles, 15(2). DOI: 10.1002/jev2.70232]
Dimitra Maoutsa retweeted
The ideal outcome for science isn't a couple of labs using proprietary models and millions of dollars to crack trophy problems, it's to have millions of AI-aided human scientists who follow their curiosity and become way more ambitious
Dimitra Maoutsa retweeted
I’ve never been able to draw, until I learned to turn math ideas into code and explore visualizing things. More than 20 years ago I started with @WolframResearch Mathematica, now ~1000 #Python scripts: from complex analysis to statistical mechanics to fractals.
#builtWithMarimo
Dimitra Maoutsa retweeted
Mammalian brain networks prioritize robustness and capability over metabolic efficiency, accepting higher energy costs to maintain function. This finding shifts the focus from simple resource conservation to the active trade-offs that shape neural architecture.
biorxiv.org/content/10.64898…
Dimitra Maoutsa retweeted
(2025) Spatiotemporal properties of cortical excitatory and inhibitory neuron activation by sustained and bursting electrical microstimulation cell.com/iscience/fulltext/S… "Electrodes in visual cortex disrupted circuits, but #ICMS still evoked activity"; #BCI #NeuroTech
ALT Graphical abstract
Dimitra Maoutsa retweeted
"Geometric Concepts in Partial Differential Equations"
- Felix Otto
youtu.be/K8-O-FdUzGs?si=A1Ms…
To gradually work my way through some of the ones which I've enjoyed:
"Optimization in Theory and Practice"
- Stephen Wright
youtu.be/Ep1TzZDOHnU?si=hlbR…
