First, congrats to the team on launching something genuinely interesting and new.
Seems like a more accurate title would be "Jev: Trading general purpose generation for fast typed inference" or something like that.
This is interesting, but the speed comparison seems misleading? A generative model that can output code in a Turing-complete language can do anything a computer can do.
Jev can only generate structured output, right? This is probably super useful for classification/routing/scoring, but it's nothing like the code generating models we're all using today for code and automation.
Also "can't hallucinate" seems wrong? Sure, it can't emit an invalid type, but it can still emit a completely wrong valid value. You can enforce structured output from an LLM too, with an appropriate harness, etc.
Assuming there's no funny business, the Doom demo is cool.
> Assuming there's no funny business, the Doom demo is cool.
The Doom demo seems very funny business. They're not feeding it video, they're feeding it a text description of what's going on in the game. It's not reading pixel data.
I think LLMs would play a lot better with that input too but Jev does seem to have a huge speed advantage; I don't know if the other models could do that in real-time.
I think the meaning of can't hallucinate in this model is that the type won't be hallucinated.
So if the generated schema is for a tool call for calculator, then the numbers will be valid numbers for sure (and not random words).
To me, it looks similar to BNF schema already introduced and implemented few years ago: generally speaking - it limits the next token that is allowed to be generated, probs are drawn from a subset tokens.
(tbh, I'm not sure why it didn't pick up as a more standard interface to LLMs, as it made a lot of sense back then, and now.)
When they say "can't hallucinate" they mean they produce a confidence value for every result, so you could see for example it has 0.1 confidence, and you can disregard the result - that'd be different from hallucinating where it believes it's correct
I would be happy enough with: only produces what it can verify with sources.
If you eg try to remember a court case (ie produce the reference via LLM token generation only), it's easy enough to check with your data whether it really exists. Similar for following links and other references.
If your data or sources are wrong, obviously your report about them will be wrong. But I wouldn't call that a hallucination.
It's not a binary thing. You can get closer or further away from that standard.
And humans also behave differently in different contexts. A conversation at the pub has more such hallucinations than a formal deposit in court. For the latter, a good lawyer will look at her shoes, when you ask him what colour her laces are.
The probability values don’t really represent confidence in modern LLMs though, especially after RLHF and RLVR.
System One says they use RLCD, Reinforcement Learning for Calibrated Decisions, which presumably has accurate probabilities as an explicit optimisation goal.
RLVR generally upweights tokens along the whole thinking trace that led to a correct answer, whether each token was "correct" or not. RLVR doesn't train a model to output an 80% likelihood, it just trains it to produce correct answers, and not to produce incorrect ones.
System One hasn't said how RLCD works, but they do say it is explicitly training models to output "calibrated" probabilities, which makes it distinct from RLVR. This is how they describe it:
> System One models are trained for calibrated decisions: their probabilities are optimized against outcomes to reflect uncertainty.
that's right, but because these models are probabilistic, it's also possible to be confidently wrong (and all future models will be smarter still and still have that possibility)
Nothing, but imagine using LLMs for a classification task
People out there are so resigned to the models being unreliable that they are really doing things like hallucinating deliberately, and then matching the hallucinations to embeddings -
I'm certainly not resigned to that, at least for classification.
Even non-frontier models are absurdly good at this in a broad sense.
Which would make it hard to judge "a model that will never produce unreliable outputs in the first place" against something that is already really, really good and exceptional in domain-specific areas with the tiniest amount of elbow grease.
What about the LLM calls though that are done midchain? In the Home Assistant video the multi-intent prompt gets split using what looks like a traditional llm model, which I'm assuming is vulnerable to classical hallucinations.
Has LLM become so synonymous with Generative Transformer that other high-parameter count models that interpret language need a different name?
For all we know this might be a non-language-generative transformer e.g. a transformer where the decoder produces confidence scores rather than language. Please provide more likely architectures if you know them, I'm genuinely curious.
I don’t think it’s misleading if you compare on the use cases they suggested. It’s faster and cheaper (no idea if higher quality), so it’s immediately interesting for certain things.
And if you buy their RLCD claims, this might be even better than huge models that know a bunch of irrelevant things.
This is likely still an LLM (in the purest definition of a language model with relatively many parameters) since the inputs are natural language, just not a generative LLM as the output is something other than more language.
How is this not a frontier model? It's bleeding edge in its own niche. It's not a frontier LLM; however, applicable to many of the things people use LLMs for.
It's nothing like a traditional LLM and so should not be compared to one. It's a heavily constrained, tiny model that can only produce a probability score or a yes/no answer over pre-defined selections. It has no long-context capacity.
I mean, imagine comparing this thing to Astra, it's hilarious. They don't even tell you what the max input size is, and they only allow 10 possible answers to choose from for the Choice mode. It's probably like a 1billion param model. They say it's "not small", but there's zero reason to believe that.
I suspect someone will be able to recreate this within a week by piecing together open-weight models.
> It's nothing like a traditional LLM and so should not be compared to one.
Frontier LLMs are expensive jack of all trades. You can absolutely compare them to purpose-built tools on any domain they touch. Engineering is all about assessing tradeoffs.
I'm biased but I wouldn't call it misleading - generating text is super awesome and flexible, (we describe that in the blog post - and I personally use string models all the time) but it's true you pay a high tax for autoregressive generation
> Also "can't hallucinate" seems wrong? Sure, it can't emit an invalid type, but it can still emit a completely wrong valid value.
that is likely true of all ML! perhaps we could debate semantics, but I don't think it's fair to say a random forest "hallucinates" in the way LLMs do
His claim was that the title is misleading, not sure how it's relevant to that claim that you use "string models" (full LLMs).
The original title before it changed less than an hour ago was:
"Jev: New frontier model 40-400x cheaper and 20-200x faster"
I'm going to agree that was misleading.
And on the second point:
>>Also "can't hallucinate" seems wrong? Sure, it can't emit an invalid type, but it can still emit a completely wrong valid value.
>that is likely true of all ML! perhaps we could debate semantics, but I don't think it's fair to say a random forest "hallucinates" in the way LLMs do"
Also going to disagree here, and I don't think it's semantics.
"Hallucination and type-safety are intrinsically related"
I'm not entirely sure why we're conflating type safety with, I guess, value or output safety.
"Would you say a linear classifier hallucinates?"
No, but it can be (and often is) mathematically correct and functionally incorrect. It doesn't help to say "a linear classifier can't hallucinate" when you get even 99% accuracy. That's 100% a semantic play, and it doesn't help when the picture of a dog is labeled cat and the response is "yeah but that's not a hallucination, only stupid LLMs do that"
And furthermore, because the model is forced to answer in a boolean (if in boolean mode), if the user input is outside of the range of a boolean, it's forced to hallucinate. It can't abstain.
User input: "Hey, have your human support agent call me, tomorrow at 5pm."
Model input: "Does the user want to speak to a human support agent?"
Output: Yes.
I imagine that your model would produce this, and I think it's fair to say this is a hallucination. A human would caveat it with: "Yes, but not right now.", your model is incapable of that. Yes is technically correct, but within the context of being in a live chat, a human would understand that the caveat is required.
No. Your launch post puts “0%” on a hallucination chart, then explains that the number comes from guaranteed schema matching.
You’ve already agreed that this doesn’t establish correctness. An approve for an unauthorized action still meets the schema guarantee.
That’s why I find the messaging misleading. You’re acknowledging the limitations in these replies while defending the broader reliability pitch.
Even granting that each answer is calibrated individually, that doesn’t establish calibration of the decision that combines them.
Sure, I can threshold a composite score, but there may be many wrong answers with the same score. An unauthorized action doesn’t become acceptable because it scores highly on the other dimensions.
I still have to define the constraints and test which wrong actions get through the complete workflow on my own data. That’s a substantial part of the work being pushed back onto the developer.
Let's say classifiers don't hallucinate. To make a fair comparison we should constrain LLMs to the same classification task. In that case, no, LLMs also don't hallucinate.
- Give Jev and LLM the same input
- Lock down both to approved/rejected/unknown (LLM restricts on decoding)
- Both can be wrong, but neither can hallucinate (invent an another option).
A hallucination in the context of LLMs is generally understood as an incorrect answer presented as factual. If you claim that "x can't hallucinate" in the context of LLMs, you're saying that x always gives accurate answers. It does not matter whether the answer is type safe. If its value is incorrect, it's a hallucination.
Just to be sure that I understand, you're saying that your model "can't hallucinate" because it only outputs a single thing, right? In this way, an LLM can't hallucinate either if I prompt it to do a classification task with a discrete set of possible outputs, right? (Assuming I reject non-conforming output. Actually, maybe what you're saying is that your system can't output non-conforming output?)
For e.g. classification tasks, even in 2026 people are doing things like hallucinating deliberately, and then matching the hallucinations to embeddings -
From a quick look at this it looks like it could easily generate natural language text by following a structured representation like UMR (Uniform Meaning Representation) or the similar representation the Abstract-Wikipedia folks will be working on for generic encyclopedic text (which will be heavily informed by Universal Dependencies). These are basically linguistically principled and frame-based counterparts to a programming language AST, that can be then converted to natural language (in a broadly language-independent way, to the extent that semantics and pragmatics make that feasible) via some sort of NLG rendering.
(To be clear, this one raw model does not support outputing a full AST directly - it wants to output "choice" among fixed options, "score" on a sliding scale, or a true/false answer (all of these with confidence scores attached), so building the AST/structure would be a code-driven (or even perhaps outside LLM-driven in some more challenging cases) multi-step affair where the model would essentially be playing a "game" of building the structured output step by step and getting a revised partial state back. But one could expect this to lead to interesting results.)
Only if you think that everyone cares about self-driving. Lots of niches require structured domains; self-driving is just one that has a lot of capital thrown at it.
I feel like the power of the approach presented here is that it gives a model a proper "language" to describe computations directly vs moving tape silliness.
I foresee this to be the path moving forward - giving AI models understanding of the computation directly(as well as compositional rules) This feels like a short path towards total software in many areas.
Agreed. It's a wildly dishonest presentation of their product from many perspectives, which is a shame because it might actually have some good use cases.
The comparison between LLM speed and Jev speed is misleading, because they're using autoregression to generate all of the type names, all of the schema, etc. A closer comparison would be if the LLM was purely outputting the raw numbers. Even then, comparisons to LLMs are pointless because you could train a transformer on the same sort of task that Jev is doing and get even better performance yet again, and a smaller model. I suspect this is some form of stripped down diffusion language model.
You really have to do a lot of hand holding here, and map out your problem space manually, and very carefully, to get any sort of accuracy. For example:
> Keep each Score to one dimension. If a description says “punctual and smart and experienced”, the question is measuring three things, and an input that is high on one and low on another can’t be placed. Confidence drops and the score means less. Split it into one Score per thing and combine them in code
If you don't perfectly represent the distributions of possible answers then you'll likely get garbage results. As far as probabilistic state machines are concerned, I'd say creating the distributions of possible answers, and their hierarchy, is the actual hard part.
One of their examples is:
- "state": "I have asked three times now. Can I please just talk to a real person?"
- "Is the customer asking for a human agent?"
Imagine the users request is: "I want your human agent to call me tomorrow at 5pm."
Human conversation is fuzzy, getting useful reliable results out of this is going to be a challenge. Of course, you could add follow up checks like: "Do they want that now, or later?" -> if later -> "Do they want that tomorrow, or the day after?" and so on... But now you're building an LLM out of if statements. I am skeptical of whether this model has much utility for fluid language interpretation - I suspect it'll only be useful for scenarios where you've tightly constrained the answer space but want to use fuzzy language to describe it. Like:
- Question to human: "Would you like a support agent RIGHT NOW?"
- Their response: Yes | Yeah | Mhmm | ye sure (any possible yes signal)
Model input: "Did they ask for a support agent?"
Still... a tiny LLM could accomplish this sort of thing without problem. And that doesn't stop someone from saying: "No, not right now. But tomorrow." - and the tomorrow would get missed. I think this is why people haven't really tried this approach much already.
Also their Doom demo is on structured state, not on images. Meaning, the enemies must be being served to the model as coordinates (or the exact angle of projectiles that hit the player), otherwise it'd have to scan every pixel of the 360 degrees to know whether an enemy is in front of the crosshair or not. You can see from the map below that it's also choosing travel checkpoints/destinations through walls. So they've severely cooked this to make it look far more capable than it is in practice, and any speed advantage that is offered here is not factoring in the shortcuts it is taking, the training on the map, and the fact that it can cheat because the structured state it is using is not bound by obstructions.
Side note - just like most people don't need an intelligent personal assistant to manage and respond their emails and book their flights, most people also don't need smart homes. Century old toggle switches are more than enough in a 3 room apartment or 5 room house unless you have a mention.
My primary beef with smart home (having tried it) is that every person that visits your home ends up confused about some element of it. A light switch that goes up and down is universally understood.
Also a quick NFC sticker in each room taking you to a small HTML site containing settings (temp, ventilation, lights, shutters, setting a alarm by the lights) has been golden.
No one will install: Shelly, AC app, look for ventilation IR controller or figure out how to use the Shelly app to turn lights into an alarm.
But tapping your phone on a NFC sticker bringing all those controls together per room in stead of per category (all lights in Shelly app. Person in room #1 has no interests in the lights in room #4 at the same time.).
One "all house" sticker next to the front door allows any last person leaving or first person entering to put the entire house in active / idle mode.
Works wonders. And as soon as local AI is quick enough the stickers will be a microphone!
We have 8 light buttons in our living room/dining room/kitchen space. It is very convenient to us that we have 1 button for turning all of them on/off at the door to upstairs (at night turn off all lights and go to bed upstairs, in the morning come downstairs and turn on all lights) - but also have 1 on/off button near our back door for when we leave/come home.
Next to that: on/off toggles a schedule where the lights are bright and cold-ish by day, and low and warm by night without us having to manually adjust each light every hour or something.
Again, need is a big word. But it's very convenient and pleasant.
This is very cool. However I don’t really want to bounce all my home automation commands to the cloud. I hope there will be an open weights approach one day. I’ve spent a lot of time setting up my local only home automation system, it would suck if it didn’t work during an internet outage, and also there are obvious privacy problems.
Rereading some things and because there's no official benchmarks, I misspoke about the ballpark comparison., but the open model's still a useful foundation to work with
Guess I'm a bit less impressed seeing that for some of the more intelligent driven+action work -- splitting requests in the video -- they had to kick out to an anthropic model.
That's fair, but it highlights how this would actually be used. It doesn't really seem like a competitor to other models but instead a way to make these real systems more enjoyable to deal with.
Haiku, to rewrite a sentence as two discreet commands.
I agree that it was notable that they delegated to an existing LLM, but I don't think it detracts much from the value proposition (not yet proven) of their demo.
Agreed but is it much easier to deal with if you need to have all of these sub processes integrated? How does one know when you need to reword a request? What if Anthropic then has a type error, then debugging that just got harder.
This, combined with contracts, could make a lot of things so much fun now!
For those who don't know (which is probably everyone but me), I ported the design-by-contract pattern in Python and combined it with LLMs. This was early 2025. I originally wrote about it here: https://leoveanu.com/2025-03-01-dbc/
. Contracts are a core feature of SymbolicAI ever since. The community seems to have loved it too (https://news.ycombinator.com/item?id=44399234).
I think I'm starting to glimpse the implications and it's gonna change agentic workloads if it holds up to scrutiny. It's too early for me to tell anything other than jot down some rough thoughts.
In short, you get blazingly fast semantic branching you can use in control flows. For contracts, I can now directly take the data model that you have to design and convert it into Jev's expected format. Or I can use Jev for semantic branching in postconditions.
If my understanding is correct, that should be doable, but I need to think more about it. It could be that with Jev I can finally “compile contracts” and better chain them into workflows, which is something I always wanted but didn't know how to do properly.
This is a very promising idea - a model that takes arbitrary text input (which can be a complex json), plus a set of questions (yes/no, multiple-choice, or score) and quickly (milliseconds) and cheaply ($0.042/MTok) answers those questions.
Unfortunately, none of this is explained in the announcement, but the documentation [0] is pretty good.
I am not an expert in this domain but as an engineer-turned-researcher, this looks a lot like GliNER with a fitting harness.
This is something I focus on in a bunch of my experiments - how to get immense value out of tiny models (<1b params). There are lots of different architectures out there and there is so much to optimize if you know what you are asking and have a grammar to constrain with.
Great to see this and I hope this is a lot on top of what is already openly available.
Looking at the example Jev use cases, it almost feels like Jev's incredible cost/task can make it competitive as a generalized "poor man's ranking" algorithm that can be useful for lean startups or any fast paced development org.
I need to rank 1000 articles and pick the 5 most relevant for the user? Jev.
I need to audit and strip out content because my user is affected by regional privacy laws (without hallucinating)? Jev.
I need to surface the 3 funniest media comments that match the user's sense of humour? Jev.
I'm sure someones working on this as we speak using an open-weight LLM base (Qwen or something would be a perfect fit).
This sort of task is a perfect fit for a very small model capable of semantic parsing. You can get away with a LOT less parameters without all the autoregressive generation and long-context reasoning.
It seems like the docs[0] are a better explanation? The comparison to llm tokens is kinda confusing.
It looks like the model takes as input a state (structured text? not sure if multi-modal) and a question (as a "Choice", "Score", or "Noul") with some additional augmentations possible. Then outputs the question's answers as appropriate (e.g. a choice, accompanying probabilities, confidence).
Edit: On the AI primer page, it looks like they do the RLCD on a pre-trained base model?
I do agree that the comparison to LLM tokens is hard to understand (also because output tokens are not comparable).
But yes, text or structured state (like a JSON with multiple pieces of text in) -> decisions out (e.g. choice maps to "match" statement, "score" maps to sorting, "noul" short for bernoulli maps to if-statements)
Small request, can we get an explanation of the naming of "noul" in the docs[0]. I tried googling, and searching the docs and didn't understand why it was called that.
(I'd also argue something like p_yes or just probability might be a simpler name, but I'm sure there's a better reason behind Bernoulli maps).
here is how I attempted to explain it to my company's AI group chat, is this roughly accurate?
"instead of autoregressive string output it instead outputs structured type-safe 'decisions' with probabilities/confidence scores, each generated in parallel
so sort of more like a Large Classification Model than a Large Language Model? or, maybe better to think of it as a sort of "shift left" in the LLM's transformer architecture, allowing you to replace the predefined token vocabulary of an LLM with a prescribed set of 'decisions' that need to be made based off the input context; and exposing those probabilities directly so they can be integrated into the system logic, instead of just sampling from top-K.
all of this while still being instruction-tuned (!!!)"
It's always been possible to build classification pipelines using LLM embeddings as the input. seems like this is a much more sophisticated / useful application of that concept
the one nuance I'd get into is I'd call it "zero-shot" over "instruction-tuned" (the latter often implies a particular distribution), but very safe for sharing
Hi - first congratulations, System One looks really promising.
The Doom demo really help me, at least, to understand how System One differs from LLMs. However the first demo (Side-by-side demonstration) - I'm struggling to understand what is going on here!
The demo is showing System One producing its output in parallel very quickly and for little cost compared to an LLM generating its answers token-by-token. The "noul" type is used to evaluate a yes/no question and return the probability that the answer is yes.
So this demo is showing System One offering much more nuanced responses and specific probabilities compared to an LLM's more crude responses (e.g. LLM shows "true" or "false" compared to "0.9" or "0.07" probabilities that the answer to some question is true).
I see this super interestingly as the "subconscious" to the llms "conscious" for lack of better terms. I'm super interested in this for broad and rapid decision making in the context of consumer agents so will be signing up for sure.
1. I am extremely on the same page
2. I do think that subconscious is not only much smarter than we give it credit for, but also much more robust than the "jagged frontier" of current LLMs
For many day-to-day computing use cases, Jev seems far better suited than an autoregressive language model, if for no other reason than it is not wasting compute thinking about anything other than how to spit out a decision.
Do you have an architectural explainer yet for Jev or are you holding that close to your chest and letting the magic rip for now?
> the model takes as input a state (structured text? not sure if multi-modal)
Input, and criteria/instructions can both be defined as structured input (JSON). This ends up being pretty powerful because the model is trained to understand structure.
I assume this isn't really for consumers/individuals currently? Kinda feels like an improved magic 8 ball.
I can't really intuit how I should think about when the model will be accurate. Is there somewhere to read more about that? I assume customers would just have some tests or talk to you.
> I assume this isn't really for consumers/individuals currently?
Unless they're hackers, no. It's not really a chat interface, it's meant for consumption by machines and composing into higher level systems (pairs great with LLMs).
> Is there somewhere to read more about that? I assume customers would just have some tests or talk to you.
We're going to release some more info on evaluations over time, and yeah, join the waitlist! We offer faster access in exchange for good memes
After much fumbling around with prompts and evals, this is exactly how I am using LLMs in production, to narrowly make choices and return structured data. Any deterministic work gets pulled out of the prompt and my goal is to narrow the model output to be as clearly defined and as minimal as possible.
Jev's focus on structured I/O and confidence scores are game changing. If this does at all what it claims, I think this is going to quickly become the new standard approach for agentic systems.
Great question! Yes, this works much like the doom player. Sensor data (LIDAR, velocity, etc.) becomes the state. You use the score primitive to operate the controls ("What level of braking should be applied" 0: None, 1: just slightly slowing down, 2: there's a suspicious cat on the side of the road you don't trust, ...
Not confidential, but not super relevant, as this is something I have learned the hard way over the past year across various projects.
A lot of people have become prompt maximalists, asking for complex multi-part solutions or dynamic workflows in a single prompt. You can get this to work sort of reliably with frontier models, but without much confidence or clarity where things might break in practice. My goal is to strip out as much determinism as possible from prompts so the LLM only needs to handle a narrow, well-informed decision, like "Pick one of these three things" and build around the answer. Sometimes you need to fill out a whole JSON payload and LLMs really actually suck at manipulating and adhering to JSON. They do ok now because labs have put in a ton of effort on making harnesses play nice with structured data. But it comes at a high token and context cost because under the hood I suspect the model is churning invalid text repeatedly until it gets around to passing some internal validation.
Example I have worked: Personal delivery app, that tracks packages from various senders using incoming emails.
I am using the single prompt approach with GPT5.4, which is free, but it’s not reliable. Using Jev I’d decompose the prompt into a bunch of smaller questions, then I’d combine the answers in software. I’m super excited to try Jev out.
Or a partially completed song, asking for the next note. I’m not sure if you’re joking, but using it for space constrained next token generation within a grammar sounds like a really neat use case.
I think what this shows is how important branding and comms are. They've captured imaginations with their demos and nomenclature, despite the arguably non-novel architecture. One forward pass, read the embedding space, train some regressors on predicate structure, [??]
I would love for things like this to be accessible via hubs like open router or AWS bedrock. It's hard to justify adding new model vendors directly with all the heightened concerns about privacy and security, but if bold new capabilities are added to a centralized already-vendor like AWS, technical people can adopt them without going through a whole compliance/purchasing/vendor review process. And an extra middleman tax is well worth it when the cost savings of the model itself can be one-two orders of magnitude.
I think the trouble is that Typesafe APIs don't fit into the normal OpenAI-style API that every other regular LLM provider users. You're not just providing unstructured text and getting unstructured text back. It would take a different request and response format than every other model on Open Router. Though you could shoe-horn it in some way, it'd be hacky.
But agreed it'd be very useful to see it deployed on other hubs, and it seems worth it to provide the bespoke API format. Perhaps Typesafe's API will end up becoming the standard for a new type of structured model, the way OpenAI's API did.
You can set privacy requirements and define an allow list. To me the main value prop is that I get one bill for all models and can quickly try new models without signing up anywhere or changing my code.
Oh! Also you can pass an array of models and if the first provider is down it automatically falls through to the next provider. More useful than it should be...
That's still ultimately privacy by contract (where you have to trust the inference providers to uphold their end of the deal), rather than privacy by design.
I'm trying to parse it down to what we had before vs what is new here.
We already had encoder models that skipped text generation for giving us a numerical output that could be computed as a probability. we also got no hallucinations and faster inference for free there. So we already had
1. "unstructured state in, probabilistic decisions out"
2. "orders of magnitude faster and more efficient"
What was hard there was to train the model head without ML expertise, and considerable amount of data.
This seems like this is a democratization of those encoders? The addition over existing encoders seems to be coming from being able to specify the output shape (up to a cardinality of 255). It is unclear to me if this is possible using Jev without additional labels for fine-tuning.
If so, that is still very impressive, but I think the faster inference and 0 hallucinations might come for free, from it not being generative.
This sounds good but so far all claims just sound like marketing terms. I'd love to see real proof. e.g. "RLCD" and "parallel sampling" have nothing to back it up.
also "70-500ms vs 3-329 seconds" are apples-to-oranges unless the LLM baseline is doing comparable work (e.g., long chain-of-thought). If Jev is skipping generation entirely for a narrow structured task, of course it's faster.
Nonetheless i want this to be true, so I'm looking forward to Jev
They have various benchmarks, e.g. how much time it takes them to do wikipedia page -> page games. Jev seems to take the same or fewer hops but in ~10x less time and for ~10x less money.
It's totally reasonable to compare against LLMs doing chain of thought if it gets comparable performance.
> If Jev is skipping generation entirely for a narrow structured task, of course it's faster
I think this is reasonable if people are actually using LLMs to solve this type of narrow structured task, which they are. The evidence is that every LLM provider has some method of forcing the output to conform to a json schema in their documentation.
BTW it was not multi model playing doom, it was passing structured input and getting structured output. Its not what I thought: frames of video passed and real time game play.
> I really have to say that I like their manifesto
Their manifesto: "you only build on top of it if it's trustworthy." - the irony of this while putting out the most misleading, dishonest marketing campaign I've seen in months for their first public appearance doesn't exactly scream "trustworthy" to me.
I am confused why they say it is not an LLM and then in the documentation it is shown as being an LLM derivative. The documentation makes it sound like they're taking a pretrained LLM and then giving it their unique post-training. How is that not an LLM?
FAQ:
Is Jev just a smaller LLM?
Jev is neither small nor an LLM, hence being off the intelligence Pareto curve.
LLM seems to have become synonymous with Generative Transformer architecture.
While this model may share much with GPT-style models on the encoder side, it clearly has a different decoder architecture. So is a high-parameter count language model an LLM even when it doesn't have a GPT-style decoder? The definitions are in flux.
I could see this being fantastic for classification tasks. Last year I shifted from using LLMs for bulk data classification tasks (1M transcripts) to generating embeddings and categorizing based on cosine similarity. It saved a ton of costs and time, but wasn't as accurate as LLMs. This seems like it can give me Terra-level classification ability with the cost/speed I need.
Overall this seems like a classifier that gives weighted scores per custom labels. It's certainly useful, but whether it brings higher quality than an LLM in structured output mode has to be seen in objective benchmarks.
Zero shot classifier indeed. Reminiscent of asking an llm a yes/no question, constraining the output to either yes or no, and looking at the logits directly
And each question is a separate single token model completion done in parallel
do you all see the use cases being similar to what you might use Fastino's Gliner models for? i see similar differentiation from general purpose LLMs in the sense that they can take natural-language input and return outputs adherent to a user-defined schema.
im thinking about how well Jev could be used to replace a current LLM-as-Judge evaluation workflows, specifically on chat transcript data (think ~1,500 tokens) i wonder if the reasoning usually required pushes it a bit out of scope. didnt see anything published about constraints on the state size, so would be curious to hear about that.
definitely seems like a modified version of GLiNER2 or 2.5:
- encoder-based (no text generation)
- multiple tasks in a single forward pass
- deterministic outputs
- constraint-based classification
I think this is a great direction -- for some kinds of users. And this makes me wonder if the 'vs' framing is misleading.
Yes, I think it's a mistake that many organizations are cramming LLMs inside of automated pipelines where the extreme generality/flexibility of the model is at odds with the fact that you're using it for a very specific task that gets repeated over and over, and needs a very specific structured output to be successful. But specifying your task carefully (as well as deciding what counts as your input state representation etc) seems like a form of programming. Something (a person or a model working in a relatively unrestricted way) will need to produce a configuration/specification for this system.
So rather than Jev vs Claude I imagine that using Claude/ChatGPT/whatever interactively to define / refine your Jev config which then runs in prod might be the happy combination?
There's a whole lot of information on this page that doesn't tell me anything about what this actually is. Can anyone spell out what the architecture is here?
They claim it's not an LLM, which I read as "not an auto-regressive token generator". I assume they are still using a transformer, otherwise they would be talking about the thing that's not a transformer, instead of all the fluff on the linked page. But they emphasize parallel generation, so is it like a text diffusion model?
Sounds like its essentially a generalized zero-shot classifier that takes and option set at runtime and works on unstructured inputs.
you pass in your "prompt" and options (described in natural language) that it can respond with, in addition to your input. it gives back that option set with a probability assigned to each one
> Input tokens: $0.042 / MTok ($42 per billion tokens).
> Output tokens: FREE (too cheap to meter).
Insane. The video demos are really compelling, in particular the speed.
> Structured outputs slot into ordinary software as fuzzy decision rules: classify, route, score, extract, or branch where hand-written logic is too brittle. The surrounding code constrains their freedom, making them easier to compose into reliable systems.
I buy this vision. A lot of LLM integration I see these days is ultimately exactly this. OpenAI-style structured outputs works decently but this would be a great improvement in cost, latency.
constrained decoding (OpenAI-style structured outputs) make models dumber unfortunately - the short+dense version is that simply masking logits is insufficient because if ever a model was assigning probability to an invalid token, the model is by definition confused. you'd be better off erroring IMO
I'd love to know if Jev is still fundamentally LLM-shaped in architecture. Like is it using a single forward pass with a learned readout over the predefined options (i.e. a discriminative head on a transformer, no decoding), or something else? I did similar things for zero-shot criterion-based classification using a 4B Qwen model but could not reach the level of intelligence they've got here. Tho speed/cheapness was similar.
The model can't reason comprehensively (e.g., like Sol XHigh would to solve a complicated problem), but it's designed to be able to answer anything a human reasonably could quickly and intuitively, i.e., system one thinking: https://en.wikipedia.org/wiki/Thinking,_Fast_and_Slow
I had a different initial confusion - it seems this company has no relation to the company formerly known as Typesafe https://en.wikipedia.org/wiki/Akka.io
It can't hallucinate, but it doesn't mean it can't make wrong decisions. Just because it adheres to a specific output format at all time, while LLMs have the output format at their mercy, then the claim of not hallucinating is made technically true.
I think that this specific part is not super interesting if your harness just recovers from invalid LLM outputs.
The latency and cost - yes, those are super interesting.
Wow, this is really cool. If this holds up to scrutiny, and has a decent context window (+16k), it suddenly changes our project's status from "cool concept, too slow and expensive to release" to "doable", just like that.
However I don't understand how are they claiming zero hallucination, how does giving confidence score fix hallucination? or am I missing something here?
I'm not sure the authors realize this is way more than "just a cool demo": if this holds up, it's going to be huge for game QA work.
Instrument your game to output properties of entities near the player and the output is the various control inputs - moment to moment gameplay gets solved. Maybe augment with a tick-by-tick controlled stepping mode if particularly twitchy - an LLM can take care of the higher level reasoning then.
The doom video is also in the article itself (headline: "Doom").
I suppose this is the same video as the one from the parent comment, but I don't know for sure - I don't have a twitter account and the above link doesn't work for me.
But when their system is given the instruction "do not fire, simply dodge" - it doesn't "simply dodge", it actually gets close to the fleshy pink demon rather than keeping its distance. Or am I misunderstanding?
This is actually pretty cool. I think the undertalked about part of this for TypeSafe is that they can always "extract"/distill the frontier of this type of task from the newest LLMs for cheap. Jev seems seems to be GPT-6-Astra/Fable 5.1 but I imagine a bunch of training data is from earlier models?
Then, you can serve it faster/cheaper than the frontier LLMs. It's basically distilling a small but extremely common use-case from LLMs and serving it. Then RLCD comes into play to update weights when a new model comes out, etc.
Any thoughts on what the next potential "cheap" win to be distilled from frontier LLMs is? I'm going to need to play around with this.
It could be used for coding if you gave it an AST.
If you work at TypeSafe please try this.
Side note: This is probably how LLMs would perform with better encoders and next-latent prediction, so eventually those will beat this architecture out. Still amazing though.
I've implemented tree-sitter in pi before, and while it works, I have no real proof it saves me tokens, or is more accurate. I think a better implementation is a model that's trained for AST's, not just "use tool, see what happens".
I'd love to do research on this when I have the time.
the hard part for coding is actually state engineering (e.g. getting your dependencies in context) - we haven't even tried it yet (because my philosophy is we should automate the easy tasks before the hard and we've been working on getting the model smart on the former)
we do think there's a lot of potential though and do want coding themed releases soon
I could see Jev being great at finding key symbols in codebase before a code generation/code review task. I sent you guys an email (to hello@) about using Jev in Code Review for www.ellipsis.dev.
I saw the CEO reply elsewhere in the comments to some other question. Maybe he can shed some light on it. My gut feeling is that this is non-trivial and they did not get this to work (yet?), otherwise I can’t come up with a good reason as to why they would not demo that as I assume half of the crowd here (myself included) would line up as customers.
Yeah it would be quite trivial to try and implement an auto regressive AST generator for STLC with Jev provided that you had bounded variable names and integers.
As you said, if it worked, they would have demoed it haha
what is the…epistemic status, for lack of a better way to put it, of the probabilities? what do they mean? what (probabilistic) guarantees do we have about, say, the responses to
- is the capital of france paris?
- it is august. is it raining in paris?
(forgive the examples; they're probably not semantically the sort of thing jev is trained to work on. but i figure the point translates to various kinds of questions that come up in "inner loop of agentic pid controller" contexts)
a normal text-generating model if asked to produce a number will also do that just fine. i assume in jev's case it was actually rled to essentially learn to express priors over things using its implicit world model, which definitely ought to help, but can we say more?
I'm trying to understand what difference does this make over LLMs.
LLMs are universal simulators, their latents model the world. So I bet if you compare their logprobs with probabilities output by this model, it will be highly correlated.
Someone should do this quick experiment. I bet there won't be enough of a meaningful difference.
This is too much for me. ML playing doom was a thing since before LLMs, decisions tree were always insanely and no one ever used then anyway, i can't see anything new in this yet everyone is treating this as a revolution. This technology was always there and quite easily accessible all along.
Seems like LLM can do everything Jev can do (just structured outputs?) but Jev is highly optimized and purpose built for it and thus way faster and cheaper. Is that a fair description?
We've already started using it for some pretty powerful decision tree stuff. We're just scratching the surface. We shipped an extension for Swamp[1] a few minutes ago and the combination is great!
The one downside is that the context window is very small (32k.) So some initial ideas we had for initial evaluation of code reviews won't fit yet in the window.
Could you use this to build a proactive memory formation and retrieval system for LLMs that runs lightning fast?
Last 32k of connect + Summary of current task: Did we learn something useful here (true/false)? What is the category to file it under? Then notify the LLM to file it away.
What class of memory might be useful here? Model gives probability to each item in the list. Short description of all memories ordered by tagged class is used in the next round. Are any of these memories useful in the current context, such that they will inform the model and help in its task (yes/no)?
I’m sure there’s some fine tuning to be had, but this sure seems like the basis for a substantially better proactive memory system that works around an existing LLM conversation.
If I’m understanding what this does and how this works (generic input, intelligent classification with probabilities, rapid and cheap), this is absolutely nuts.
Very cool! Can you explain when I would use this vs. training a standard ML model on my data? Suppose I had a fraud dataset with features like customer ID, amount, merchant, online or in-person, etc. - I can't imagine that a general model like Jev would predict this more accurately or cheaply than even a basic XGBoost model trained on my dataset (one that I could build in a few minutes by asking Codex to build it). Where does Jev add value here?
pasting it here:
zero-shot + general == programmable
I would assume any extreme scale narrow task could then be fine-tuned for, but we'll see - I suspect putting it all in shared cognitive core has bit maintainability/generalization benefits
It's very misleading. If I'm actually playing a game I don't get the coordinates of enemies sent back to me so that I can feed into my mouse to snap my crosshair to. It's looking through walls too, because it's working off structured state in text form. You could re-create this whole demo without using AI. Have an LLM generate the state machine for you and no model is required to run it.
It's a bit hastily put together, but I made a dspy fork where you can add a decorator to automatically use TypeSafe where possible on Signatures. It shows a fair bit of what actual, hands on usage looks like.
The output tokens are just responses to your inputed questions and their probability. So relatively few output tokens. No unstructured text back in the response.
We had early access and found it to be pretty useful. Having a second form of verification, where you can ask multiple questions (in the form of Nouls) raised our confidence in the outputs of other models. [0] IMHO This type of model works incredibly well in concert with LLMs, not as a replacement.
I think the doom demo uses a text representation of the world and it's basically, "projectile coming your way" -> "Strafe". "Enemy ahead" -> "shoot. So it works well when spawned in a room of enemies (as we see in the video).
If self driving is red means stop, green means go, and stay in your lane - then it would work great, but having to actually think and test which maneuver is optimal for a given situation while weighting safety, road rules, random unexpected actions and getting to your destination, I think it's a much bigger problem. A bigger model specifically trained on that maybe would do great, but then the output is not the constraint anymore.
But I haven't tried the model, so I 'm just ballparking and could be very wrong.
The model doesn't have image input capabilities (yet, it seems from the post), so for the Doom demo, a harness is extracting a bunch of structured information from the game (map layout, enemy locations, player ammo, health, etc) and providing it as a massive JSON blob to the model so it can make its decisions. This model _could_ be hooked up to make the decisions for a self-driving car, but it would need to be fed a structured blob of the situation around it, so all the computer vision problems of self-driving are still there. And that's before you get into the confidence and accuracy of this model.
It might be boring, but I can see exactly how I could use this right now to improve my agentic rag.[0] In two months I am supposed to deal with a giant corpus, while still maintaining responsive chat UX. I have been working my butt off to make our first big client happy. This could really help solve the chunk ranking problem.
[0] assuming the policies are compatible with sensitive production workloads, some time in the near future.
Haven't seen any docs or so. Is this actually a general model, or does it need training on the the data set it answers? Finding it suspicious you never see some kind of prompt.
Looks promising. I'm building an AI video editor and multi tool calls take >30s using Gemini. This would be a a game changer if Jev can take that down to single digits at p95.
I can see the value in this but looks like there's going to be trouble in communicating the difference between this and a regular LLM, and also proving the potential cost savings in using this to replace existing systems that are using LLMs with frameworks like langgraph, as this can't be a drop in replacement and would require a significant amount of re-architecting/reengineering of systems to get the type system to work
I was thinking about something similar (maybe) - generally speaking, embeddings for LLMs tend to learn real world concepts - things like 'fruit' or 'France' or 'city' as directions in embeddings.
But in things like programming, most concepts are abstract - 'if hungry eat an apple' in programming terms would look like
'if hunger > 50 {apples--; hunger-=30;}'
and compilers work with 'concept erasure' - to them, tokens (which are like llm tokens) look like
'if var1 > 50 {var2--;var1-=30}'.
They don't care about how these things map to real concepts. So all the embedding directions used to encode real-world concepts are just noise to LLMs when programming. This greatly reduces dimensionality and training costs. So does a token representation tuned for programming constructs, rather than natural language would probably have a more efficient encoding.
Current models go beyond the simple embedding because you start to encode groups of concepts in the context-aware part of the model (attention heads or any other method). So it is never simply words/tokens in isolation anymore.
If we could come up with a system to classify the probabilities across a large number of candidate words (or components thereof) then this could actually be good at producing text, one element at a time. We could call these elements 'tokens' and picking the right one could be called something like 'decoding'. Crazy idea but hear me out...
On a more serious note, it will be fascinating to see how this different spin on modelling inference will create new paradigms or slot into existing ones.
While I understand that accelerating development isn't necessarily the target for this, and it's not at all intended to generate code the way many of us are...
I think this could be pretty decent in CI? There's a lot of "flakes" I've mediated that this could have handled much more efficiently. Maybe observability as well, triggering elevated logging and other initial measures?
I think I missed why is this faster? What I’m reading here is it’s similar to constrained decoding but I’m not seeing the explanation of why it’s able to get those results.
insane doom demo
i wonder what the limits of its intelligence are? i'm guessing it's not great at reasoning tasks, it seems breaking down the problem helps significantly, but how much does a problem need to be broken down for reliable performance?
also this would be huge if it could run locally but it seems like there's no intention to do that at the moment
Hm, would be good to understand the architecture better. Is this answering just from a world model informed prior? How informed is it by the information in the prompt? I can't see this maintaining calibration across all domains and all types of structured output.
Is there anything published on how it maintains calibration? Or when you say "outputs calibrated probabilities" you mean "as calibrated as frontier LLM models, just cheaper" - which is a different claim; as LLM's aren't particularly well calibrated
Hasn't there been a lot talk about Astra's opaque reasoning capabilities (being able to think through complex questions without using a chain of thought)?
Given that, can't you just replicate Jev by telling Astra "here is the question, you must make a multiple choice decision / output a score between 1-10, please answer directly in a single word, no reasoning allowed"?
(Edit: Ok, Jev is much cheaper in input tokens so these two aren't directly comparable at all)
I like the idea of System one models but all LLMs so far work as system 1 thinking because humans generate speech subconsciously with system 1.
System 2 thinking requires consciousness which AI does not have, so even reasoning models are still system 1 thinking as system 1 in humans has reasoning with heuristics.
Its limited but most people navigate the world with it completely, so it's enough for AI.
Interesting concept, I can't see a reason to use a generalist classifier over an api rather then just training my own? If it was open weights I would probably mess around with it.
You know you're too old when you see the company name and think! Oh I wonder what Martin Odeskey , Jonas Bonér and co are up to. Wait, didn't they become lightbend... Altho this comment takes away from what these guys are doing which legitimately sounds interesting.
You could theoretically ask “what is the next appropriate character?” and add the entire ascii charset but i doubt it’d work well and you’d be implementing autoregressive churn across network latency…
strings (and all sequential data structures) are not allowed at all - this is how we make sure all outputs can be computed in parallel (thus no output token cost)
This makes me think of Expressions of Change [1], a project that aimed to make updates to a program a first-class primitive in a programming language. A model like this can't output code directly, but perhaps it would be well suited to select from the small set of discrete operations on code envisioned by the EoC author?
Thanks for the early access! I was testing the Lisp idea out in the playground, but I don't think the model is smart enough right now to generate actual code. I tried having Jev finish generating the code for a Fibonacci number function, but it kept wanting to create a literal number instead of refer to a variable which is a number. This happened both when I gave Jev the current program as a string and when I gave Jev the program as structured data.
Maybe I'm just not doing a very good job at prompting Jev, but I think right now it's not quite capable enough to generate Lisp code.
Huh this looks fantastic. The Doom demo really sold for me that this could be a great tool for accelerating QA at my gamedev studio. Signed up for early access.
This is basically a zero-shot classifier that can accept raw text (or structured text) as an input, and is able to classify that text as accurately (they claim) as a frontier-level LLM. I have workflows this would be useful for, looking forward to it showing up on OpenRouter.
Funny how the authors are asserting that "doing the right task > data > compute > algorithms" while simultaneously releasing AI model for calibrated decision making, which if they work, would mean that "compute > doing the right task"
Congrats on the launch! What's different between Jev and Microsoft's Guidance package? https://github.com/guidance-ai/guidance Is it a diffusion generator under the hood?
Signed up for the beta! :) would love to put this through some real-world shootouts against traditional LLMs to see where this type of model really excels.
I’m guessing it might be able to replace maybe 40-70% of LLM calls for a given pipeline depending on the business task, cutting the API costs on those calls by an order of magnitude.
I am positive I know exactly how this works, I made something similar a few months back. But the problem is without generation you are extremely limited in the use cases. And while the model can't hallucinate, it can still be wrong. It just can't make up data.
Yeah saying it can't hallucinate is crazy. It can still forward a billing query to the dev department incorrectly. It can still get an obvious yes/no question completely wrong
Last year I also had a rather similar idea, but dropped it before I went very far in working on it. I wonder if you and I had similar ideas?
1. Start with an LLM, so that your model understands natural language.
2. Replace RoPE with a tree embedding scheme, and causal attention with a sparse attention on the graph structure. (You could use full attention... but it's cheaper to use graph attention.)
3. Chop off the final unembedding layer, replacing it with a projection down to two scalars, one for logits and one for confidence.
4. Each option of a choice is represented by a number of tokens in leaf position; average these tokens' logit outputs to get the option's logit. Average all of the confidences from all of the options to get the choice's confidence.
5. Train the logits by KL divergence from a true distribution (or NLL on samples from a true distribution).
6. Train the confidences on a subset of the data in which you know the entire true distribution.
The hardest part is getting real world data for workflows, but I wildly speculate that you can get by with only ~50,000 documents if you first adapt domains using synthetic data.
So an encoder-only model with a classifier trained on the heads or something? DeepSeek recently switched to an encoder-decoder architecture in an attempt to get the best of both worlds (fast prefill while preserving generation capability), I wonder if that might be the future?
That was the first thing that come into my head. OK I can train very simple model, that can generate json's for specific tasks, so what?
How we can be sure that this "limited use cases" not just overfitting for particular outputs (or even distillation?)
> We deliberately chose not to publish performance against public benchmarks. In fact, we plan to only have one-off evals when we make product updates.
lol, I bet they would publish them if their score on those benchmarks were good.
Finetuning a language model for decision classification (with probabilities) is already well-understood. What specifically changes in the training objective with RLCD? Are its benefits isolated from Jev’s new architecture/parallelism?
Will need hands on to truly tell, but the doom demo seems very promising. If it can play that with text descriptions of where stuff is by distance and degrees in a 3D context then many GUI automation tasks should be easily doable
That's kinda the goal. Imagine all the automation in the world being able to embed intelligence directly inside it - factories could route based on more complicated questions, hardware could anticipate your needs. Customer support could be done without humans 90% of the time.
The whole page reads like it was vibe-written by an AI. If I'd built something as disruptive as this claims to be, I'd have spent at least fifteen minutes writing the announcement myself.
Every time I see 'we' in an announcement like this, I picture one guy alone in his basement.
If it work as good as they say it does, confidence score + really fast response when you want very fast response, basically..
To me it is a crime against humanity to not open source it.
Just get the money from cloud inference and cloud agentic sessions or whatever but open source it.
This tech, a good harness, a good model provider, and you have basically a AGI building machine.
But also effectively this is a classification model. It excels at specific certain types of workloads, and obviously will fail at others. Not really sure how one benchmarks this tbf. I can see their argument on why this requires a novel specific eval for whatever your usecase is. A consistent "global" benchmark might be hard to do
The technology and the results are very handwavy. What is RLCD exactly ? What are scores on benchmarks compared to LLMs ?
This website does not inspire confidence at all, it all sounds like a marketing piece. I wish it was true, some kind of text-prompted classifier with LLM performance would be cool, but I can't trust it with what we are given.
Can HN have a tag for open-weight vs closed-source models please? The progress is nice, but if it is not released at least in papers or open-weight? These are just ads?
Okay, so it doesn't output text, that much is understood. What are the inputs like? I'm assuming maybe a text input? maybe an AST definition? Really hard to tell how this works at all from the demos, especially since we can't really try it out.
Why did they pick the name System One? It's not really explained what "System One tasks" and "System One shaped queries" are. Things that need a fast response?
Does this imply it's a very small model? I couldn't find anything about the model itself.
> we assume there is a correct compute graph (a “workflow” represented in code) and use the predictions of the largest, smartest, and most expensive external models as reference probabilities.
...
Rephrased: every model gets the same workflow. We test how they compare to the average of the smartest models (in this case, Astra and Fable).
They assume there is a correct graph, but they don't compare to that, they compare to the average of the smarts models? So the smartest models are getting it wrong but you compare that anyway as a benchmark? So the outcome is "how much of a Fable am I getting" etc. Why not compare the actually correct thing?
But then even on this hand constructed eval, the first plot is showing Jev at less than Sonnet 5 accuracy. It is barely better than Luna. There are two Opus 5's and two Sonnet 5's without explanation. What is the plot showing?
Super intrigued by this - large scale automation using LLMs is quite annoying due to deprecation cycles of models from frontier labs and cost of running your own being prohibitive when you have a blend of them.
Not in the traditional sense of a coding agent, but we think there's a ton of opportunity in using it for context management ("do we _really_ need to pass all these tokens to the agent?"), semantic linting ("how does this score against this AGENTS.md: <...>"), etc.
If you zoom in (especially on the large title), you'll see that the text is a semi-transparent gray with a black internal outline. It seems like all the typography is SVG-rendered. Actually insane. I've never seen this before. Not even the most vibeslopped websites have that.
this might finally be smart enough and fast enough for jarvis. hard to feel like iron man when your assistant takes 8 seconds to decide to pause your music
Is there a bottleneck which would hinder putting this architecture in charge of a humanoid? Would it be able to operate continuously, for example in conjunction with an LLM for long-term reasoning? Doom seemingly works extremely well.
Parallel inference where you don't want a subagent seems niche. But there is a lot of random things where businesses ultimately want some kind of score instead of generating something.
I think the interesting thing would be seeing if prompt injections still work with this kind of model.
we have played with this! the fascinating thing we've found so far is that adversarial examples for our model are quite different from that of LLMs so that they work even better together
> LLMS
> Strings / generated text. Strings are flexible and can be anything: chat responses, code, hallucinations, refusals, or even type-safe structured values. To be used by software, responses need to be parsed + validated. There is also always some risk that the AI goes off the rails.
> Jev
> Type-safe structured values. Possible outputs and structure are defined in advance. The model never makes type errors. All answers are accompanied with calibrated probabilities and confidence scores.
I mean, this isn't even remotely comparable to LLMs so why compare? Also, why are they bringing up AGI given there approach is so restrictive that what they're building literally cannot have the creativity required for AGI? The video is 100% marketing slop...
The bulk of the application of LLMs is that they generate reasonably reliable text which doesn't need to be defined in advanced. I'm sure there is a niche for this and congrats to the team, but please let's not hype this as if it's the next big thing in AI...
It looks like a specialized encoder-only(-ish) transformer with scalar and ordinal output heads. Acausal in effect, maybe? Probably not even autoregressive?
I'd use this as a tool an LLM can use for specialized tasks. It's not AI in itself.
First, congrats to the team on launching something genuinely interesting and new.
Seems like a more accurate title would be "Jev: Trading general purpose generation for fast typed inference" or something like that.
This is interesting, but the speed comparison seems misleading? A generative model that can output code in a Turing-complete language can do anything a computer can do.
Jev can only generate structured output, right? This is probably super useful for classification/routing/scoring, but it's nothing like the code generating models we're all using today for code and automation.
Also "can't hallucinate" seems wrong? Sure, it can't emit an invalid type, but it can still emit a completely wrong valid value. You can enforce structured output from an LLM too, with an appropriate harness, etc.
Assuming there's no funny business, the Doom demo is cool.
> Assuming there's no funny business, the Doom demo is cool.
The Doom demo seems very funny business. They're not feeding it video, they're feeding it a text description of what's going on in the game. It's not reading pixel data.
I think LLMs would play a lot better with that input too but Jev does seem to have a huge speed advantage; I don't know if the other models could do that in real-time.
I think the meaning of can't hallucinate in this model is that the type won't be hallucinated.
So if the generated schema is for a tool call for calculator, then the numbers will be valid numbers for sure (and not random words).
To me, it looks similar to BNF schema already introduced and implemented few years ago: generally speaking - it limits the next token that is allowed to be generated, probs are drawn from a subset tokens.
(tbh, I'm not sure why it didn't pick up as a more standard interface to LLMs, as it made a lot of sense back then, and now.)
When they say "can't hallucinate" they mean they produce a confidence value for every result, so you could see for example it has 0.1 confidence, and you can disregard the result - that'd be different from hallucinating where it believes it's correct
if it puts a high confidence value on a wrong answer, thats still hallucinating, no?
llm hallucinations are high probability tokens that are incorrect vs the real world
Correct, they have not made a universal all-knowing omniscient oracle, which is what would be required for "can't hallucinate".
That seems like a weird standard.
I would be happy enough with: only produces what it can verify with sources.
If you eg try to remember a court case (ie produce the reference via LLM token generation only), it's easy enough to check with your data whether it really exists. Similar for following links and other references.
If your data or sources are wrong, obviously your report about them will be wrong. But I wouldn't call that a hallucination.
There isn’t a single human in this world and hasn’t ever been that meets your happy-enough standard. Make of it what you will.
Why is that at all relevant?
Humans are known to hallucinate a lot. Ask 10 different witnesses at a crime scene what they saw and they'll all report different things.
A good, non-hallucinating LLM would only report things for which it has evidence. It would consult the facts every single time.
It's a pain in the butt for humans to fact-check everything but LLMs can quickly look up all kinds of stuff. That's what makes them useful.
It's not a binary thing. You can get closer or further away from that standard.
And humans also behave differently in different contexts. A conversation at the pub has more such hallucinations than a formal deposit in court. For the latter, a good lawyer will look at her shoes, when you ask him what colour her laces are.
Not to be tooo pedantic, but a bot that assigned 0 confidence to everything wouldn’t hallucinate.
A calculator either gets the right answer or doesn’t answer.
It wouldn’t have to be all knowing as long as it knew perfectly what it doesn’t know
A quantum calculator answers in distributions.
Yes, there is no magic sauce here that makes stochastic output binary if that’s what people are looking for.
Right and so maybe we should stop saying "can't hallucinate" when it can by definition.
It’s not what people are looking for, but what they wrongly claim.
Technically speaking when you send the prefix “The capital of France is “ into an LLM it will also produce probabilities across its whole vocabulary.
The probability values don’t really represent confidence in modern LLMs though, especially after RLHF and RLVR.
System One says they use RLCD, Reinforcement Learning for Calibrated Decisions, which presumably has accurate probabilities as an explicit optimisation goal.
How is that different from RLVR?
RLVR generally upweights tokens along the whole thinking trace that led to a correct answer, whether each token was "correct" or not. RLVR doesn't train a model to output an 80% likelihood, it just trains it to produce correct answers, and not to produce incorrect ones.
System One hasn't said how RLCD works, but they do say it is explicitly training models to output "calibrated" probabilities, which makes it distinct from RLVR. This is how they describe it:
> System One models are trained for calibrated decisions: their probabilities are optimized against outcomes to reflect uncertainty.
… which they could provide in their APIs but are vehemently opposed to because it makes distillation much easier, and faster.
Yeah. Yet another reason why open-weight models are better. If I want to use the logits, I can.
I read "hallucinations" as "generates novel output with no grounding/source". i.e. "it just made something completely up".
I believe their "accuracy" metric (sonnet 5 level) is where "right/wrong" is measured.
that's right, but because these models are probabilistic, it's also possible to be confidently wrong (and all future models will be smarter still and still have that possibility)
Yeah but what stops it from producing confidently incorrect outputs...
Nothing, but imagine using LLMs for a classification task
People out there are so resigned to the models being unreliable that they are really doing things like hallucinating deliberately, and then matching the hallucinations to embeddings -
https://softwaredoug.com/blog/2026/08/10/hypothetical-classi...
You could do that or you could just... use a model that will never produce unreliable outputs in the first place.
But we're going from "Apple" to "Apple: 99% - trust me". It could still be an image of an orange :)
It's pretty darn smart. If you did want to hack on it in earnest and find out for yourself, send me an email - nathan@typesafe.ai
I'm certainly not resigned to that, at least for classification.
Even non-frontier models are absurdly good at this in a broad sense.
Which would make it hard to judge "a model that will never produce unreliable outputs in the first place" against something that is already really, really good and exceptional in domain-specific areas with the tiniest amount of elbow grease.
Speed and cost look good though (for now)!
What about the LLM calls though that are done midchain? In the Home Assistant video the multi-intent prompt gets split using what looks like a traditional llm model, which I'm assuming is vulnerable to classical hallucinations.
That's really funny when you consider that generative models also don't hallucinate if you check up on them on every token generated?
Has LLM become so synonymous with Generative Transformer that other high-parameter count models that interpret language need a different name?
For all we know this might be a non-language-generative transformer e.g. a transformer where the decoder produces confidence scores rather than language. Please provide more likely architectures if you know them, I'm genuinely curious.
I don’t think it’s misleading if you compare on the use cases they suggested. It’s faster and cheaper (no idea if higher quality), so it’s immediately interesting for certain things.
And if you buy their RLCD claims, this might be even better than huge models that know a bunch of irrelevant things.
What was misleading was the original title:
"Jev: New frontier model 40-400x cheaper and 20-200x faster"
I'm not the gatekeeper of who gets to call themselves a frontier model, but I don't think most people would count Jev in that group. It sounds false.
If their specific claims hold up, then it would make more sense to say something like:
"Advanced the speed/cost frontier for structured decisions"
I dunno, I would consider Waymo and Tesla to have frontier models.
I think AlphaFold and related are also frontier models.
Being an LLM does not seem like the qualifier for frontier.
This is likely still an LLM (in the purest definition of a language model with relatively many parameters) since the inputs are natural language, just not a generative LLM as the output is something other than more language.
The inputs aren't natural language. https://docs.typesafe.ai/primitives
The inputs are natural language, they're just also structured into a tree. The first example on that very page shows natural language instructions:
Large language models are not the only type of model.
How is this not a frontier model? It's bleeding edge in its own niche. It's not a frontier LLM; however, applicable to many of the things people use LLMs for.
It's nothing like a traditional LLM and so should not be compared to one. It's a heavily constrained, tiny model that can only produce a probability score or a yes/no answer over pre-defined selections. It has no long-context capacity.
I mean, imagine comparing this thing to Astra, it's hilarious. They don't even tell you what the max input size is, and they only allow 10 possible answers to choose from for the Choice mode. It's probably like a 1billion param model. They say it's "not small", but there's zero reason to believe that.
I suspect someone will be able to recreate this within a week by piecing together open-weight models.
> It's nothing like a traditional LLM and so should not be compared to one.
Frontier LLMs are expensive jack of all trades. You can absolutely compare them to purpose-built tools on any domain they touch. Engineering is all about assessing tradeoffs.
Well, not quite a week: https://x.com/harshagundal/status/2100044305536889015 - apparently it took him 2 hours.
nali shwana
I'm biased but I wouldn't call it misleading - generating text is super awesome and flexible, (we describe that in the blog post - and I personally use string models all the time) but it's true you pay a high tax for autoregressive generation
> Also "can't hallucinate" seems wrong? Sure, it can't emit an invalid type, but it can still emit a completely wrong valid value.
that is likely true of all ML! perhaps we could debate semantics, but I don't think it's fair to say a random forest "hallucinates" in the way LLMs do
His claim was that the title is misleading, not sure how it's relevant to that claim that you use "string models" (full LLMs).
The original title before it changed less than an hour ago was:
"Jev: New frontier model 40-400x cheaper and 20-200x faster"
I'm going to agree that was misleading.
And on the second point:
>>Also "can't hallucinate" seems wrong? Sure, it can't emit an invalid type, but it can still emit a completely wrong valid value.
>that is likely true of all ML! perhaps we could debate semantics, but I don't think it's fair to say a random forest "hallucinates" in the way LLMs do"
Also going to disagree here, and I don't think it's semantics.
Type safety is not factual correctness.
> Type safety is not factual correctness.
I very much agree with this and want to hone in on where do actually disagree. Would you say a linear classifier hallucinates?
> I very much agree with this
From the intro blog
"Hallucination and type-safety are intrinsically related"
I'm not entirely sure why we're conflating type safety with, I guess, value or output safety.
"Would you say a linear classifier hallucinates?"
No, but it can be (and often is) mathematically correct and functionally incorrect. It doesn't help to say "a linear classifier can't hallucinate" when you get even 99% accuracy. That's 100% a semantic play, and it doesn't help when the picture of a dog is labeled cat and the response is "yeah but that's not a hallucination, only stupid LLMs do that"
Hallucinations were defined in the context of text generation models so your question does not really make sense.
IMO your system can make mistakes that are similar in spirit to hallucination (i.e. answering with a false answer instead of abstaining to answer).
And furthermore, because the model is forced to answer in a boolean (if in boolean mode), if the user input is outside of the range of a boolean, it's forced to hallucinate. It can't abstain.
User input: "Hey, have your human support agent call me, tomorrow at 5pm."
Model input: "Does the user want to speak to a human support agent?"
Output: Yes.
I imagine that your model would produce this, and I think it's fair to say this is a hallucination. A human would caveat it with: "Yes, but not right now.", your model is incapable of that. Yes is technically correct, but within the context of being in a live chat, a human would understand that the caveat is required.
To be fair - you’re crafting a deliberately bad model input for a contrived example.
No. Your launch post puts “0%” on a hallucination chart, then explains that the number comes from guaranteed schema matching.
You’ve already agreed that this doesn’t establish correctness. An approve for an unauthorized action still meets the schema guarantee.
That’s why I find the messaging misleading. You’re acknowledging the limitations in these replies while defending the broader reliability pitch.
Even granting that each answer is calibrated individually, that doesn’t establish calibration of the decision that combines them.
Sure, I can threshold a composite score, but there may be many wrong answers with the same score. An unauthorized action doesn’t become acceptable because it scores highly on the other dimensions.
I still have to define the constraints and test which wrong actions get through the complete workflow on my own data. That’s a substantial part of the work being pushed back onto the developer.
hallucinations are not wrong answers, that's why we use a different term
id say yes. a linear classifier that classifies between red and yellow balls will hallucinate on blue.
linear regressions hallucinate in the simpson's paradox.
the model output can be quite confident and not representative of reality
Let's say classifiers don't hallucinate. To make a fair comparison we should constrain LLMs to the same classification task. In that case, no, LLMs also don't hallucinate.
- Give Jev and LLM the same input
- Lock down both to approved/rejected/unknown (LLM restricts on decoding)
- Both can be wrong, but neither can hallucinate (invent an another option).
A hallucination in the context of LLMs is generally understood as an incorrect answer presented as factual. If you claim that "x can't hallucinate" in the context of LLMs, you're saying that x always gives accurate answers. It does not matter whether the answer is type safe. If its value is incorrect, it's a hallucination.
Just to be sure that I understand, you're saying that your model "can't hallucinate" because it only outputs a single thing, right? In this way, an LLM can't hallucinate either if I prompt it to do a classification task with a discrete set of possible outputs, right? (Assuming I reject non-conforming output. Actually, maybe what you're saying is that your system can't output non-conforming output?)
Yeah that's precisely correct.
For e.g. classification tasks, even in 2026 people are doing things like hallucinating deliberately, and then matching the hallucinations to embeddings -
https://softwaredoug.com/blog/2026/08/10/hypothetical-classi...
With TypeSafe it just picks the class (actually probabilities across classes), reliably every single time.
404?
https://softwaredoug.com/blog/2026/08/10/hypothetical-classi...
From a quick look at this it looks like it could easily generate natural language text by following a structured representation like UMR (Uniform Meaning Representation) or the similar representation the Abstract-Wikipedia folks will be working on for generic encyclopedic text (which will be heavily informed by Universal Dependencies). These are basically linguistically principled and frame-based counterparts to a programming language AST, that can be then converted to natural language (in a broadly language-independent way, to the extent that semantics and pragmatics make that feasible) via some sort of NLG rendering.
(To be clear, this one raw model does not support outputing a full AST directly - it wants to output "choice" among fixed options, "score" on a sliding scale, or a true/false answer (all of these with confidence scores attached), so building the AST/structure would be a code-driven (or even perhaps outside LLM-driven in some more challenging cases) multi-step affair where the model would essentially be playing a "game" of building the structured output step by step and getting a revised partial state back. But one could expect this to lead to interesting results.)
> I'm biased but I wouldn't call it misleading
- @CompleteSkeptic
Very strange.
> Jev can only generate structured output, right? This is probably super useful for classification/routing/scoring,
My first thought was that it would be ideal for robotics? As in control of limbs, general planning, route finding, etc.
That's a great point! It quite looks like the System 1 model of Physical Intelligence
Um, Isn't SELF DRIVING the elephant in the room?
Only if you think that everyone cares about self-driving. Lots of niches require structured domains; self-driving is just one that has a lot of capital thrown at it.
I feel like the power of the approach presented here is that it gives a model a proper "language" to describe computations directly vs moving tape silliness.
I foresee this to be the path moving forward - giving AI models understanding of the computation directly(as well as compositional rules) This feels like a short path towards total software in many areas.
Agreed. It's a wildly dishonest presentation of their product from many perspectives, which is a shame because it might actually have some good use cases.
The comparison between LLM speed and Jev speed is misleading, because they're using autoregression to generate all of the type names, all of the schema, etc. A closer comparison would be if the LLM was purely outputting the raw numbers. Even then, comparisons to LLMs are pointless because you could train a transformer on the same sort of task that Jev is doing and get even better performance yet again, and a smaller model. I suspect this is some form of stripped down diffusion language model.
You really have to do a lot of hand holding here, and map out your problem space manually, and very carefully, to get any sort of accuracy. For example:
> Keep each Score to one dimension. If a description says “punctual and smart and experienced”, the question is measuring three things, and an input that is high on one and low on another can’t be placed. Confidence drops and the score means less. Split it into one Score per thing and combine them in code
If you don't perfectly represent the distributions of possible answers then you'll likely get garbage results. As far as probabilistic state machines are concerned, I'd say creating the distributions of possible answers, and their hierarchy, is the actual hard part.
One of their examples is:
- "state": "I have asked three times now. Can I please just talk to a real person?"
- "Is the customer asking for a human agent?"
Imagine the users request is: "I want your human agent to call me tomorrow at 5pm."
Human conversation is fuzzy, getting useful reliable results out of this is going to be a challenge. Of course, you could add follow up checks like: "Do they want that now, or later?" -> if later -> "Do they want that tomorrow, or the day after?" and so on... But now you're building an LLM out of if statements. I am skeptical of whether this model has much utility for fluid language interpretation - I suspect it'll only be useful for scenarios where you've tightly constrained the answer space but want to use fuzzy language to describe it. Like:
- Question to human: "Would you like a support agent RIGHT NOW?"
- Their response: Yes | Yeah | Mhmm | ye sure (any possible yes signal)
Model input: "Did they ask for a support agent?"
Still... a tiny LLM could accomplish this sort of thing without problem. And that doesn't stop someone from saying: "No, not right now. But tomorrow." - and the tomorrow would get missed. I think this is why people haven't really tried this approach much already.
Also their Doom demo is on structured state, not on images. Meaning, the enemies must be being served to the model as coordinates (or the exact angle of projectiles that hit the player), otherwise it'd have to scan every pixel of the 360 degrees to know whether an enemy is in front of the crosshair or not. You can see from the map below that it's also choosing travel checkpoints/destinations through walls. So they've severely cooked this to make it look far more capable than it is in practice, and any speed advantage that is offered here is not factoring in the shortcuts it is taking, the training on the map, and the fact that it can cheat because the structured state it is using is not bound by obstructions.
Here is their docs by the way: https://docs.typesafe.ai/ - so you can understand how it works.
Always exciting to see people working on novel models, rather than the Nth version of the same slightly tweaked LLM.
I'm very curious how much ressources are needed to run such a model. This could be a complete game changer for local applications.
Wasn't really till seeing this home assistant demo they have (https://www.loom.com/share/18c4dbcf8db546dfb2d7f2ef018e78e4) that the value really clicked for me.
Seems really cool.
That's good.
Side note - just like most people don't need an intelligent personal assistant to manage and respond their emails and book their flights, most people also don't need smart homes. Century old toggle switches are more than enough in a 3 room apartment or 5 room house unless you have a mention.
My primary beef with smart home (having tried it) is that every person that visits your home ends up confused about some element of it. A light switch that goes up and down is universally understood.
My smarthome has regular switches and wifi.
There's no reason to not do both.
Also a quick NFC sticker in each room taking you to a small HTML site containing settings (temp, ventilation, lights, shutters, setting a alarm by the lights) has been golden.
No one will install: Shelly, AC app, look for ventilation IR controller or figure out how to use the Shelly app to turn lights into an alarm.
But tapping your phone on a NFC sticker bringing all those controls together per room in stead of per category (all lights in Shelly app. Person in room #1 has no interests in the lights in room #4 at the same time.).
One "all house" sticker next to the front door allows any last person leaving or first person entering to put the entire house in active / idle mode.
Works wonders. And as soon as local AI is quick enough the stickers will be a microphone!
Need is a big word. Convenience is also a factor.
We have 8 light buttons in our living room/dining room/kitchen space. It is very convenient to us that we have 1 button for turning all of them on/off at the door to upstairs (at night turn off all lights and go to bed upstairs, in the morning come downstairs and turn on all lights) - but also have 1 on/off button near our back door for when we leave/come home.
Next to that: on/off toggles a schedule where the lights are bright and cold-ish by day, and low and warm by night without us having to manually adjust each light every hour or something.
Again, need is a big word. But it's very convenient and pleasant.
This is very cool. However I don’t really want to bounce all my home automation commands to the cloud. I hope there will be an open weights approach one day. I’ve spent a lot of time setting up my local only home automation system, it would suck if it didn’t work during an internet outage, and also there are obvious privacy problems.
GLiClass is performant, and its zero-shot classification scores are in the same ballpark as the Terra-level results Jev points to.
https://github.com/knowledgator/gliclass
Thanks!
Rereading some things and because there's no official benchmarks, I misspoke about the ballpark comparison., but the open model's still a useful foundation to work with
This video makes a better job at explaining what it is about vs the marketing ones. Thanks for sharing.
Guess I'm a bit less impressed seeing that for some of the more intelligent driven+action work -- splitting requests in the video -- they had to kick out to an anthropic model.
That's fair, but it highlights how this would actually be used. It doesn't really seem like a competitor to other models but instead a way to make these real systems more enjoyable to deal with.
Haiku, to rewrite a sentence as two discreet commands.
I agree that it was notable that they delegated to an existing LLM, but I don't think it detracts much from the value proposition (not yet proven) of their demo.
Agreed but is it much easier to deal with if you need to have all of these sub processes integrated? How does one know when you need to reword a request? What if Anthropic then has a type error, then debugging that just got harder.
That really helped figuring out what this thing does, thank you.
This, combined with contracts, could make a lot of things so much fun now!
For those who don't know (which is probably everyone but me), I ported the design-by-contract pattern in Python and combined it with LLMs. This was early 2025. I originally wrote about it here: https://leoveanu.com/2025-03-01-dbc/ . Contracts are a core feature of SymbolicAI ever since. The community seems to have loved it too (https://news.ycombinator.com/item?id=44399234).
I think I'm starting to glimpse the implications and it's gonna change agentic workloads if it holds up to scrutiny. It's too early for me to tell anything other than jot down some rough thoughts.
In short, you get blazingly fast semantic branching you can use in control flows. For contracts, I can now directly take the data model that you have to design and convert it into Jev's expected format. Or I can use Jev for semantic branching in postconditions.
If my understanding is correct, that should be doable, but I need to think more about it. It could be that with Jev I can finally “compile contracts” and better chain them into workflows, which is something I always wanted but didn't know how to do properly.
Eager to test. On the waiting list.
love it. send me an email and i'll try to get you moved up on the list? nathan@typesafe.ai
This is a very promising idea - a model that takes arbitrary text input (which can be a complex json), plus a set of questions (yes/no, multiple-choice, or score) and quickly (milliseconds) and cheaply ($0.042/MTok) answers those questions.
Unfortunately, none of this is explained in the announcement, but the documentation [0] is pretty good.
[0]: https://docs.typesafe.ai/concepts/how-to-build-with-system-o...
API example[0] makes it clear how it'd be used:
[0]: https://docs.typesafe.ai/sdk/pythonI am not an expert in this domain but as an engineer-turned-researcher, this looks a lot like GliNER with a fitting harness.
This is something I focus on in a bunch of my experiments - how to get immense value out of tiny models (<1b params). There are lots of different architectures out there and there is so much to optimize if you know what you are asking and have a grammar to constrain with.
Great to see this and I hope this is a lot on top of what is already openly available.
Looking at the example Jev use cases, it almost feels like Jev's incredible cost/task can make it competitive as a generalized "poor man's ranking" algorithm that can be useful for lean startups or any fast paced development org.
I need to rank 1000 articles and pick the 5 most relevant for the user? Jev.
I need to audit and strip out content because my user is affected by regional privacy laws (without hallucinating)? Jev.
I need to surface the 3 funniest media comments that match the user's sense of humour? Jev.
More like:
I need to ...? -> Open-weight model.
I'm sure someones working on this as we speak using an open-weight LLM base (Qwen or something would be a perfect fit).
This sort of task is a perfect fit for a very small model capable of semantic parsing. You can get away with a LOT less parameters without all the autoregressive generation and long-context reasoning.
you don't say - https://huggingface.co/harshatheg/Qwen-2.5-1B-RLCD
Crazy, looks like this was just published a few hours after the TypeSafe post!
Wonder if this could lead to better recommendation algorithms.
It seems like the docs[0] are a better explanation? The comparison to llm tokens is kinda confusing.
It looks like the model takes as input a state (structured text? not sure if multi-modal) and a question (as a "Choice", "Score", or "Noul") with some additional augmentations possible. Then outputs the question's answers as appropriate (e.g. a choice, accompanying probabilities, confidence).
Edit: On the AI primer page, it looks like they do the RLCD on a pre-trained base model?
[0]:https://docs.typesafe.ai/concepts/system-one
CEO here - that is right!
I do agree that the comparison to LLM tokens is hard to understand (also because output tokens are not comparable).
But yes, text or structured state (like a JSON with multiple pieces of text in) -> decisions out (e.g. choice maps to "match" statement, "score" maps to sorting, "noul" short for bernoulli maps to if-statements)
Small request, can we get an explanation of the naming of "noul" in the docs[0]. I tried googling, and searching the docs and didn't understand why it was called that.
(I'd also argue something like p_yes or just probability might be a simpler name, but I'm sure there's a better reason behind Bernoulli maps).
[0] https://docs.typesafe.ai/primitives/noul#noul
here is how I attempted to explain it to my company's AI group chat, is this roughly accurate?
"instead of autoregressive string output it instead outputs structured type-safe 'decisions' with probabilities/confidence scores, each generated in parallel
so sort of more like a Large Classification Model than a Large Language Model? or, maybe better to think of it as a sort of "shift left" in the LLM's transformer architecture, allowing you to replace the predefined token vocabulary of an LLM with a prescribed set of 'decisions' that need to be made based off the input context; and exposing those probabilities directly so they can be integrated into the system logic, instead of just sampling from top-K.
all of this while still being instruction-tuned (!!!)"
It's always been possible to build classification pipelines using LLM embeddings as the input. seems like this is a much more sophisticated / useful application of that concept
very accurate!
the one nuance I'd get into is I'd call it "zero-shot" over "instruction-tuned" (the latter often implies a particular distribution), but very safe for sharing
Hi - first congratulations, System One looks really promising.
The Doom demo really help me, at least, to understand how System One differs from LLMs. However the first demo (Side-by-side demonstration) - I'm struggling to understand what is going on here!
I was confused at first too, but it makes more sense when you read about their primitives. E.g. https://docs.typesafe.ai/primitives/noul
The demo is showing System One producing its output in parallel very quickly and for little cost compared to an LLM generating its answers token-by-token. The "noul" type is used to evaluate a yes/no question and return the probability that the answer is yes.
So this demo is showing System One offering much more nuanced responses and specific probabilities compared to an LLM's more crude responses (e.g. LLM shows "true" or "false" compared to "0.9" or "0.07" probabilities that the answer to some question is true).
I think that's to demonstrate its speed
I see this super interestingly as the "subconscious" to the llms "conscious" for lack of better terms. I'm super interested in this for broad and rapid decision making in the context of consumer agents so will be signing up for sure.
1. I am extremely on the same page 2. I do think that subconscious is not only much smarter than we give it credit for, but also much more robust than the "jagged frontier" of current LLMs
(shilling my blog post on that jaggedness: https://www.completeskeptic.com/p/lies-damned-lies-and-bench...)
For many day-to-day computing use cases, Jev seems far better suited than an autoregressive language model, if for no other reason than it is not wasting compute thinking about anything other than how to spit out a decision.
Do you have an architectural explainer yet for Jev or are you holding that close to your chest and letting the magic rip for now?
architecture is close to the chest for now, but we have talked about writing a paper
I don't want to shill my blog too much, but I will say data is probably far most interesting than architecture: https://www.completeskeptic.com/p/the-bitterest-lesson
It's unclear if the context extends as the conversation grows?
> the model takes as input a state (structured text? not sure if multi-modal)
Input, and criteria/instructions can both be defined as structured input (JSON). This ends up being pretty powerful because the model is trained to understand structure.
e.g.: https://docs.typesafe.ai/primitives/advanced#structured-inst...
> not sure if multi-modal
just JSON... for now :)
> outputs the question's answers as appropriate
correct!
I assume this isn't really for consumers/individuals currently? Kinda feels like an improved magic 8 ball.
I can't really intuit how I should think about when the model will be accurate. Is there somewhere to read more about that? I assume customers would just have some tests or talk to you.
> I assume this isn't really for consumers/individuals currently?
Unless they're hackers, no. It's not really a chat interface, it's meant for consumption by machines and composing into higher level systems (pairs great with LLMs).
> Is there somewhere to read more about that? I assume customers would just have some tests or talk to you.
We're going to release some more info on evaluations over time, and yeah, join the waitlist! We offer faster access in exchange for good memes
After much fumbling around with prompts and evals, this is exactly how I am using LLMs in production, to narrowly make choices and return structured data. Any deterministic work gets pulled out of the prompt and my goal is to narrow the model output to be as clearly defined and as minimal as possible.
Jev's focus on structured I/O and confidence scores are game changing. If this does at all what it claims, I think this is going to quickly become the new standard approach for agentic systems.
we hope so! the bigger hope is to not just eat LLM market share, but to allow for people to use AI much more in the inner loop of software
I'm sure you've thought of self-driving. How does the model work in that space?
Great question! Yes, this works much like the doom player. Sensor data (LIDAR, velocity, etc.) becomes the state. You use the score primitive to operate the controls ("What level of braking should be applied" 0: None, 1: just slightly slowing down, 2: there's a suspicious cat on the side of the road you don't trust, ...
Full disclosure, I am not they :=)
But the real problem in self driving isn't the decision making but object description. That is, computer vision if with cameras.
Decision making isn't that of a bottleneck I suppose.
This smells like a tool a more broadly capable LLM would take advantage of extremely well.
Curious what your use case is if not confidential.
Not confidential, but not super relevant, as this is something I have learned the hard way over the past year across various projects.
A lot of people have become prompt maximalists, asking for complex multi-part solutions or dynamic workflows in a single prompt. You can get this to work sort of reliably with frontier models, but without much confidence or clarity where things might break in practice. My goal is to strip out as much determinism as possible from prompts so the LLM only needs to handle a narrow, well-informed decision, like "Pick one of these three things" and build around the answer. Sometimes you need to fill out a whole JSON payload and LLMs really actually suck at manipulating and adhering to JSON. They do ok now because labs have put in a ton of effort on making harnesses play nice with structured data. But it comes at a high token and context cost because under the hood I suspect the model is churning invalid text repeatedly until it gets around to passing some internal validation.
Example I have worked: Personal delivery app, that tracks packages from various senders using incoming emails.
I am using the single prompt approach with GPT5.4, which is free, but it’s not reliable. Using Jev I’d decompose the prompt into a bunch of smaller questions, then I’d combine the answers in software. I’m super excited to try Jev out.
So in theory you could feed it incomplete text, and then ask it for the probabilities of what the next character could be?
I think the joke here is getting missed
Strings trigger us
Or a partially completed song, asking for the next note. I’m not sure if you’re joking, but using it for space constrained next token generation within a grammar sounds like a really neat use case.
Feed the generated note back into the input for the next query and you have ... autoregression?
If you provide it an AST of the english language, yes.
you could, but it the model is not optimized for text
this is complex, but generating text is highly complicated and requires mode dropping to make long cohesive text
was wondering same
Side note: it took me more time than I would like to admit to realize that Diogo Almeida isn’t a satirical version of the name Dario Amodei
That would have to default to Wario Amodei.
I feel like should be Cario Amodei. The D to C flip a rotation of the M to W flip
Flip both, Cario Vmodei.
It wasn't until the demo videos that I realized the post wasn't satirical.
thought the same lol
Well now I’m rooting for them!
I think what this shows is how important branding and comms are. They've captured imaginations with their demos and nomenclature, despite the arguably non-novel architecture. One forward pass, read the embedding space, train some regressors on predicate structure, [??]
I would love for things like this to be accessible via hubs like open router or AWS bedrock. It's hard to justify adding new model vendors directly with all the heightened concerns about privacy and security, but if bold new capabilities are added to a centralized already-vendor like AWS, technical people can adopt them without going through a whole compliance/purchasing/vendor review process. And an extra middleman tax is well worth it when the cost savings of the model itself can be one-two orders of magnitude.
I think the trouble is that Typesafe APIs don't fit into the normal OpenAI-style API that every other regular LLM provider users. You're not just providing unstructured text and getting unstructured text back. It would take a different request and response format than every other model on Open Router. Though you could shoe-horn it in some way, it'd be hacky.
But agreed it'd be very useful to see it deployed on other hubs, and it seems worth it to provide the bespoke API format. Perhaps Typesafe's API will end up becoming the standard for a new type of structured model, the way OpenAI's API did.
It can be shoehorned to work with OpenAI's newer Responses format.
Isn't openrouter the exact opposite of caring about security and privacy?
I guess you can choose your provider still? But isn't the point that the lowest bidder is doing inference?
You can set privacy requirements and define an allow list. To me the main value prop is that I get one bill for all models and can quickly try new models without signing up anywhere or changing my code. Oh! Also you can pass an array of models and if the first provider is down it automatically falls through to the next provider. More useful than it should be...
That's still ultimately privacy by contract (where you have to trust the inference providers to uphold their end of the deal), rather than privacy by design.
I always setup guard rails so that only zdr providers are used.
They don't like adding stealth startups :(
The thing is, in this climate it's hard to believe such tech will remain secret for long.
So, assuming this is not vaporware, this would raise the tide for everyone because it shows what's possible.
I'm trying to parse it down to what we had before vs what is new here.
We already had encoder models that skipped text generation for giving us a numerical output that could be computed as a probability. we also got no hallucinations and faster inference for free there. So we already had
1. "unstructured state in, probabilistic decisions out" 2. "orders of magnitude faster and more efficient"
What was hard there was to train the model head without ML expertise, and considerable amount of data.
This seems like this is a democratization of those encoders? The addition over existing encoders seems to be coming from being able to specify the output shape (up to a cardinality of 255). It is unclear to me if this is possible using Jev without additional labels for fine-tuning.
If so, that is still very impressive, but I think the faster inference and 0 hallucinations might come for free, from it not being generative.
I can't see how this is different from a fine tuned LFM2.5 encoder
This sounds good but so far all claims just sound like marketing terms. I'd love to see real proof. e.g. "RLCD" and "parallel sampling" have nothing to back it up.
also "70-500ms vs 3-329 seconds" are apples-to-oranges unless the LLM baseline is doing comparable work (e.g., long chain-of-thought). If Jev is skipping generation entirely for a narrow structured task, of course it's faster.
Nonetheless i want this to be true, so I'm looking forward to Jev
Edit: I really have to say that I like their manifesto https://typesafe.ai/manifesto
They have various benchmarks, e.g. how much time it takes them to do wikipedia page -> page games. Jev seems to take the same or fewer hops but in ~10x less time and for ~10x less money.
It's totally reasonable to compare against LLMs doing chain of thought if it gets comparable performance.
> If Jev is skipping generation entirely for a narrow structured task, of course it's faster
I think this is reasonable if people are actually using LLMs to solve this type of narrow structured task, which they are. The evidence is that every LLM provider has some method of forcing the output to conform to a json schema in their documentation.
love that you love the manifesto! letting the first batches off the waitlist now, but we do have some early users describing their experience (https://x.com/danshipper/status/2099947471518474522)
Congrats ! Really excited for the team.
It's not an LLM though it's a frontier model on structured data
Did you see the video where it plays Doom, it made it click for me
i've seen it play minecraft as well, what im not sure here is what is the thing that produces the JSON and keeps track of the objects
can jev play battlefield six for example
BTW it was not multi model playing doom, it was passing structured input and getting structured output. Its not what I thought: frames of video passed and real time game play.
so what? put an LLM on Cerebras and get its responses faster, and put Jev on Cerebras and gets its responses even faster
> I really have to say that I like their manifesto
Their manifesto: "you only build on top of it if it's trustworthy." - the irony of this while putting out the most misleading, dishonest marketing campaign I've seen in months for their first public appearance doesn't exactly scream "trustworthy" to me.
What do you find dishonest?
I am confused why they say it is not an LLM and then in the documentation it is shown as being an LLM derivative. The documentation makes it sound like they're taking a pretrained LLM and then giving it their unique post-training. How is that not an LLM?
FAQ: Is Jev just a smaller LLM?
Jev is neither small nor an LLM, hence being off the intelligence Pareto curve.
Image in documentation: https://mintcdn.com/ts-docs/aFVnpmCIX68NpsV1/images/ai-prime...
LLM seems to have become synonymous with Generative Transformer architecture.
While this model may share much with GPT-style models on the encoder side, it clearly has a different decoder architecture. So is a high-parameter count language model an LLM even when it doesn't have a GPT-style decoder? The definitions are in flux.
I could see this being fantastic for classification tasks. Last year I shifted from using LLMs for bulk data classification tasks (1M transcripts) to generating embeddings and categorizing based on cosine similarity. It saved a ton of costs and time, but wasn't as accurate as LLMs. This seems like it can give me Terra-level classification ability with the cost/speed I need.
This ought to work better than SpamAssassin, I'm sure
yup just joined the waiting list with a very similar use case in mind
Overall this seems like a classifier that gives weighted scores per custom labels. It's certainly useful, but whether it brings higher quality than an LLM in structured output mode has to be seen in objective benchmarks.
Zero shot classifier indeed. Reminiscent of asking an llm a yes/no question, constraining the output to either yes or no, and looking at the logits directly
And each question is a separate single token model completion done in parallel
do you all see the use cases being similar to what you might use Fastino's Gliner models for? i see similar differentiation from general purpose LLMs in the sense that they can take natural-language input and return outputs adherent to a user-defined schema.
https://fastino.ai/blog/gliner2-5-span-free-information-extr...
im thinking about how well Jev could be used to replace a current LLM-as-Judge evaluation workflows, specifically on chat transcript data (think ~1,500 tokens) i wonder if the reasoning usually required pushes it a bit out of scope. didnt see anything published about constraints on the state size, so would be curious to hear about that.
definitely seems like a modified version of GLiNER2 or 2.5: - encoder-based (no text generation) - multiple tasks in a single forward pass - deterministic outputs - constraint-based classification
I think this is a great direction -- for some kinds of users. And this makes me wonder if the 'vs' framing is misleading.
Yes, I think it's a mistake that many organizations are cramming LLMs inside of automated pipelines where the extreme generality/flexibility of the model is at odds with the fact that you're using it for a very specific task that gets repeated over and over, and needs a very specific structured output to be successful. But specifying your task carefully (as well as deciding what counts as your input state representation etc) seems like a form of programming. Something (a person or a model working in a relatively unrestricted way) will need to produce a configuration/specification for this system.
So rather than Jev vs Claude I imagine that using Claude/ChatGPT/whatever interactively to define / refine your Jev config which then runs in prod might be the happy combination?
There's a whole lot of information on this page that doesn't tell me anything about what this actually is. Can anyone spell out what the architecture is here?
They claim it's not an LLM, which I read as "not an auto-regressive token generator". I assume they are still using a transformer, otherwise they would be talking about the thing that's not a transformer, instead of all the fluff on the linked page. But they emphasize parallel generation, so is it like a text diffusion model?
Sounds like its essentially a generalized zero-shot classifier that takes and option set at runtime and works on unstructured inputs.
you pass in your "prompt" and options (described in natural language) that it can respond with, in addition to your input. it gives back that option set with a probability assigned to each one
yes and can do many of those in parallel
I would guess a tiny stripped down text diffusion model. It only has 32k context, and for choice mode it can only select from 10 choices.
> and for choice mode it can only select from 10 choices.
Rip there goes my excitement. I have a task that something like this would be great for but the list of options is a zero or two larger than that xd
They said that it works with up to 255 options.
you can still chain them
> Input tokens: $0.042 / MTok ($42 per billion tokens).
> Output tokens: FREE (too cheap to meter).
Insane. The video demos are really compelling, in particular the speed.
> Structured outputs slot into ordinary software as fuzzy decision rules: classify, route, score, extract, or branch where hand-written logic is too brittle. The surrounding code constrains their freedom, making them easier to compose into reliable systems.
I buy this vision. A lot of LLM integration I see these days is ultimately exactly this. OpenAI-style structured outputs works decently but this would be a great improvement in cost, latency.
thanks a ton!
constrained decoding (OpenAI-style structured outputs) make models dumber unfortunately - the short+dense version is that simply masking logits is insufficient because if ever a model was assigning probability to an invalid token, the model is by definition confused. you'd be better off erroring IMO
This is best thing to use for decision making, evaluation, classification. If I'm not wrong.
Can I put it as Air Traffic Controller? With similar error rates as humans?
That would be the litmus test.
"Does not hallucinate" is not the same as "is never wrong".
So the ATC test could be the benchmark.
Not hallucinating is easy when you don't produce strings.
Hallucinating as we use the word really only applies to generative AI. Non generative AIs can't hallucinate, they can just be wrong.
I'd love to know if Jev is still fundamentally LLM-shaped in architecture. Like is it using a single forward pass with a learned readout over the predefined options (i.e. a discriminative head on a transformer, no decoding), or something else? I did similar things for zero-shot criterion-based classification using a 4B Qwen model but could not reach the level of intelligence they've got here. Tho speed/cheapness was similar.
This is a good product but the naming/branding is pretty unfortunate.
Typesafe.AI sounds like some typescript/structured output type of tool…
What even is “system one” ?
IMO the product/tech is really there, just needs better communication.
The model can't reason comprehensively (e.g., like Sol XHigh would to solve a complicated problem), but it's designed to be able to answer anything a human reasonably could quickly and intuitively, i.e., system one thinking: https://en.wikipedia.org/wiki/Thinking,_Fast_and_Slow
I wonder how well it can play chess, or go.
I had a different initial confusion - it seems this company has no relation to the company formerly known as Typesafe https://en.wikipedia.org/wiki/Akka.io
I mean, it's a structured output model that (apparently) can't hallucinate. I don't mind the name.
It can't hallucinate, but it doesn't mean it can't make wrong decisions. Just because it adheres to a specific output format at all time, while LLMs have the output format at their mercy, then the claim of not hallucinating is made technically true.
I think that this specific part is not super interesting if your harness just recovers from invalid LLM outputs.
The latency and cost - yes, those are super interesting.
Wow, this is really cool. If this holds up to scrutiny, and has a decent context window (+16k), it suddenly changes our project's status from "cool concept, too slow and expensive to release" to "doable", just like that.
Just joined the waitlist, excited to try it out!
afaik Jev's context window is 32k
Great to see something new..
However I don't understand how are they claiming zero hallucination, how does giving confidence score fix hallucination? or am I missing something here?
https://x.com/completeskeptic/status/2099925682726002904?s=4...
The doom demo is quite cool
The doom demo is also in the article, for anyone that doesn't want to go to X.com. :)
I'm not sure the authors realize this is way more than "just a cool demo": if this holds up, it's going to be huge for game QA work.
Instrument your game to output properties of entities near the player and the output is the various control inputs - moment to moment gameplay gets solved. Maybe augment with a tick-by-tick controlled stepping mode if particularly twitchy - an LLM can take care of the higher level reasoning then.
Direct link to the Doom video tweet:
https://x.com/completeskeptic/status/2099925687465570372
The doom video is also in the article itself (headline: "Doom").
I suppose this is the same video as the one from the parent comment, but I don't know for sure - I don't have a twitter account and the above link doesn't work for me.
I linked to the tweet that has the video because if you are not signed in you cannot see the whole thread of tweets.
I can see the individual tweets in the browser while not signed in though.
But when their system is given the instruction "do not fire, simply dodge" - it doesn't "simply dodge", it actually gets close to the fleshy pink demon rather than keeping its distance. Or am I misunderstanding?
i think its just telling the model that it cant output a shoot action
Link to a raw MP4 of the video, from the parent article: https://framerusercontent.com/assets/rlL7ImEbISFoYt3IJEHHfvj...
It's in the parent article under a section named "Doom" in case that asset URL ever changes.
This is actually pretty cool. I think the undertalked about part of this for TypeSafe is that they can always "extract"/distill the frontier of this type of task from the newest LLMs for cheap. Jev seems seems to be GPT-6-Astra/Fable 5.1 but I imagine a bunch of training data is from earlier models?
Then, you can serve it faster/cheaper than the frontier LLMs. It's basically distilling a small but extremely common use-case from LLMs and serving it. Then RLCD comes into play to update weights when a new model comes out, etc.
Any thoughts on what the next potential "cheap" win to be distilled from frontier LLMs is? I'm going to need to play around with this.
> AI Map Reduce over Big Data
> Search for relevant information over giant corpuses
Do you mean as an alternative to embeddings?
This has the potential to be huge for computer use.
OpenAI has been teasing how fast computer use is with their models running on Cerebras chips but the difference here is a burning hole in your pocket.
Like which elements to select? Similar to the doom and wikipedia runs?
It could be used for coding if you gave it an AST.
If you work at TypeSafe please try this.
Side note: This is probably how LLMs would perform with better encoders and next-latent prediction, so eventually those will beat this architecture out. Still amazing though.
I've implemented tree-sitter in pi before, and while it works, I have no real proof it saves me tokens, or is more accurate. I think a better implementation is a model that's trained for AST's, not just "use tool, see what happens".
I'd love to do research on this when I have the time.
Cool project!
That's what I was insinuating through "better encoder"; the model creating more efficient representations of ASTs using something like JEPA
the hard part for coding is actually state engineering (e.g. getting your dependencies in context) - we haven't even tried it yet (because my philosophy is we should automate the easy tasks before the hard and we've been working on getting the model smart on the former)
we do think there's a lot of potential though and do want coding themed releases soon
I could see Jev being great at finding key symbols in codebase before a code generation/code review task. I sent you guys an email (to hello@) about using Jev in Code Review for www.ellipsis.dev.
I saw the CEO reply elsewhere in the comments to some other question. Maybe he can shed some light on it. My gut feeling is that this is non-trivial and they did not get this to work (yet?), otherwise I can’t come up with a good reason as to why they would not demo that as I assume half of the crowd here (myself included) would line up as customers.
Yeah it would be quite trivial to try and implement an auto regressive AST generator for STLC with Jev provided that you had bounded variable names and integers.
As you said, if it worked, they would have demoed it haha
im not seeing it.
youd ask it to pick a location on the ast to add something from the grammar?
i dont see how this stays confined well enough? make a new output space every time? does that end up auto-regressive?
what is the…epistemic status, for lack of a better way to put it, of the probabilities? what do they mean? what (probabilistic) guarantees do we have about, say, the responses to
- is the capital of france paris?
- it is august. is it raining in paris?
(forgive the examples; they're probably not semantically the sort of thing jev is trained to work on. but i figure the point translates to various kinds of questions that come up in "inner loop of agentic pid controller" contexts)
a normal text-generating model if asked to produce a number will also do that just fine. i assume in jev's case it was actually rled to essentially learn to express priors over things using its implicit world model, which definitely ought to help, but can we say more?
I'm trying to understand what difference does this make over LLMs.
LLMs are universal simulators, their latents model the world. So I bet if you compare their logprobs with probabilities output by this model, it will be highly correlated.
Someone should do this quick experiment. I bet there won't be enough of a meaningful difference.
This is too much for me. ML playing doom was a thing since before LLMs, decisions tree were always insanely and no one ever used then anyway, i can't see anything new in this yet everyone is treating this as a revolution. This technology was always there and quite easily accessible all along.
Seems like LLM can do everything Jev can do (just structured outputs?) but Jev is highly optimized and purpose built for it and thus way faster and cheaper. Is that a fair description?
We've already started using it for some pretty powerful decision tree stuff. We're just scratching the surface. We shipped an extension for Swamp[1] a few minutes ago and the combination is great!
The one downside is that the context window is very small (32k.) So some initial ideas we had for initial evaluation of code reviews won't fit yet in the window.
1: https://swamp-club.com/extensions/@swamp/typesafe-ai
Could you use this to build a proactive memory formation and retrieval system for LLMs that runs lightning fast?
Last 32k of connect + Summary of current task: Did we learn something useful here (true/false)? What is the category to file it under? Then notify the LLM to file it away.
What class of memory might be useful here? Model gives probability to each item in the list. Short description of all memories ordered by tagged class is used in the next round. Are any of these memories useful in the current context, such that they will inform the model and help in its task (yes/no)?
I’m sure there’s some fine tuning to be had, but this sure seems like the basis for a substantially better proactive memory system that works around an existing LLM conversation.
If I’m understanding what this does and how this works (generic input, intelligent classification with probabilities, rapid and cheap), this is absolutely nuts.
Very cool! Can you explain when I would use this vs. training a standard ML model on my data? Suppose I had a fraud dataset with features like customer ID, amount, merchant, online or in-person, etc. - I can't imagine that a general model like Jev would predict this more accurately or cheaply than even a basic XGBoost model trained on my dataset (one that I could build in a few minutes by asking Codex to build it). Where does Jev add value here?
you will need to collect data for every decision/usecase and then train a model. But this can be used for different use cases with just a prompt.
Founders response to a similar question on X: https://x.com/CompleteSkeptic/status/2100067328620896408?s=2...
pasting it here: zero-shot + general == programmable
I would assume any extreme scale narrow task could then be fine-tuned for, but we'll see - I suspect putting it all in shared cognitive core has bit maintainability/generalization benefits
a) Is this available on Bedrock? b) Does it support structured outputs? c) What about trying it out?
Doom demo is beyond impressive, even scary
It's very misleading. If I'm actually playing a game I don't get the coordinates of enemies sent back to me so that I can feed into my mouse to snap my crosshair to. It's looking through walls too, because it's working off structured state in text form. You could re-create this whole demo without using AI. Have an LLM generate the state machine for you and no model is required to run it.
As a zero-shot classifier, I expect that effectiveness is dependent on the data trained upon.
Jev input … > Unstructured data (e.g. text) with an emphasis on structured program state.
What pre-training data/model is Jev based on? Surely result effectiveness is dependent (outside of one’s own input as “state”) on that?
They never show exactly how they use it? Only a bunch of animations of it 'working'. Would like to see the actual code used for the demos!
It's a bit hastily put together, but I made a dspy fork where you can add a decorator to automatically use TypeSafe where possible on Signatures. It shows a fair bit of what actual, hands on usage looks like.
https://github.com/typesafeainate/dspy-typesafeify
DSPy seems like the right comparison and this is the first comment I've seen mentioning it.
Thanks for putting this together. I'm surprised the cost saving is so little though. I expected much more based on the post.
The doom demo shows the program state / query.
This is quite the paradigm shift. Can't wait to get my hands on it.
> [others] Output tokens: ~5x more expensive than input tokens.
> [them] Output tokens: FREE (too cheap to meter).
I'm very confused by this.
The output tokens are just responses to your inputed questions and their probability. So relatively few output tokens. No unstructured text back in the response.
They're not doing autoregression, so all the outputs are computed in one big forward pass. Very cheap.
I think OP is confused about "others" vs "them".
they’re talking about two totally different things, right?
it's our output tokens that are free (under the system one / jev column)
We had early access and found it to be pretty useful. Having a second form of verification, where you can ask multiple questions (in the form of Nouls) raised our confidence in the outputs of other models. [0] IMHO This type of model works incredibly well in concert with LLMs, not as a replacement.
[0] https://goodstartlabs.com/research/verification-is-the-bottl...
Don't your numbers suggest DeepSeek V4.1 Flash, for $100 more, gets you to slightly better agreement?
The Doom demo looks impressive but was it a fine-tuned model? It's the difference between a cool demo and revolutionary tech.
Shouldn't self-driving be a piece of cake if it works this well for Doom? Or what am I missing?
I think the doom demo uses a text representation of the world and it's basically, "projectile coming your way" -> "Strafe". "Enemy ahead" -> "shoot. So it works well when spawned in a room of enemies (as we see in the video).
If self driving is red means stop, green means go, and stay in your lane - then it would work great, but having to actually think and test which maneuver is optimal for a given situation while weighting safety, road rules, random unexpected actions and getting to your destination, I think it's a much bigger problem. A bigger model specifically trained on that maybe would do great, but then the output is not the constraint anymore.
But I haven't tried the model, so I 'm just ballparking and could be very wrong.
The model doesn't have image input capabilities (yet, it seems from the post), so for the Doom demo, a harness is extracting a bunch of structured information from the game (map layout, enemy locations, player ammo, health, etc) and providing it as a massive JSON blob to the model so it can make its decisions. This model _could_ be hooked up to make the decisions for a self-driving car, but it would need to be fed a structured blob of the situation around it, so all the computer vision problems of self-driving are still there. And that's before you get into the confidence and accuracy of this model.
Driving is more complicated than Doom, and it doesn't look that great at Doom to me.
Super cool! Instantly joined the waitlist.
It might be boring, but I can see exactly how I could use this right now to improve my agentic rag.[0] In two months I am supposed to deal with a giant corpus, while still maintaining responsive chat UX. I have been working my butt off to make our first big client happy. This could really help solve the chunk ranking problem.
[0] assuming the policies are compatible with sensitive production workloads, some time in the near future.
Haven't seen any docs or so. Is this actually a general model, or does it need training on the the data set it answers? Finding it suspicious you never see some kind of prompt.
Edit: never mind, found https://docs.typesafe.ai/introduction/quickstart by now
1. yes a general model 2. no training at all 3. but it is focused on "System 1" tasks (more human judgment, less math reasoning)
It's very generalized. Can't wait until everyone can see it.
Looks promising. I'm building an AI video editor and multi tool calls take >30s using Gemini. This would be a a game changer if Jev can take that down to single digits at p95.
I can see the value in this but looks like there's going to be trouble in communicating the difference between this and a regular LLM, and also proving the potential cost savings in using this to replace existing systems that are using LLMs with frameworks like langgraph, as this can't be a drop in replacement and would require a significant amount of re-architecting/reengineering of systems to get the type system to work
I was thinking about something similar (maybe) - generally speaking, embeddings for LLMs tend to learn real world concepts - things like 'fruit' or 'France' or 'city' as directions in embeddings.
But in things like programming, most concepts are abstract - 'if hungry eat an apple' in programming terms would look like
'if hunger > 50 {apples--; hunger-=30;}'
and compilers work with 'concept erasure' - to them, tokens (which are like llm tokens) look like
'if var1 > 50 {var2--;var1-=30}'.
They don't care about how these things map to real concepts. So all the embedding directions used to encode real-world concepts are just noise to LLMs when programming. This greatly reduces dimensionality and training costs. So does a token representation tuned for programming constructs, rather than natural language would probably have a more efficient encoding.
Current models go beyond the simple embedding because you start to encode groups of concepts in the context-aware part of the model (attention heads or any other method). So it is never simply words/tokens in isolation anymore.
If we could come up with a system to classify the probabilities across a large number of candidate words (or components thereof) then this could actually be good at producing text, one element at a time. We could call these elements 'tokens' and picking the right one could be called something like 'decoding'. Crazy idea but hear me out...
On a more serious note, it will be fascinating to see how this different spin on modelling inference will create new paradigms or slot into existing ones.
While I understand that accelerating development isn't necessarily the target for this, and it's not at all intended to generate code the way many of us are...
I think this could be pretty decent in CI? There's a lot of "flakes" I've mediated that this could have handled much more efficiently. Maybe observability as well, triggering elevated logging and other initial measures?
I think I missed why is this faster? What I’m reading here is it’s similar to constrained decoding but I’m not seeing the explanation of why it’s able to get those results.
insane doom demo i wonder what the limits of its intelligence are? i'm guessing it's not great at reasoning tasks, it seems breaking down the problem helps significantly, but how much does a problem need to be broken down for reliable performance? also this would be huge if it could run locally but it seems like there's no intention to do that at the moment
This will be insane for tool usage, and probably where the major economics for day-to-day usage will be.
The goal is going to be to use llms to distill operations down to some dsl, and pass it into something like Jev.
Hm, would be good to understand the architecture better. Is this answering just from a world model informed prior? How informed is it by the information in the prompt? I can't see this maintaining calibration across all domains and all types of structured output.
Is there anything published on how it maintains calibration? Or when you say "outputs calibrated probabilities" you mean "as calibrated as frontier LLM models, just cheaper" - which is a different claim; as LLM's aren't particularly well calibrated
Hasn't there been a lot talk about Astra's opaque reasoning capabilities (being able to think through complex questions without using a chain of thought)?
Given that, can't you just replicate Jev by telling Astra "here is the question, you must make a multiple choice decision / output a score between 1-10, please answer directly in a single word, no reasoning allowed"?
(Edit: Ok, Jev is much cheaper in input tokens so these two aren't directly comparable at all)
the edit is right - jev would be cheaper, faster, and more self-consistent (in general)
we actually use astra (and fable) in this way for our evals: evals.typesafe.ai
someone on the team cooked hard on that and it shows example traces comparing our model to opus/sol
This is amazing. I really could use this.
I like the idea of System one models but all LLMs so far work as system 1 thinking because humans generate speech subconsciously with system 1.
System 2 thinking requires consciousness which AI does not have, so even reasoning models are still system 1 thinking as system 1 in humans has reasoning with heuristics.
Its limited but most people navigate the world with it completely, so it's enough for AI.
Interesting concept, I can't see a reason to use a generalist classifier over an api rather then just training my own? If it was open weights I would probably mess around with it.
You know you're too old when you see the company name and think! Oh I wonder what Martin Odeskey , Jonas Bonér and co are up to. Wait, didn't they become lightbend... Altho this comment takes away from what these guys are doing which legitimately sounds interesting.
A question I have, with the type { output: string }, would the model not become a LLM? And if it does, shouldn’t it cost the same as a LLM for output?
I don’t see this as an option in their website
You could theoretically ask “what is the next appropriate character?” and add the entire ascii charset but i doubt it’d work well and you’d be implementing autoregressive churn across network latency…
strings (and all sequential data structures) are not allowed at all - this is how we make sure all outputs can be computed in parallel (thus no output token cost)
This makes me think of Expressions of Change [1], a project that aimed to make updates to a program a first-class primitive in a programming language. A model like this can't output code directly, but perhaps it would be well suited to select from the small set of discrete operations on code envisioned by the EoC author?
[1]: www.expressionsofchange.org
A few questions:
1. Do you provide any kind of largest common subtree caching for cheaper input?
2. Have you tried auto-generating Lisp programs structurally?
3. Have you tried augmenting a Lisp language with a `choice` function that makes choices given a prompt, the environment, and the continuation stack?
(1) Nope, it's always the same input token cost
(2-3) No, but that's kind of a sick cook ... Want to get access and try it? nathan@typesafe.ai
Thanks for the early access! I was testing the Lisp idea out in the playground, but I don't think the model is smart enough right now to generate actual code. I tried having Jev finish generating the code for a Fibonacci number function, but it kept wanting to create a literal number instead of refer to a variable which is a number. This happened both when I gave Jev the current program as a string and when I gave Jev the program as structured data.
Maybe I'm just not doing a very good job at prompting Jev, but I think right now it's not quite capable enough to generate Lisp code.
Link: https://console.typesafe.ai/playground?share=shr_148e1248984...
Huh this looks fantastic. The Doom demo really sold for me that this could be a great tool for accelerating QA at my gamedev studio. Signed up for early access.
This is basically a zero-shot classifier that can accept raw text (or structured text) as an input, and is able to classify that text as accurately (they claim) as a frontier-level LLM. I have workflows this would be useful for, looking forward to it showing up on OpenRouter.
exactly right!
So is it a structured data-based language model? Or is there a model and a harness? Hopefully they’ll open up and explain more.
it is just a model, no harness yet ;)
it is a structured data model, but technically not a language model (it doesn't generate language)
Funny how the authors are asserting that "doing the right task > data > compute > algorithms" while simultaneously releasing AI model for calibrated decision making, which if they work, would mean that "compute > doing the right task"
Congrats on the launch! What's different between Jev and Microsoft's Guidance package? https://github.com/guidance-ai/guidance Is it a diffusion generator under the hood?
Signed up for the beta! :) would love to put this through some real-world shootouts against traditional LLMs to see where this type of model really excels.
I’m guessing it might be able to replace maybe 40-70% of LLM calls for a given pipeline depending on the business task, cutting the API costs on those calls by an order of magnitude.
Very cool. LLMs have been borderline unusable as functions for the longest time, very excited for this direction.
From the person in the video regarding issues with benchmarks in general, and for LLMs. Also their approach. Good article.
https://substack.com/home/post/p-215252866
This sort of stuff almost sends shivers down my spine, it's like i'm looking 5 years into the future.
join the discord! we love forward thinkers
There was no "AI Winter"
Is this a Markov/Diffusion model with some sort of external Engram memory? If so, this could be extremely interesting.
Seems like "some" of LLMs tasks are now Jev tasks.
I am positive I know exactly how this works, I made something similar a few months back. But the problem is without generation you are extremely limited in the use cases. And while the model can't hallucinate, it can still be wrong. It just can't make up data.
Yeah saying it can't hallucinate is crazy. It can still forward a billing query to the dev department incorrectly. It can still get an obvious yes/no question completely wrong
Last year I also had a rather similar idea, but dropped it before I went very far in working on it. I wonder if you and I had similar ideas?
1. Start with an LLM, so that your model understands natural language.
2. Replace RoPE with a tree embedding scheme, and causal attention with a sparse attention on the graph structure. (You could use full attention... but it's cheaper to use graph attention.)
3. Chop off the final unembedding layer, replacing it with a projection down to two scalars, one for logits and one for confidence.
4. Each option of a choice is represented by a number of tokens in leaf position; average these tokens' logit outputs to get the option's logit. Average all of the confidences from all of the options to get the choice's confidence.
5. Train the logits by KL divergence from a true distribution (or NLL on samples from a true distribution).
6. Train the confidences on a subset of the data in which you know the entire true distribution.
The hardest part is getting real world data for workflows, but I wildly speculate that you can get by with only ~50,000 documents if you first adapt domains using synthetic data.
Are you able to share how it works in that case?
I'll tell you this. Output isn't too cheap to meter, there is no decoder.
So an encoder-only model with a classifier trained on the heads or something? DeepSeek recently switched to an encoder-decoder architecture in an attempt to get the best of both worlds (fast prefill while preserving generation capability), I wonder if that might be the future?
That was the first thing that come into my head. OK I can train very simple model, that can generate json's for specific tasks, so what? How we can be sure that this "limited use cases" not just overfitting for particular outputs (or even distillation?)
Except this, this thing looks like revolution.
Wild that it doesn't generate text. I wonder how its technology compares to Tesla's FSD stack.
> We deliberately chose not to publish performance against public benchmarks. In fact, we plan to only have one-off evals when we make product updates.
lol, I bet they would publish them if their score on those benchmarks were good.
Congrats on the release!
Finetuning a language model for decision classification (with probabilities) is already well-understood. What specifically changes in the training objective with RLCD? Are its benefits isolated from Jev’s new architecture/parallelism?
Will need hands on to truly tell, but the doom demo seems very promising. If it can play that with text descriptions of where stuff is by distance and degrees in a 3D context then many GUI automation tasks should be easily doable
I like it. What is it?
If it is so cheap, why such a limited release?
If I understand correctly, it can play chess and rubic cube better than LLM ? ( may be go too ? )
ok I've searched, it may not, but it can do "driving car" and "trade 0-DTE Option" much better.
So... a classifier model?
This puts the human even more out of the loop I'll guess?
That's kinda the goal. Imagine all the automation in the world being able to embed intelligence directly inside it - factories could route based on more complicated questions, hardware could anticipate your needs. Customer support could be done without humans 90% of the time.
Is the tradeoff of the parallel output that we don't get arbitrary string generation? like output # of tokens is fixed ahead of time?
Either way, really cool and impressive.
Yeah, it doesn't output strings, just decisions/answers.
Non-hallucinated ones at that.
The whole page reads like it was vibe-written by an AI. If I'd built something as disruptive as this claims to be, I'd have spent at least fifteen minutes writing the announcement myself. Every time I see 'we' in an announcement like this, I picture one guy alone in his basement.
unfortunately all hand-written :( my chief-of-staff does unironically handwrite em dashes though
For what it’s worth, I didn’t get that impression, and even noticed a couple typos ;)
I really appreciated the hand-written release. Thank you.
If it work as good as they say it does, confidence score + really fast response when you want very fast response, basically.. To me it is a crime against humanity to not open source it. Just get the money from cloud inference and cloud agentic sessions or whatever but open source it. This tech, a good harness, a good model provider, and you have basically a AGI building machine.
Golem, if you read this, add me on battle.net (europe) Sansviande#2540 and let's talk. Give me 1 minute.
“Crime against humanity” buddy please
This looks and feels a lot like productionized conformal prediction
Can this be used in conjunction with a text-generating LLM for better quality code generation?
What's the difference compared to just taking an embedding and feed forward a simple net trained for the task?
> Extraordinary claims require extraordinary evidence so see below for the receipts.
Yes, that’s the kind of attitude I want to see in these model releases
But the evidence is not there...
Indeed, they talk as skeptics but don’t offer a ton of evidence, other than a couple videos of demos. A live demo would be far more convincing.
They gesture at not using benchmarks for some reason...
https://typesafe.ai/blog/antibenchmaxxing
But also effectively this is a classification model. It excels at specific certain types of workloads, and obviously will fail at others. Not really sure how one benchmarks this tbf. I can see their argument on why this requires a novel specific eval for whatever your usecase is. A consistent "global" benchmark might be hard to do
Would it be fair to say that this is tailored for tool-selection subagents?
The technology and the results are very handwavy. What is RLCD exactly ? What are scores on benchmarks compared to LLMs ?
This website does not inspire confidence at all, it all sounds like a marketing piece. I wish it was true, some kind of text-prompted classifier with LLM performance would be cool, but I can't trust it with what we are given.
Can HN have a tag for open-weight vs closed-source models please? The progress is nice, but if it is not released at least in papers or open-weight? These are just ads?
If I’m understanding correctly, this will work well for self driving cars?
Any relation / inspiration to GLiClass?
Okay, so it doesn't output text, that much is understood. What are the inputs like? I'm assuming maybe a text input? maybe an AST definition? Really hard to tell how this works at all from the demos, especially since we can't really try it out.
inputs are structured program state. there is an example at around second 30 of the doom demo
(though ideally everyone gets off the waitlist and can try it out for themselves )
Why did they pick the name System One? It's not really explained what "System One tasks" and "System One shaped queries" are. Things that need a fast response?
Does this imply it's a very small model? I couldn't find anything about the model itself.
Maybe: https://thedecisionlab.com/reference-guide/philosophy/system...
Bingo. It's a Psychology term for the part of our brain that reacts instinctively rather than thoughtfully and logically
The eval is baffling me
> we assume there is a correct compute graph (a “workflow” represented in code) and use the predictions of the largest, smartest, and most expensive external models as reference probabilities. ... Rephrased: every model gets the same workflow. We test how they compare to the average of the smartest models (in this case, Astra and Fable).
They assume there is a correct graph, but they don't compare to that, they compare to the average of the smarts models? So the smartest models are getting it wrong but you compare that anyway as a benchmark? So the outcome is "how much of a Fable am I getting" etc. Why not compare the actually correct thing?
But then even on this hand constructed eval, the first plot is showing Jev at less than Sonnet 5 accuracy. It is barely better than Luna. There are two Opus 5's and two Sonnet 5's without explanation. What is the plot showing?
I gave up.
Is there a downloadable technical report somewhere?
Super intrigued by this - large scale automation using LLMs is quite annoying due to deprecation cycles of models from frontier labs and cost of running your own being prohibitive when you have a blend of them.
Am I the only one struggling to parse the distinction System One (the system/harness?) and Jev (the model?)?
Can this be used in practice to write code?
Not in the traditional sense of a coding agent, but we think there's a ton of opportunity in using it for context management ("do we _really_ need to pass all these tokens to the agent?"), semantic linting ("how does this score against this AGENTS.md: <...>"), etc.
Funny how it can do everything but not chat. Sort of how when I was a kid I thought of a medicine that could cure any disease except the common cold.
This is potentially huge and can crash the Big Two's stock prices or block their IPOs completely.
Woof, that page is hard to read. I don't understand what they've done to the way text is rendering but it's not great for my eyes.
If you zoom in (especially on the large title), you'll see that the text is a semi-transparent gray with a black internal outline. It seems like all the typography is SVG-rendered. Actually insane. I've never seen this before. Not even the most vibeslopped websites have that.
I don't know if they changed it since your comment, but it's all just text to me
I could put this to use today.
I think we'll see a bunch of different architectures over the next five years.
this might finally be smart enough and fast enough for jarvis. hard to feel like iron man when your assistant takes 8 seconds to decide to pause your music
I’m not understanding what this is. It’s a faster cheaper LLM?
Is there a bottleneck which would hinder putting this architecture in charge of a humanoid? Would it be able to operate continuously, for example in conjunction with an LLM for long-term reasoning? Doom seemingly works extremely well.
defining the workflow such that the operation is a set of relevant questions
um what is going on with the outfit changes in the launch video...
https://x.com/CompleteSkeptic/status/2099925682726002904
Can't tell if they're just having fun or if it is ai-generated. On the verge of not being able to tell. Voice sounds a little synthetic.
definitely not AI-generated - this is my real wardrobe
we also thought the voice at the end was AI-ish, but apparently that's a real voice actor but slightly sped up
The whole video seemed generated to me
Pretty sure that's an intended joke.
Reminds me of this: https://www.reddit.com/r/ITcrowd/comments/tg05j1/i_cant_beli...
oooooh it can play Doom!
forget LLM benchmaxxing sidequests, I'm sold on the real benchmark
Parallel inference where you don't want a subagent seems niche. But there is a lot of random things where businesses ultimately want some kind of score instead of generating something.
I think the interesting thing would be seeing if prompt injections still work with this kind of model.
we have played with this! the fascinating thing we've found so far is that adversarial examples for our model are quite different from that of LLMs so that they work even better together
What is it about the rendering of this page that is so... off? It almost looks like the entire thing is a <canvas> element.
edit: looks like a framer export where there is a text stroke being applied :|
Looks like a great model for NLP.
my sons name is also Jev
this is interesting, so not an LLM but can be used in these use cases that LLM's have been shoehorned into
https://docs.typesafe.ai/concepts/use-case-map
"While Jev gives up string generation, it’s optimized for structured outputs and can’t hallucinate"
Ouh! Any open weights models that can do this yet?? If not, how much longer? I have a Mac Studio coming soon.
> Outputs
> LLMS > Strings / generated text. Strings are flexible and can be anything: chat responses, code, hallucinations, refusals, or even type-safe structured values. To be used by software, responses need to be parsed + validated. There is also always some risk that the AI goes off the rails.
> Jev > Type-safe structured values. Possible outputs and structure are defined in advance. The model never makes type errors. All answers are accompanied with calibrated probabilities and confidence scores.
I mean, this isn't even remotely comparable to LLMs so why compare? Also, why are they bringing up AGI given there approach is so restrictive that what they're building literally cannot have the creativity required for AGI? The video is 100% marketing slop...
The bulk of the application of LLMs is that they generate reasonably reliable text which doesn't need to be defined in advanced. I'm sure there is a niche for this and congrats to the team, but please let's not hype this as if it's the next big thing in AI...
Creativity is not required for AGI, that's maybe the only thing that is not required for AGI actually.
What a sad world would you live in if you don't keep creativity for the humans.
If this is true means, AI Stock bubble burst. (For good)
It looks like a specialized encoder-only(-ish) transformer with scalar and ordinal output heads. Acausal in effect, maybe? Probably not even autoregressive?
I'd use this as a tool an LLM can use for specialized tasks. It's not AI in itself.
If this is true, then AI Stock Bubble burst (for Good)
"is this the real thing or is just fantasy"
This makes me kind of nervous for the whole AI thing now. Are people gonna lose their jobs, etc.? so much of the economy is now built on top of LLMs.
What's different about this particular model that worries you?
it doesn't require nearly as much compute as normal LLMs. anything depending on increased datacenter and compute spending would be threatened.
a world where people eat so they can feed Big Computer sucks. We need Little Computer, driving robots in the fields.
I definitely agree with this, but the transition is gonna suck for some.
Cool I guess. Definitely not worth the 1000+ points though.