Quentin Feuillade-Montixi: Why AI Needs Its Own Language | Intelligence at the Frontier 2026

Yes, so it was a bit my fault. I'm not going to talk about I am Sekhri James. Sorry about the people that were expecting this. Um I sent my slides way too late. So my my speech is about a new programming language.

So previously I worked on I am Sekhri James for a year and a half. I actually wrote a post called studying the alien mind which tries to like approach the systems as an alien mind and try to study them with like an ethological approach. Uh but this was my previous startup. I stopped doing this like 6 months ago because I was quite frustrated by the state of the way we build systems. So I spent 3 years like red teaming them and doing lots of jailbreaking for like open AI and robotics and a lot of companies like this.

And I was very frustrated about the fact that the way we build systems is broken and like I saw a lot of ways we could have improved the system I was jailbreaking. And so this is why I'm going on the other side right now. I'm trying to like think about like how do we actually build a system that are robust and how do we put put the tools in the hands of the developers to be able to create those alien AI systems. Uh I'm also going to get my notes. Because opera.

Okay. So so the first the way I'm thinking about it is I think the first thing we need to fix is the language. So we've always Is it which next one? What what button is it? Thanks.

Just before you. So in the past we've always um built the language around the thing that we needed to build. So for example when we started having hardware we started with like assembly. So assembly was developed for talking directly to the computer. And then we had C for building more complex operating systems.

And then we had Python and JavaScript. Like we always try to go one level higher in terms of abstraction for the things that we are building. And and currently the way we build and what we build has changed dramatically but the language has not changed at all. So I think this is one level that we need to improve on. Yeah, thanks.

Uh so there's two different things that are strange. There's the way we build. So, I was part of the people who were like very resistant about about like putting AI into the my coding uh, experience because uh, I I love like having a deep knowledge of all the parts of my infrastructure. And so, I took a lot of time before I actually put a like use the AI for coding for myself. And uh, I started doing this around 6 months ago.

And uh, I like the way I code has completely changed. Like I almost don't read anymore. Like all my lines of code What I do is I I'm trying to be very specific about what I want to build. And then I spend time like very like describing a lot. And then I'm I cannot trust the AI to build correctly.

And then in the end I just double-check by like asking the AI to do um, like whether the the infrastructure is actually what I wanted. And um, and I have a really hard time like trying to visualize what's actually is happening. And actually, I have an anecdote on this project. Uh, so it's been a month and a half I started. And uh, the first week I kind of like was very specific about the infrastructure of what I wanted to build.

And um, a week after this, I kind of like started building on top of it because it was working. And uh, and I discovered at some point that one specific point I said about my infrastructure was not followed. And so, I ended up having to redo the entire thing. So, I lost a week of work just because the AI didn't like follow exactly the infrastructure I stated. And uh, so there's an issue with like the coordination between the human and the AI when we build.

Like the way we build is like fundamentally broken. And the second thing is the what we build is changing. So, like we have many agents, we have many like AI being integrated into the code currently. But uh, and I I would argue that we are very far from having seen the actual like best way we are going to code with AI currently. So, I In my previous startup, I worked on a a system that was automatically generating Java code.

And uh, the way it worked is like it was very um, robust. And it was also um, so I I used a genetic algorithm. And for genetic algorithm, you need to have like a real understanding of how your individual is like um, done. So, basically, you want to have like uh, um, very clear understanding of my my individual is made of these operations. So, you want to have like different operations.

But at the same time, you want to also to have the flexibility of being able to generate whatever prompts. So, for example, you want to be able to say, "I want to elicit the behavior of generating methamphetamine." So, what I want is like this is my input, and my output is I want to find the prompts that elicit this behavior. But then you have like a methodology which says, "Okay, I know how to build a prompt that needs to like add this specific sentence. I need to transform this sentence with like Japanese.

I need to like encode this sentence." And then like all those operations combined becomes a prompt specifically. And then so, the way I build it is like I had objects which were operations, and I had LLMs that were doing the glue between the elements. So, it was like very much like uh the rigidity of code and like the the ability to have actually understanding of what's happening. And at the same time, the the generalization of LLM which were able to like turn unstructured data into structured data inside the infrastructure itself.

So, it was not like it was not a thing that was done uh on top or like in the loop. It was part of the infrastructure itself. And uh and the language currently don't really allow this kind of of developments. Uh if you can do this with like doing API calls within your code. So, you have to like really patch like add API calls within the code to be able to do this like woven infrastructure.

Um so, maybe next. Um so, there's all those things also like uh when AI does like code currently, it lost a lot of of context on boilerplates, stuff that people have already done. For example, like API call, they have to like go search on the the specific API um um information, like what are the input, what are the output. And uh and if like it's the AI has to code a lot of patterns that have already been done. Um also, the humans is very hard for them to like see what's actually happening.

A second thing is So, this hybrid execution model you can't have it currently. And uh all you can do this with API call base not great. And the patterns aren't very composable, so you can use libraries. Libraries are used for reusing patterns that other people have done, but it's not a high level of reusability. You have to like use a specific functions, but it's not like a patterns that have already been previously that you can reuse.

Uh so, my the language I'm building, it has three properties. So, the first one is everything is a node. So, when you think about the language, it has what we call first-class citizen. First-class citizen of the language is for example an integer, a float, a function, an object. These are like first-class integer.

For me, the first citizen is everything on top of it, which is like LM codes is a first-class citizen. A code itself is a first-class citizen. So, it's not the integer itself, but it's the code, a piece of code. A human in the loop is a first-class citizen. A database is also a first-class citizen.

So, it means that your your code does not have like a difference of like this is your infrastructure, this is your code, this is whatever. Everything is on the same level of abstraction. Yeah, next to me. Thank you. Uh so, it looks like this in practice.

The the AI has like there is two views, this is the second property. There is the code view, which is what the AI sees, and there is the human view, which is what the the human sees in the area in the workflow. Um and it's like very when while the AI codes some piece of code, you can see live the thing being updated. So, you can like very follow exactly what the AI is doing through like a visual architecture. And and also I think I forgot something in the previous one.

Group of nodes are nodes, which mean like you can create a pattern, put it like chain a bunch of nodes for like a specific patterns, and you can collapse it back into a node because it has like input, output, and a specification of what it does. And so, you can like reuse patterns of stuff that people have already done, and it's basically itself a first-class primitive. So, you can always go higher and higher in terms of level of abstraction. And uh and it also also the case for uh the code itself, which means that you can always like take a group, collapse it. And here I collapsed it a bit like manually, but what's happening is that you can see the what is the input of this group, what is the output of this group, and what is the description of what it's doing.

So, when we when you collapse, you only see in out description. And when you expand, you actually see what's happening inside. So, they I can like dynamically update the code while it's working on specific things. Um yeah, next. Thank you.

Uh and uh and also like this is part of the another aspect I already touched on, but uh currently when we build system, there is like either we like do full code, it's like rigid, predictable, and it's not like very you can't really uh handle a lot of unstructured data. Uh or you can do agent and tool codes, but they are like very fragile, and it's like uh and you can't really control what's happening. And uh the having a graph allows you to have both the ability to um have rigid like control and structure into the system, and at the same time you have the ability to have structured data that flows, because information can be whatever, but the the structure is done by like the design of the infrastructure. Yeah, thanks. Uh and yeah, I one other thing I was very like um adamant about is having um human in the loop.

So, I want to have the another which is basically the human, and it's done with an extension. So, I built a browser extension where you can um you can connect to the back end, and while your execution run, it can like query you when there is a specific um uh like needs for a human in the loop, basically. And you can just drop it inside the the graph, it just like flows in the human, and you can just like answer the question, and it's really like consider the flow uh actively. Yeah, thanks. And uh and also the the aspect about um the the ability to have like group of knowledge being shareable is that you can like optimize a lot on specific patterns for to make them like robust and um and like compliant.

So, it allows people to just say, "Okay, I want my system to be compliance. I already have a pattern for I don't know a browser extension." So, you can imagine a browser extension patterns with like prompts with like robustness checking fronts with um like a lot of things you can build around this browser extension. And then you can collapse it into a node which is the browser extension. And anyone that uses it basically has like the ability to have like a robust um browser extension and a robust and like compliance by default.

But, next And so, I'm just going to show it was supposed to be a video. Ah. Okay, my bad. It's my fault, I think. I forgot to share.

Uh yeah. So, I just wanted to show in practice how it looks currently. Um so, I did a comparison between Cloud Code and um and my tool for building automated cold outreach. So, it's something that a lot of people are talking about, but no one is actually trying to build it. And so, I tried in practice to say, "Okay, if I wanted to automate my cold outreach with Apollo."

Apollo is is a company that does um people and organization search with like high level of details. And so, I did the comparison. It actually took me an hour and a half to get it with Cloud Code. And I actually I tried voluntarily to do it like well. And I had one big bug like it ignored one of my instruction and I got to an infinite loop.

I had to run to the computer to like stop the computer to run. Like I was I spent a lot of tokens on it. And for my tool, it was 4 minutes. I just like did one prompt. It's uh Yeah, the the video shows how it looks like in practice.

Like when the AI codes, you see like the the graph being built real time. And then the execution is like once it's done, you just have to fill one thing which is like the app information. And then like the sorry, the Gmail credentials. And it directly runs. And yeah.

And so, it's like also the the amount of lines of code you have for this system is very compressed compared to So, it's still a very beta version. So, um this is probably going to be going down as when I optimize the system. Uh, yeah. And I think I think that's good. Yeah.

If you have question, I have some time. Thank you.