Quite a claim in this new work from Goodfire. Among many points, they argue that some of the main examples of superposition and polysemanticity in large models were misunderstood. This definitely deserves some deep, rigorous peer review. already-clean.goodfire.com
Who's Who of AI
Shahab Bakhtiari
Why they matter
Researcher with public evidence across AI research.
- AI signals
- 7 past 30d
- Sources
- 6 distinct domains
- Discussions
- 8 past 30d
- Latest signal
- 6d ago
|| assistant prof at University of Montreal || leading the systems neuroscience and AI lab (SNAIL: https://www.snailab.ca/) 🐌 || associate academic member of Mila (Quebec AI Institute) || #NeuroAI || vision and learning in brains and machines
What they're sharing
Neural Networks Are Already Clean already-clean.goodfire.com
Temporal Difference Learning for Diffusion Models arxiv.org
Borges and AI arxiv.org
Neurodata Without Boredom: Benchmarking Agentic AI for Data Reuse arxiv.org
View on Bluesky →
Partial success in closing the gap between human and machine vision proceedings.neurips.cc
seq-JEPA: Autoregressive Predictive Learning of Invariant-Equivariant World Models proceedings.neurips.cc
Building pro-worker AI | Brookings brookings.edu
Expert consensus report: Ways generative AI can support and threaten learning in K-20 U.S. education cehd.umn.edu
View on Bluesky →
What AI can learn from hunger thetransmitter.org
View on Bluesky →
Articles & links
Shahab Bakhtiari reposted
Temporal Difference Learning for Diffusion Models (ICML 2026) arxiv.org/abs/2606.15048 By Yangchen Pan (my former PhD student) and co-authors. It reformulates diffusion training as a Markov reward process and introduces a TD objective to encourage temporal consistency across d…
AI Weekly's analysis
→
- The paper introduces a temporal difference objective that penalises inconsistency across the full denoising trajectory rather than only at adjacent time steps.
- It reframes diffusion as a Markov reward process and denoising as a policy evaluation problem, unifying discrete-time and continuous-time formulations.
- Reported FID gains are strongest when the number of sampling steps is small, the regime where few-step samplers and low-compute serving live.
Read full analysis →
View on Bluesky →
Shahab Bakhtiari reposted
Are you aware of the paper Borges and AI by Bottou and Schölkopf? arxiv.org/abs/2310.01425
AI Weekly's analysis
→
- The paper proposes understanding LLMs through Jorge Luis Borges rather than the science-fiction imagery used by both AI proponents and opponents.
- Authors Léon Bottou and Bernhard Schölkopf argue the sentient-machine and paperclip-apocalypse pictures may not describe the phenomenon well.
- Submitted to arXiv in late September 2023, the essay is a framing argument rather than an empirical result.
Read full analysis →
View on Bluesky →
Shahab Bakhtiari reposted
New preprint with @lingqiz.bsky.social: Neurodata Without Boredom: Benchmarking Agentic AI for Data Reuse arxiv.org/abs/2605.12808 1/10
I think the evidence is growing and gaps are closing; this is from a few years ago: proceedings.neurips.cc/paper_files/..., and things've been improving since then. Of course, there are skeptics ...
I’m thinking even beyond image labeling—more in the realm of recent self-supervised learning methods (e.g., what we proposed here: proceedings.neurips.cc/paper_files/...). There’s really cool work bringing these concepts into visual neuroscience (we and others have been doing …
Acemoglu et al.’s case for "pro-worker AI" makes so much sense. The common counter to #AI job loss is that tech always creates new work. Acemoglu argues that new jobs don't appear spontaneously, especially if we don't actively build AI to complement workers. www.brookings.edu/…
Shahab Bakhtiari reposted
What happens when you ask 30 experts in the learning sciences about the effects of Generative AI on learning? You get this report, which has insights and recommendations based on expert thinking and experience, not hype. The full report is worth a read, but here are a few high…
Shahab Bakhtiari reposted
One path toward AI safety, then, is to take a page from evolution and find a way to install caring so that it retains leverage over knowing. #neuroskyence #neuroAI By @tonyzador.bsky.social www.thetransmitter.org/neuroai/what...
Their own posts
Recent commentary
I’ve been thinking… now that AI is taking on groundbreaking, award-winning problems across science, what are those equivalent problems in #neuroscience? What would one of these models actually have to do to leave us all in awe? 🧠🤖
Astra’s most striking leap across several fronts (eg robotic control) boils down to spatial intelligence. I used to think genuine spatial understanding was impossible for language models without physical real-world interaction. Well … Astra proved that wrong or it's far more than just a mere LLM
Their network
In Shahab Bakhtiari's orbit
Center = Shahab Bakhtiari. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.
Are you Shahab Bakhtiari? Show it.
Add the Who’s Who of AI badge to your site or bio. It links back to this profile.
Markdown: [](https://aiweekly.co/whos-who/person/shahabbakht-bsky-social)