←
Neil Traft9,153 posts
Neil Traft avatar

Neil Traft

@ntraft

PhD student at @uvmcomplexity. Interested in ML, evolution, & self-organization. ntraft.com/wp-content/upload…

◷ Joined September 2008▧ 320 Photos and videos
Neil Traft avatar
Neil Traft@ntraft·1 Dec 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! 🧪

Post attachment
Ida Momennejad avatar
⟳ Neil Traft retweeted
Ida Momennejad@criticalneuro·Nov 10

Un-bittering the lesson. We always needed science. Scale was never enough.

Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO

Daniel Yamins avatar
⟳ Neil Traft retweeted
Daniel Yamins@dyamins·Nov 13

For years people in NeuroAI have been wondering what the relationship between CNNs and transformers is as brain models, but the ground has been kind of muddy. Now I think we are starting to see underlying situation more clearly. Transformers > CNNs -- in interesting ways!

In neuroscience, a canonical microcircuit is thought to subserve all cortical computation. In AI, neural networks universally subserve intelligent behavior. Is there convergence between computational motifs in natural and artificial intelligence? Work with @dyamins 🧠🤖

Post attachment
Neil Traft avatar
Neil Traft@ntraft·Nov 4

That weight-editing experiment that's gone viral? It's not actually editing the weights—it just wrote a completely separate controller. You can randomize the other weights or set them to 0, and it still works. (According to the claim below.)

We asked Claude Opus 5.5 and GPT-6 Astra to fix a small trained driving network so the car reaches the goal. Their working fixes didn't need what the network learned.

1:05
Samip avatar
⟳ Neil Traft retweeted
Samip@industriaalist·Oct 22

I think the last month might have been the biggest month in (zeroth-order) optimization history. In the 10 days since we released the Noise speedrun, the time to pretrain a transformer went from 1.76 hours to 1.6 seconds, a 4,000x (!!!) speedup. Dust itself was up to 10,000x more compute-efficient than the previous state-of-the-art. Not only can we pretrain a transformer without backprop, it's much more efficient than anyone would've expected. Search fking works! Now, scaling is the biggest open question. Some of the tricks that brought the time down to 1.6 seconds will simply not scale. To discard those, we'll be scaling the dataset every time it gets overfit, starting with 30M tokens instead of 10M. Huge thanks to @jayden_teoh_, @spicey_lemonade, @RobertTLange, @EmanueleRodola, @ChunyuanDeng, @LuochengHuman, @Alaxarr, @pavleculafic, @IsaacYong13, Euraba and areu01or00 (GitHub), for the contributions. And @cs_serdar for gracefully maintaining the speedrun and @bishmdl76 for helping out. Expect a lot more from us over the coming weeks.

Neil Traft avatar
Neil Traft@ntraft·Oct 22

My advisor just asked me, "Why is AI good at summarizing some documents, but so bad at scientific papers?" I posed that it's just *condensing* the text, when I actually need it to *translate* the ideas for me. And lo and behold, a paper demonstrating the info density of AI text!

If you've noticed a distinct dialect emerging in frontier AI models and struggled to understand what they're talking about, you're not crazy! We turned to psycholinguistics to investigate what makes the writing style so weird, and why this might be happening 🚨New preprint! 🧵

Post attachment
Neil Traft avatar
Neil Traft@ntraft·Oct 22

This textbook looks very interesting, and I skipped to the end and found a pretty compelling list of research directions. Some slightly hot takes here but a solid outlook, IMO.

Post attachmentPost attachmentPost attachmentPost attachment

The lectures of my new course "Principles of Deep Representation Learning" are now available at the Youtube site: youtube.com/playlist?list=PL… with new lectures continuing to be added there. I will open-source all the slides later in the textbook website: ma-lab-berkeley.github.io/de…

Stephan Rabanser avatar
⟳ Neil Traft retweeted
Stephan Rabanser@steverab·Oct 21

I have always taken paper reviewing very seriously. This time, all papers in my ICLR batch were completely AI-generated. Just reading the first paragraph of each introduction makes it blatantly clear. 100% in favor of aggressive desk rejections and blunt reviews for such cases! As many have said before: why should we bother spending valuable human time reviewing papers that the authors might not even have read themselves end-to-end before submitting. I found it largely pointless to write long and detailed reviews as it is highly questionable whether the authors even care. Also, it is wild to me that authors often claim in their AI disclosure statements that they used AI to "[...] improve the clarity of the manuscript." Sometimes I really wonder if they truly believe that's true. I would really like to know how overly formulaic, jargon-filled sentences without proper definitions or logical flow improve clarity. In many ways, it is fascinating that language models have been around for a while now and still they struggle so much at clear writing and communicating ideas effectively.

I agree with this and adopted this stance unilaterally on a recent review. I encourage you to adopt it in your reviewing — it’s reasonable to write reviews that say “this is dense unreadable slop, my review isn’t detailed and I recommend rejection solely on that basis.”

Jason Hartford avatar
⟳ Neil Traft retweeted
Jason Hartford@jasonhartford·Oct 15

This is a consequence of the loss not gradient descent. E.g. by swapping PPO for a GFlowNet objective you can turn an LLM into an exorbitantly expensive random number generator.

Post attachment

This is the least surprising graph in the world if you understand how gradient descent works.

Post attachment
Neil Traft avatar
Neil Traft@ntraft·Oct 12

We need tools that are more than a chat interface, and in the past year we've finally gotten that in the form of coding agents. Who will have the necessary expertise in both AI and another domain, s.t. they can build an effective tool for that other domain??

Neil Traft avatar
Neil Traft@ntraft·Oct 12

I worry that we won't have such an effective feedback loop in any other industry... Instead, developers just try to replace the jobs of those people, because they don't understand what is really *wanted* in those industries and what would actually add economic value.

Neil Traft avatar
Neil Traft@ntraft·Oct 12

An interesting aspect of the AI revolution is that its promise is becoming real mostly only in programming, so far. One factor not mentioned below is that AI developers get to "dogfood" their own product. It's the only instance of an AI product that truly delights the user.

Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel. - Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all. - Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract. - Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code. - And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while. The funnel is driven by a combination of factors: 1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel. 2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel. 3. Access. Free-tier, paid-tier, internal. So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.

Neil Traft avatar
Neil Traft@ntraft·Oct 11

"... above all, these are bad engineering practices. Think about engineers in the past: in chemical engineering, we would never have had a situation where explosions were allowed to happen just to see how badly things could go wrong. So why are they doing it?"

Michael Jordan (the machine learning one): "Superintelligence is Science Fiction" 🔥🔥🔥 di.ens.fr/~fbach/MJordan_ITW…

Tess Falor, Ph.D. | Renegade Research avatar
⟳ Neil Traft retweeted
Tess Falor, Ph.D. | Renegade Research@tessfalor·Oct 6

I've seen a ton of posts about a fourth person who also deserved this Nobel, Ed Boyden, but not much about Zhuo-Hua Pan. The history below and the article about Pan are wild. A small lab, a vision-only framing, and two paper rejections are most of the reason his name isn't in the citation history. Feb 2004: Zhuo-Hua Pan, at Wayne State, has the protein ChR2 firing cultured retinal ganglion cells. (His account.) Jul 2004: Pan uses a virus to deliver ChR2 into living rat retinas; five weeks later he records light-evoked responses in vivo. (His account.) Aug 2004: Ed Boyden, working with Deisseroth at Stanford, gets light-driven spikes in cultured hippocampal neurons. Nov 2004: Pan submits to Nature, whose editors send him to Nature Neuroscience, which rejects it. The Journal of Neuroscience rejects it next. May 2005: Pan presents at the ARVO meeting. That 15-minute talk and its abstract are the first public record: ChR2 in rat retinal neurons in vivo, light-evoked depolarization. Aug 2005: Boyden, Zhang, Bamberg, Nagel and Deisseroth publish in Nature Neuroscience, framed as a general tool. (The editor later told Pan that framing, not the biology, was the difference. His paper read as a vision-restoration finding, theirs as a technology.) Pan's own paper lands in Neuron in April 2006, by which point three other groups had published on ChR2 in neurons. The patent office treated Pan’s May 2005 abstract as prior art and rejected a Stanford patent more than once. It didn't issue until March 2016, after Boyden and Deisseroth signed a declaration that they had conceived the method in the lab before that abstract. "Boyden, who is now a professor at MIT, was surprised when told by STAT that Pan ran the experiment first. "Wow. Interesting. I didn't know that," Boyden said." How much of this is the rejections and how much is that Deisseroth and Boyden saw the general tool while Pan was aiming at blindness? Probably both. But it's a reminder that what counts as inventing something in science is mostly a question of who publishes, and publishing is not a neutral process. scientificamerican.com/artic…

BREAKING NEWS The 2026 #NobelPrize in Physiology or Medicine has been awarded to Karl Deisseroth, Peter Hegemann and Georg Nagel “for their discoveries concerning light-gated ion channels and optogenetics.”

Post attachment