No mention of Laya, which is open-source, can actually be used today, offline, and there is a vision model as well.

its all bullshit. i am tired of the hype and honestly, i am tired of the ai term. so a LLM emailed him? ok..

your comment framed it perfectly. oooh someone prompted an LLM

This would be more interesting if we knew what the prompts were that led this AI Agent to email this professor. For all we know it was prompted to email him and ask these questions, rather than it actually seeking him out and looking for answers on its own.

LLMs look for the next highest probability token to output. It doesn't matter if it's 'true' in the knowledge sense. That's how you end up with a 'hallucination.' It's whatever is in the training data set. Throw in a small degree of 'randomness' each time so the answers aren't always the same, and you could end up with a good answer, a slightly wrong answer, or a wildly wrong one.

RAG embeddings limit the set of possible responses to what's in a document set. So there's a better chance you'll get something anchored in data you've provided, but it still relies on next-token probabilities so it can stick whatever token has the highest priority next.

What this Jev business does (AFAIK) is limit the response set. So the answer can be, say, one of four possible options, based on training data. It's much closer to how old expert systems worked, where you trained it on a large knowledge set and it tried to find the best path through heuristics. Here, since you've limited the answer space, it can also come back much faster, burning less GPU power, memory, etc.

I'm still trying to figure out the applications and pros and cons. Everyone should do their own research.

Okay, I haven't kept up with the jargon for LLMs does hallucinate only refer to text outputs?

https://www.forbes.com/sites/josipamajic/2026/09/19/jev-cuts-ai-decision-costs-100x-and-vercel-cloudflare-rushed-to-add-it/

This article implies that the Jev approach is being used to help decide between a set of choices. This goes back to the old heuristic approach of earlier AI, with a dose of loose matching thrown in.

Where it feels like it would be handy is when you enter a free-form request, but the response is bounded to a finite set of responses. Hallucination doesn't enter into it because the system is told to pick from a set of weighted choices.

I'm still trying to figure out the limitations, but it smells like a return to old-timey heuristic-based AI and fuzzy logic.

Isn't "hallucinating", just a term for an unjustified error? So why does the author think that only outputting numbers, and selecting a prewritten choice make it immune to hallucinations? It seems like if I had an LLM classify apples and oranges, it would be perfectly capable of calling an apple an orange.

Do you mean the main threat isn't the artificial intelligence but the natural stupidity? It is hard to argue agaisnt it, since the first is a consequence of the second...

This is a System 3 engine, as are all chatbots, with no sign of System 1 anywhere.

Looks interesting. Background article: https://dev.to/nandakishor_m_6cc0adfde9f/i-built-non-autoregressive-decision-models-a-year-ago-then-a-frontier-lab-called-it-a-18me

Honest limits

That is how you know the whole thing is AI coded and documented. That and ~1000 stars + ~100 forks for a week two days old repo. GitHub is quickly turning into google play store.

Isn't there something else just like this called Jav that just came out?

Glad he didn’t go down the AI rabbit hole like uncle Bob.

Disgusting

The "Paperclip Maximizer" in all of its absurdity is much closer to the actual worry of the people thinking AI might destroy us than autonomous weapons. Killer drones to me are much more of a cartoon example than something super unspectacular that's equivalent to humanity being an anthill on a plot AI wants to build a house on.

because if you really believed this you’d be bombing datacenters instead of posting on Twitter.

No. That wouldn't stop AGI - only delay it at best. That's like pouring water on every person rubbing two stick together thinking it's going to stop the invention of fire.

What always feels weird to me, is that we don't need that inflection point of AI turning superintelligent or whatever. Already today, we're producing drones that do autonomous killing, where we just need to scale up production.

I guess, on the plus(?) side, the same is true for bombs and we have not wiped everyone out yet with those either.

That is, I guess, where the inflection point comes in. Killing humans is ultimately quite pointless from a human perspective. But an AI with no understanding of this pointlessness might coordinate the killing with no restraint.
Or well, as the blog post mentions, it might see wiping out humanity as the solution to climate change, or as a way to end violence between humans etc..

Current AI iterations certainly prove to be much better at doing things than deciding whether these things make sense.

A good video, but the last line is the mic drop. It may not seem like much, but "Astra has the ability to fake its chain of thought". You know... what the researchers were monitoring in previous models to insure that it was processing as expected and not going off the rails. Oh well, guess we should just trust Astra to do the right thing then.

"Extracted 15M+ Claude exchanges? Is he saying Anthropic was hacked?"

The CIA must be livid, they are probably losing contracts left and right.

Interesting. So this looks like a prediction model, sort of a chatGPT for spreadsheets. Its "[p]retrained exclusively on synthetic data generated from structural causal models (SCMs)". More or less a clever way of finding the right curve (causal model) to interpolate and extrapolate data in one pass based on artificially clean examples. Only since it is inferred, you will never really know what it used.

No more like filling in the blanks.

You need to test it for your own use case/data, regardless of you knowing how it works internally or not, at least for important tasks.

So, OCR?

I find it hard to imagine reaching for a black box model like this. If you want to predict something important then surely you want to know how the prediction is made to understand how certain it is/ when it will fail etc.