Neil Traft

Just wait till you hear about 0's decision...

Even here, they miss the point. They somehow don't realize there is no universal alignment. Our species is a collective—a collective of collectives, even. To be "aligned" is just to be part of the collective; an integrated whole. Which requires us *all* to participate in building & integrating AI.

OpenAI, of all companies, should know this.

Ah, but they were not *exactly* created to further AI technology. They were created specifically to secure against the mythical AI End Times, the Apocalypse. So they closed down rather than opening up.

The best thing a company could possibly do is to create an ecosystem around its technology. Everyone claims to want to be "a platform, not a product". But their behavior belies their big words.

"Platform" should mean an open foundation for building. Not a limited-use API gateway.

With its focus on short term profitability, the US has forgotten how to make markets competitive. And companies have forgotten that they are only *vehicles* for the final product, not ends in themselves.

As long as we agree that no state space is truly markovian, I can get behind that! 😛

What do you mean exactly? Can a state space be fully observable, yet incomplete? Or are you just referring to unknown transition function or something else?

Good news, those are all partially observable, so you get to keep those! 😉

In retrospect I should have said RL should never be *framed* any other way... Except for maybe a chess game.

kermit the frog is using a typewriter in a messy room . ALT: kermit the frog is using a typewriter in a messy room .

I’m very excited to announce that I’ve just signed a contract with @princetonupress.bsky.social for a new book, tentatively titled “The Genomic Code” 📖 😊

In reality, all environments are only partially observable, so this is the only regime in which RL should be evaluated.

Hmm, I guess at that point I'd call it something other than "LLM" and "Transformer"; even if we accidentally use the same name it seems to me that would constitute a significantly different system. So I still think this is a core justification for why the current paradigm will never be enough.

Guessing you're talking about a situation where there are multiple equally-good opportunities to turn, so Google should say something like, "if current light is green, keep going; else, take the turn now"?

I've always held that the ideal attitude for a PhD is, "the experience of doing a PhD is worthwhile in and of itself, regardless of what comes after". Even if you continue as a researcher, you may never have as much freedom to learn/explore as you do now. (Assuming a good advisor.)

Send like a useful framework. Looking forward to reading this new position paper.

Wonder if a truck and a car count as different embodiments? Probably still workable even if not. But maybe sedan & pickup = same embodiment; tractor-trailer = different embodiment.

I just realized that the "AI scientist" vision is all about this "collecting of facts", like packrats or crows.
bsky.app/profile/ntra...

And you would think that this would finally be our chance to focus on that now that we have language models! They could actually be very good at explanation… If they actually “understood”, that is.

It's quite revealing... in tech circles nowadays we are so fixed on the conception of "science as innovation", completely leaving behind the other half of science, fundamental understanding.

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! 🧐

So, you need *some* adaptability to changes in body type. But I don't know exactly how far "one brain, any embodiment" folks would want to push that. It does sound extreme.

On the flip side, if each brain is so specific to its embodiment that, e.g., every time you want to tweak the sensors on your self-driving car you need to *completely re-gather all data from all cities and make all new maps*, then this would be pretty devastating to that use case, no?

Ehhhh, no I unfortunately don't think so, because if someone is doing "the right thing for the wrong reasons" then they will absolutely be led astray into doing the wrong thing, usually sooner rather than later. Depends on the severity of their wrongness, perhaps.

Team from the University of Vermont + External faculty posing for a picture by a river

VCSI executive director announcing location for the conference in 2026, standing on a large stage with colleagues

female scholar presenting a parallel talk, engaging with the audience

PhD student presenting a research poster

Proud of our team's talks and posters at #IC2S2 in Norrköping Sweden - what a great week. We're excited to host next year in Vermont!

Women Dating Safety App 'Tea' Breached, Users' IDs Posted to 4chan “DRIVERS LICENSES AND FACE PICS! GET THE FUCK IN HERE BEFORE THEY SHUT IT DOWN!” the thread read before being deleted.

We are going to spend the next few years finding out exactly *why* it was a horrible idea to unleash a tsunami of vibe-coded apps made by idiots and scammers on an unprepared world.

Welcome to the Entirely Foreseeable AI Consequences Era.

Apparently Papers With Code was abruptly sunsetted for reasons that are unclear. 😢

Trending Papers - Hugging Face Your daily dose of AI research from AK

😞 So sad to see paperswithcode is discontinued – but grateful that the @hf.co team is, as always, stepping up to support the community!

It was an incredible resource for use cases, common themes in papers, and visualizing how models have improved on evals over time:

huggingface.co/papers/trend...

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")

There are people, in tech (and now in the government!), who will mislead you about what current AI models are capable of. If we don't call them out, they'll drag us all down.

more evidence for “a bunch of AI cultists dismantled large parts of the federal government because they believed that their LLMs were intelligent and could do the work better” thesis

He's absolutely drunk with power... and summer tomatoes