Sign in

Neil Traft

@ntraft.bsky.social
442 followers 260 following 291 posts

PhD student at the Vermont Complex Systems Institute. Interested in ML, evolution, self-organization, & collective intelligence. ntraft.com t.co/ja3fdtRLdM

PostsRepliesMedia
Neil Traft @ntraft.bsky.social · 29/12/2025
Astonishing: the default behavior of Chrome's "Listen to this page" feature is NOT to simply read you the page, LIKE YOU WOULD EXPECT, but instead to make one of those goofy NotebookLM AI podcasts... and I had to search the Help docs to figure out how to change it. Thanks, Google.
000
Neil Traft @ntraft.bsky.social · 16/12/2025
Random observation: did you know this?! I knew about how Einstein's general relativity is required to make GPS navigation work, but I'm not as familiar with how quantum mechanics factors into semiconductor technology. Remarkable.
220
Neil Traft @ntraft.bsky.social · 06/12/2025
I thought this was a fascinating way to motivate model merging: to be able to *interpolate* between model classes, rather than be forced into discrete choices. (models trained on different datasets, models of different sizes, etc.) "Interpolation and alignment are two sides of the same coin."
042
Neil Traft @ntraft.bsky.social · 06/12/2025
Did you know that you could create hybrids across different neural net architectures—*without* updating the weights?? Come check out my poster at the UniReps @unireps.bsky.social workshop at 3:45pm today at #NeurIPS! (Or you can already browse the posters on the wall in 20D throughout the day.) 😃
041
Neil Traft @ntraft.bsky.social · 01/12/2025
Link to the full paper: openreview.net/pdf?id=e4wKQ... I think there's a lot of very interesting potential applications of "model stitching". Have a look. Plus, you get to feel like a mad scientist. IT'S ALIVE!
010
Neil Traft @ntraft.bsky.social · 01/12/2025
I’ll have a poster at @unireps.bsky.social 🔵🔴 It’s a terrific workshop at the intersection of neuroscience and DL: across both biological and artificial NNs, how can we measure / compare / align / merge neural representations? Should be fascinating! Check it out if you're around! unireps.org/2025
List of topics for the UniReps workshop.
101
Neil Traft @ntraft.bsky.social · 01/12/2025
So stoked to be going to my first #NeurIPS!! It’s crazy that in 10+ years of robotics and AI, I've never been to the great Lollapalooza of Machine Learning. 🥳🎆 I’m presenting my Frankensteinian efforts to stitch together parts of different neural networks! 🧪
Abstract for the paper, "Bridging Large Gaps in Neural Network Representations with Model Stitching".
120
Neil Traft @ntraft.bsky.social · 17/10/2025
Translation: You have no new messages.
LinkedIn email: "You have 1 new message"
010
Neil Traft @ntraft.bsky.social · 28/09/2025
GPT now has ads. 💀 They are voluntary... for now.
120
Neil Traft @ntraft.bsky.social · 12/09/2025
Every single part of this website is solid gold. 😂 Thank you!
Every second you don't subscribe, another potential future human life is lost.

Stop being a mass murderer: <subscribe entry field>
060
Neil Traft @ntraft.bsky.social · 03/08/2025
In reality, all environments are only partially observable, so this is the only regime in which RL should be evaluated.
240
Neil Traft @ntraft.bsky.social · 29/07/2025
I just realized that all these hyped-up “AI scientist” concepts all concentrate on the idea of *discovery*—new algorithms, new materials, new products—without exception. Not a one of them focus on *understanding*—arguably the larger role of a scientist! 🧐
140
Neil Traft @ntraft.bsky.social · 25/07/2025
Been studying Mixture of Experts recently. It's baffling to learn that typically different tokens are routed to *different* experts. How can you be an "expert" at a particular word?? I suspect this name lends us very poor intuition for what these models actually do! (shows 2 tokens and 4 "experts")
021
Neil Traft @ntraft.bsky.social · 28/02/2025
Don't know which pub you're referencing (having a hard time finding on my phone), but he also discusses this in his "Studying Artificial Life With Cellular Automata". I.e., molecules are parallel so alife should be too. (Later mentions why traditional computers are not so parallel, even w many CPU.)
120
Neil Traft @ntraft.bsky.social · 15/02/2025
That process for making a decision? THAT'S the free will! That's just you. Deciding.
If	the	neural	 computations	are	causally	sensitive	to	semantic	content,	rather	than	
detailed	syntax,	and	those	semantics	relate	to	person-level	 concepts,	and	all	that	
information	is	integrated	in	a	hugely	contextually	interdependent	way,	and	is	
used	to	direct	behavior	over	nested	timescales,	in	ways 	that	cannot	be	either	
algorithmically	 or	physically	pre-specified,	based	on	criteria	configured	into	the	
circuits	 derived	from	learning,	which	embody	reasons	of 	the	person and	not	any	
of	their parts,	then	I	would	say	that	*just	is the	person deciding	what	to	do*.
273
Neil Traft @ntraft.bsky.social · 02/01/2025
Yeah I think I got a pretty good one... #bsky_builds_bots
041
Neil Traft @ntraft.bsky.social · 28/12/2024
Post your cookies. (Better late than never.)
Almond crescents, raspberry jam thumbprints, candied ginger molasses cookies, and matcha latte cookies.
020
Neil Traft @ntraft.bsky.social · 16/12/2024
Sutton's Bitter Lesson is that simple algorithms win *IF* they can scale. But the whole point is we don't all agree on what will scale.
000
Neil Traft @ntraft.bsky.social · 11/12/2024
WHY SO EFFECTIVE? MORE IN PAPER!
010
Neil Traft @ntraft.bsky.social · 03/12/2024
A new foundational infrastructure is needed to achieve the Internet as the participatory, self-organized collective we initially dreamed it would be... rather than the centralized, uncustomizable, ad-choked Internet we've ended up with.
Screenshot of part of the linked thread.
034
Neil Traft @ntraft.bsky.social · 19/11/2024
Look, I do machine learning, but I'm not happy about it.
110
Neil Traft @ntraft.bsky.social · 25/09/2024
The technology we could have vs. the technology we got.
000