This New Frontier Model Is Not An LLM.

Author: Tim Carambat · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: This New Frontier Model Is Not An LLM. · Channel: Tim Carambat · 15:07 · Watch on YouTube

Topics: Task and workflow automation, Model inference and optimization, AI models

Summary

Timothy Carambat discusses Jev, a new “System 1” model from Type Safe AI, which its creators describe as an alternative to traditional LLMs optimized for automation. Rather than generating conversational text, System 1 models are presented as producing fast, constrained decisions with probabilities and confidence scores. Carambat explains how this approach could help with classification, tool selection, and other tasks where predictable outputs matter, while noting that the model’s accuracy, size, and ability to run locally remain uncertain. He sees these models as potential complements to LLMs: frontier models for chat and reasoning, and System 1 models for structured automation tasks.

Key points

The launch video presents System 1 models as foundation models optimized for automation rather than natural conversation.
Carambat describes their output as constrained decisions with probabilities and confidence scores, rather than generated text.
The proposed parallel processing is presented as a way to make responses faster than token-by-token LLM generation.
Constrained outputs could support tasks such as spam classification, categorization, and selecting among available actions.
Carambat argues that using LLMs for these tasks can waste tokens and introduce latency, especially at scale.
He says restricting the available choices can prevent out-of-bounds answers, but emphasizes that accuracy still needs to be tested.
Possible applications include model routing, tool selection, filtering, and other automation tasks.
Whether Jev is small enough and effective enough to run on consumer or local hardware remains unknown.
Carambat imagines System 1 models working alongside LLMs, with LLMs handling chat and reasoning and System 1 models handling structured decisions.

Chapters

Start	Chapter	Summary
00:00	Introducing System 1 models	Carambat introduces Jev and plays its launch video, which describes System 1 models as an automation-focused alternative to traditional LLMs.
03:39	Why structured tasks are a poor fit for chat models	He explains why using frontier LLMs for classification and decisions can be costly and unpredictable, and describes the appeal of deterministic outputs.
06:14	Constrained choices and parallel output	Using email spam detection and game actions as examples, Carambat contrasts token-generating LLMs with System 1 models that select from defined options.
10:24	Accuracy, probabilities, and automation	He notes that constrained choices do not resolve questions about accuracy, then explains how probability outputs could feed into automated workflows.
11:58	Potential uses and open questions	Carambat discusses model routing and other applications, considers local use as an open possibility, and frames System 1 models as complements to LLMs.

Quotes

They don’t like reason. They don’t talk to you. It’s not chat. — 04:41

It instantly takes the prompts and from whatever options are available gives you essentially a confidence score — 09:20

And that’s what these models are for. — 11:26

Use Frontier for Frontier intelligence. — 14:03

Concepts and entities

Jev, System 1 model, TypeSafe AI, RLHF, BERT, Qwen3.5 0.8B, Timothy Carambat

Full transcript (original language)

00:00 Hey everyone, Timothy Carambat, long time no talk. Uh I got it like I’ve always done with this channel, I make videos when there’s things I want to talk about. Uh that’s really the impetus of this channel. I’m not an influencer and in fact actually I never know when a video inspiration or news is going to strike and like for example I am unfortunately in SF right now at the same time as Dreamforce. So I’m just hold up in my hotel just working for the day and there’s this thing that recently came out and that’s what I want to talk

00:31 about today. Talking about how this next model is not a traditional LLM but I see a huge amount of potential for where this is going. And so without wasting too much time, let’s talk about this new uh System 1 model or Jev. You may have seen this on Twitter, let me just show you kind of uh I guess the the marketing materials for this and then talk about why this is important. And so I think just to start off for anybody who has no idea what I’m about to talk about, uh we

01:02 can just watch the launch video. I think the launch video is pretty good but it it it is confusing for people who maybe are familiar with you know ChatGPT and LLMs in general right now. >> I’m Diego Alameda, founder of Type [music] Safe AI and at OpenAI I co-created ChatGPT and RLHF, the post-training algorithm behind most frontier AI. Our team trained the first models to be superhuman at instruction following, what we now call chat. >> [music] >> And we asked ourselves, are models that

01:34 are superhuman at chat AGI? The answer was obviously not, but the trillion-dollar question is why not? RLHF has led to LLMs that are optimized for human preferences and include issues such as mode dropping, overconfidence, and an overall lack of reliability. These flaws mean that LLMs require humans in the loop and almost no true automation can be done. >> [music] >> At Type Safe, we’ve spent 2 years building in stealth, and we’re finally ready to share our new type of foundation model that’s optimized for

02:04 automation. System 1 models. With a new architecture, [music] new sampler, and new training algorithm, reinforcement learning for calibrated decisions. The improvements are clear if you see them side by side. Ask a System 1 model a ton of structured questions, just like you would an LM. Get the answers [music] back near instantly. Meanwhile, LMs take hundreds of times longer to finish responding. Look at how the LM generates sequentially, which is great for a natural conversation, but [music] totally useless for computers. LLMs

02:36 extract intelligence from the tiny straw of auto regression, and similar to the jump that transformers made over RNNs, we are replacing sequential computation with parallel, because [music] that is obviously the future. Our System 1 models output decisions with probabilities and confidence instead of words, and they can’t hallucinate. They’re a lot more like code, reliable, fast, self-consistent, and type-safe. This opens up a whole new world of possibilities and applications. Today, [music] we’re releasing Jev, the first

03:07 public System 1 model. It’s 100 times faster, so real-time AI is finally possible. 100 times cheaper with input tokens priced at $42 per billion tokens, and output tokens are free because they’re finally too cheap to meter. Jev is super smart, and its intelligence per dollar is literally off the [music] charts. This is just the beginning. We’re excited to see what you build with Jev. >> [music] >> As we say at Type-Safe, we’re building prod, not god.

03:39 >> Now, that video was a lot to take in, and I don’t I have applied for the waitlist access, and like you did all that stuff, but I also have friends who do have access, and they have shared with me results, and like we kicked around the can on some things. So, I’m actually pretty well informed with the actual behavior of the model and I’m excited about it because when you talk about these models, it is not like it’s like a tool, right? Right now, everyone takes an AI model, any off-the-shelf model or frontier model and just throws

04:10 everything at it hoping for the best output, right? If it needs to help you write emails, it helps you write emails, but if it needs to make a decision or categorize or classify something, people will throw those problems at frontier models because taking a classifier model and fine-tuning it is just so much work and a lot of you need a lot of data, there’s a lot of testing, there’s weird edge cases. So, some people just throw everything at frontier models. This isn’t a good thing to do. It does work,

04:41 but it costs a lot of money at scale. This is I think where you can categorize system one models. They they don’t like reason. They don’t talk to you. It’s not chat. It’s more of like a deterministic way of using all of the benefits of a model like the kind of understanding of the world and language, but being able to get a deterministic, absolutely deterministic output where if you give it, “Hey, your only options are A, B, and C.” It can only ever choose A, B,

05:12 and C. There’s some other things that they talked about in this video like the pricing. The pricing is per billion tokens, which is insane, but the one thing that I’m mostly interested in being local, right? Is what does it take to run this thing? Because if inference or I’m sorry, if input tokens are so cheap, they’re pricing it by the billions and output tokens are so cheap, they don’t even charge for them. To me, that seems like a type of model

05:44 that could run on anything. If you look at traditional classification models like like BERT or something like that, if anybody knows what that is, just like a model that’s like, “Hey, I gave you input, give me an output.” And it’s been trained well enough to do that. Uh those work really well. You can run them. They’re very small. You can run them on CPU. You can run them on like little Raspberry Pis. They run quite well. Uh but they’re they’re still slow and they’re not deterministic. They can hallucinate. Anyway, to help kind of uh I guess like square this model, uh I

06:14 came up with a like a drawing, I guess you could say, of kind of a traditional problem that I see people throwing at frontier models that this model could reliably help with so that you understand what it is that they’re working on. So, this is a very simple way of thinking about it, right? In blue is the prompt. Is this email spam? Now, typically, if you were to ask that to a frontier model, it would go and be like, “Well, looking at this email, blah blah blah blah blah.” If you’re building

06:44 something that maybe does this in the background, right? Maybe it’s like a Chrome extension that goes to your server and it looks at an email and it does all this other stuff or maybe it’s like part of your email service and you can like filter whatever it is that you’re building, you cannot afford to be wasting tokens on latency. So, “Is this email spam?” You would say, “Return only JSON or a one or zero for true and false.” The model, if it’s a reasoning model, is going to go and think anyway and then give you that final output. And

07:16 this is assuming that the model is smart enough, not heavily quantized, and isn’t going to in in general just start responding with regular text anyway. A lot of people have been doing this. This is kind of similar to how tool calling actually works today. It’s the model just producing tokens that then get parsed into whatever it is that the output is or the JSON block that you’re looking for. This is a this this works, but it is it is slow, right? It’s the same as making every single token. And there’s a

07:47 similar class of problems where it’s maybe the selection is not so binary. It’s not a yes or no or a true or false. It’s a list of options that are specific to what you are doing. So, if I was, you know, doing a game, right? It would be like, “What’s the next action the user should take?” I would list all of the available options. And then the model would again, if it is a reasoning model, think, look at the screen, ingest all of this stuff, process all these tokens, and hopefully give me just the text of

08:19 one of the items that I output. Keep in mind, it’s the same thing. [snorts] It’s still going over all of the context. It’s still going over the entire process. And you are just hoping and relying on the fact that this model will not hallucinate, not it also follow your instructions, and ultimately just give you a static option. This works well when you have, you know, frontier models running on multi-million dollars of GPUs or bi- billions of dollars of GPUs. But when you start doing this stuff locally

08:50 or with heavily quantized models, it this is where I think a lot of people on this channel specifically probably have seen these models fall apart. Where instead of saying, “Move forward,” it says, “I think the user should move forward.” And then your code breaks because now it’s like, “Oh, okay. Well, now I got to go.” It’s not giving me exactly what I wanted. These are just kind of the quirks, I guess you would say, of an LLM. And just to really kind of drive home this point, this is where I’m talking about how this parallel processing also of data

09:20 accelerates things dramatically. And so, what you can see is TypeChat, this the Jev System 1 model, TypeChat is the name of the company building these models. It instantly takes the prompts and from whatever options are available gives you essentially a confidence score, right? This is similar to how LLMs predict tokens, right? It’s like a probability, like this token has the highest probability of occurring next. But, because the amount of options that it can choose from are constrained, it can

09:51 be parallelized very effectively. And compared to something like a GPT 5.6 tera, you can see that it actually goes through and generates the tokens as if it was, you know, building a sentence essentially of giving you the actual response you need. Given a defined number of inputs and constraining the output to only whatever it is the options that I’m giving you, makes it to where essentially you have a, and I’m using quotes here, hallucination-free model. Because the

10:24 model cannot give you an out-of-bounds option. If you give it three options, it will only ever return the highest probability of those three options. Now, its accuracy is the issue. If you give it, for example, three options that have nothing to do with the input, uh I have not done enough testing in my my friend hasn’t done enough testing uh to be able to determine like, how does that work? Does it just give everything no probability, or does it just give the most, I guess you would say sensible of the

10:55 nonsense, the highest probability? There’s there’s some things to work out here. But, that’s the way that it can be thought of. And so, just to kind of illustrate this again, you could have a task like categorize this email, and you attach the email. You tell it the only options available are spam, phishing, promotion, FYI, whatever it is you want to do. And what you would get back fundamentally is essentially a probability matrix of whatever the highest probability option is, and deterministically, you can just grab the

11:26 highest one or the lowest one or whatever it is that you’re looking for, and then continue on with your code. And that’s what these models are for. They are for using the benefits of AI and training in this kind of world understanding, but not having to go through all of the, I guess you would say, pains and bottlenecks of traditional LLMs while still getting that deterministic output that actually makes automation possible, right? If you have code that you cannot reliably determine what the output will be 100% of the time, you don’t have

11:58 code, you have a bug in disguise. I know I’ve given examples here of like, you know, classifying an email or like is something yes or no, very binary things. But you could go farther with this. Um one of the most obvious areas for this is model routing. So if you were to give it, like let’s just say based on a complexity factor or something like that, give the user prompt to the system one models and or the history or whatever it is that you want to do and say, “Okay, I want to classify the

12:30 complexity of this model based on most complex, like medium complexity, low complex, whatever.” You can do model routing based on this, too. The fact is is that these models, I imagine, are quite small and can run, I would imagine, very effectively on consumer hardware given how they work. The parallelization though may be the issue here, but regardless, we don’t know enough yet. But the idea is that imagine basically now we have deterministic systems that have all the benefits of AI. We can use them as tools. We can

13:01 build more reliable tools, and this is going to wind up, in my opinion, becoming a fundamental concept to building a system around AI. You’ll always have your LLM to do like the, you know, the chat essentially, right? The reasoning, the chat, the understanding, um but you’ll have these models to do, I guess you would say like the ugly work, right? Choosing which tools to call, choosing where the model should go, classifying, is this good, bad, yes, no, uh those kind of things. That’s where this

13:32 shines, but that’s also where a lot of tokens get wasted. So, I’m I’m I’m actually really excited. I really want to try these models uh just to see how they work and kind of like put them maybe through some use cases, but also what are these models? Uh can we run these locally? Cuz if we can run these things locally, the idea of like an effective like personal intelligence on your computer is I mean, that’s that’s just like a no-brainer. Um there’s so many things I’m thinking of in my head of just like, “Wow, like I I I have a

14:03 million things that I would love to plug this model in for like anything LLM.” There’s several tools there. And I’m sure you’re thinking like, “Oh man, if I had a deterministic way where if I gave it something, it could only ever give me the answers I expected.” Yeah, like that’s really nice. You get the benefits of both. Use Frontier for Frontier intelligence. Use this for like tool calling, selection, filtering, categorization without having to train a custom model to do it. That is the key point. You could probably take, you know, Qwen 3.5 0.8B, fine-tune it to

14:36 only give you a specific input or output um or a specific output or type of output. But at the end of the day, it’s still an LLM, and at some point it’s going to get over quantized, it’s going to get over context, going to fail you. And this might not. So, either way. I just want to make a video about this. I thought it was really interesting. Again, I just you know, I never know when I want to make a video. I didn’t even know I was going to make a video. I’m in a random hotel room, and I didn’t even bring my mic. So, I apologize for the audio quality of this video, but either way, hopefully you find this interesting. I’d love to hear y’all’s

15:07 thoughts on it, the tools you’re building, maybe use cases. Honestly, I was just outlining a few uh just from the limited experience I have, but let me know what you think about these system one models. I’m going to link the uh website and stuff in the description, so you can do your own research. But, yeah. cool news. I like this. Anyway, thanks. Bye.