TypeSafe AI Jev vs. Laya (what they are and the controversy so far)

Author: Adam Gardner · 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: TypeSafe AI Jev vs. Laya (what they are and the controversy so far) · Channel: Adam Gardner · 06:53 · Watch on YouTube

Topics: AI agents, AI models, Open source and software ecosystems

Summary

The video explains Jev by TypeSafe as a fast decision engine that classifies structured inputs rather than engaging in open-ended chat. It then outlines Nandakishor Mukkunnoth’s claim that he developed a similar approach earlier and released it as the open-source project Laya. The presenter describes the dispute as a contrast between Mukkunnoth’s account of being overlooked and TypeSafe’s description of Jev as the first public system-one model. The video closes by suggesting viewers try Jev, Laya, or both, while noting the presenter’s experience and uncertainty about whether this kind of model will endure.

Key points

Jev takes textual input and returns answers from defined options, such as true/false or multiple-choice outcomes.
The presenter describes Jev as a decision engine intended for use within other workflows and applications, rather than as a chat interface.
In the blog-post example, Jev could assess brand fit, suggest a publication month, and identify a suitable staff reviewer.
The presenter says Jev can answer these questions in parallel, making it fast and inexpensive, and says its output tokens are believed to be free.
Nandakishor Mukkunnoth says he worked on a similar architecture in March 2025, published papers, released model weights and a dataset, and shared a package and approach publicly.
Mukkunnoth says TypeSafe launched Jev in September 2026 with the same non-autoregressive decision concept, but without technical papers, open weights, or open training datasets.
Mukkunnoth says he built Laya as an open, horizontal decision model after addressing limitations in his earlier approach.
The presenter notes that TypeSafe’s homepage calls Jev the first public system-one model and says the company promises not to train or fine-tune on retained user data.
The presenter recommends trying Jev or Laya, especially with a GPU, and says Laya did not work on his Mac while Jev impressed him.

Chapters

Start	Chapter	Summary
00:00	What Jev does	The presenter introduces Jev and explains how it answers structured classification questions, using a company blog post as an example.
02:05	Jev as a workflow decision engine	The video describes Jev’s intended role in other apps and workflows before introducing the claim that a similar architecture was previously developed as Laya.
03:06	Mukkunnoth’s account of Laya	The presenter quotes Mukkunnoth’s account of his earlier research and his criticism of TypeSafe’s Jev launch.
05:09	TypeSafe’s position and the presenter’s take	The presenter notes TypeSafe’s wording and data-retention statements, then shares his experience trying the projects and his thoughts on their future.

Quotes

Jev is not an LLM that you can chat back and forth with. — 00:00

So Jev is a decision engine and you use it inside other workflows and apps. — 02:05

I worked on this literally one year back in March 2025. — 03:06

Instead of staying bitter, I decided to take everything I learned — 04:38

I’ve been very impressed with Jev, to be honest. — 06:19

Concepts and entities

Jev, Laya, TypeSafe AI, Nandakishor Mukkunnoth, Diogo Almeida, decision model, System 1, reinforcement learning

Full transcript (original language)

00:00 Jev by TypeSafe is the hottest project this week in AI, but there’s something else kicking off too, laya. I’ve been experimenting with Jev for the last few days, and I’ve got to say, it’s really impressive. For those of you not aware, Jev is not an LLM that you can chat back and forth with. What you can do is provide textual input and a set of outcomes, either true or false, or sort of a multi-choice answer that the question can then go into one of those buckets.

00:30 Basically, if you’ve got some data and you can already tell that a human could answer that with a yes or a no, or a multiple choice answer, Jev can evaluate it too. Let’s work through an example. Imagine you get a draft blog post from a colleague and you need to evaluate the document. Does the writing style fit our corporate standard? For example, I’m British, so I write ISE. The corporate standard might be IZE and all the wording and terminology. Does that actually fit our corporate standard? Is the topic on brand for us?

01:01 Based on what focus areas we have as a company each month, between now and the end of the year, which month should we publish this blog post in? And which of our technical staff should be approached to validate or co-author this blog post? Basically, looking at the previous staff blog posts, which would be the best staff member to have a look at this blog post and say yes or no. Now, for each of these questions, we humans can imagine an answer based on the corporate style and spelling guide, for example, the blog post is either going to fit or it needs work.

01:34 Based on what topics we’re focusing on as a company in the next 12 months, as I say, we can judge whether this post feels right to be published and when. Based on the focus areas for each month, we can say that this blog post should probably go out in September. And based on which staff authored and validated previous blog posts, we’ll probably be able to guess the best staff member to look over this post before it goes out into the world. Jev can answer all of those questions and it can answer them in parallel,

02:05 meaning it’s lightning quick and it’s cheap. And actually, I believe the Outlook, the Outlook, the output tokens are actually free. So Jev is a decision engine and you use it inside other workflows and apps. You probably won’t be interacting with Jev directly. There’ll be kind of a UI, a wrapper around it. Jev’s going to provide the answers and you’ll use those answers to drive other processes, setting the publish date for the blog in a calendar or contacting a member of staff who can review the blog,

02:35 you know, sending that email out before publication, whatever the case may be. And that’s where the story might have ended. A very, very useful project with some really clever engineering behind it by a well-cashed up Silicon Valley lab, sure to be acquired or acqui-hired in the next couple of months. Except we discover Nandakishor Mukkunnoth and his claim to have built the Jev architecture a year ago, which is now open-sourced as Laya. So what’s going on?

03:06 What’s the truth? Well, the good news is actually we can hear it from his own words due to his LinkedIn posts and a post on dev.to Here’s what he says. Everyone in AI right now is talking about a new kind of model, an architecture that is not auto-regressive, does not generate text, and gives lightning-fast probability productions over a structured JSON schema. Seeing the hype online feels both validating and deeply frustrating. I worked on this literally one year back in March 2025.

03:36 I spent months of hard work, sweat, and sleepless nights building it, published a paper, released the model weights on Hugging Face, published the open dataset, built a PiPi package, and posted the whole approach on Reddit. Then, in September 2025, I published a second paper, laying out the exact framework for schema-based decisions guided by reinforcement learning. For anyone curious, the guiding brain in my system was always reinforcement learning, not just an embedding model or an auto-regressive LLM.

04:06 And then, in September 2026, a well-funded frontier lab called TypeSafe AI, founded by Diogo Almeida, a co-inventor of ChatGPT and OpenAI, launched Jev. They proposed the exact same non-auto-regressive decision concept as if it was a brand new scientific breakthrough, except they launched without technical papers, without open weights, and with zero open training datasets. It is incredibly frustrating when something you poured your heart and soul into for months as an open-source researcher

04:38 gets overlooked because it was built for a vertical use case, while a well-funded lab packages the same core idea horizontally and gets all the glory. But that is the open-source story in general. He then goes on to say, Instead of staying bitter, I decided to take everything I learned from my March 2025 and September 2025 papers, fix every architectural limitation of the old approach, and build a completely open, horizontal, system-one decision model, RL agent.

05:09 Now, I am going to skip over the rest because it’s full of very technical details and maths that, frankly, I don’t understand. But I wanted to put this on your radar in case you’ve heard about Jev and are wanting to try out something open-source. I’ve also linked to everything, including all of the posts and HuggingFace and all of that stuff, in the video description. Now, in TypeSafe’s defense, their homepage does say the first public system-one model. As in, the first public system-one model.

05:42 And you’ve got to assume here that lawyers have been over every single word of this website and copy to make sure that they are compliant. TypeSafe do retain your data, but they promise not to train or fine-tune on it. Now, if you’re an Enterprise customer, you can obviously contact them to negotiate a zero-data-retention contract. So, there you are. Now, you’re up to speed with the whole Jev / Laya situation. Basically, the original producer, maintainer, developer has open-sourced Laya because he says and feels that his solution was a year ago,

06:19 and JEV has just come along and kind of taken it. Definitely do try out one or the other, or both, especially if you have a GPU. I couldn’t get Laya to work on my, frankly, quite piddly Mac, but I’ve been very impressed with Jev, to be honest. It’ll take a little getting used to now that we’re all used to kind of chatting backwards and forwards with LLMs. You kind of need to unlearn some of those bad habits of just chatting, and I guess we will see in the future where this goes, and whether this is the next low-overhead, low-cost step that we’ve all been waiting for and all needing,

06:53 or it’s just another flash in the pan.