This talent is worth more since AI

3 Quebec entrepreneurs explore how to build AI agents, protect your digital sovereignty, and identify what machines will never be able to do in your place.

Spotify Apple Podcasts Key points 01Building your AI agents without being a coderJonathan explores Claude Code and builds his own AI agents without being a developer. What used to be reserved for devs is opening up to everyone · and that changes everything.

Julien introduces the concept: your digital talent can be reproduced by machines. Your organic talent, on the other hand, is irreplaceable. The key is to identify and pilot both.

Lovable, Base44, Replit, Cursor: Jonathan and Vincent break down why these tools generate generic output. Pixel-perfect fidelity stays out of reach without precise annotation work.

Don’t give everything to the same model. Split your data between OpenAI, Claude and Gemini. Vincent is building a local AI agent at Crewdle to protect sensitive data.

Vincent builds the infrastructure that lets companies deploy AI in a reliable, sovereign and fixed-price way.

30 years in design, three years reinventing everything because of AI. Jonathan teaches companies to declare their intentions before touching a tool.

15 years in VFX (RodeoFX, Scanline/Netflix), now coaching creative entrepreneurs. His question: does your talent protect you, or does it trap you?

LinkedIn What we talk about In this second episode of Blind Spot, Jonathan, Vincent and Julien dive into the topics that define AI in 2026: how to build your own agents with Claude Code, why vibe coding has concrete limits, and how to protect your digital sovereignty without sacrificing performance.

Vincent Lamanna (Crewdle) has turned his team into a network of autonomous agents where every employee pilots their own AI tools. Jonathan Bélisle has explored design and transformation through AI for 30 years and warns against the illusion of full automation. Julien Klein, after 15 years in VFX at Scanline/Netflix, develops the concept of organic talent: what makes us irreplaceable in the face of machines.

Vincent Lamanna has built an AI infrastructure at Crewdle that transforms companies, and is developing in parallel a 100% local AI agent to protect sensitive data. Jonathan Bélisle spent 30 years in design and helps companies avoid confusing speed with depth when it comes to AI. Julien Klein managed 150 people in VFX at Scanline/Netflix before dedicating himself to creative coaching, and proposes the concept of organic talent as a shield against automation.

Together, they explore why processes remain essential even in the era of AI agents, how to tell your digital talent (what machines will replace) apart from your organic talent (what they will never do), and why choosing a single AI model is a major strategic mistake. The conversation moves from the technical to the philosophical, from the operational day-to-day to the geopolitics of language models.

Three complementary perspectives, one single certainty: your organic talent is what AI will never be able to do in your place.

Artificial intelligence doesn’t replace talents. It amplifies them. And that is precisely where the trap hides.

For years, the recipe for professional success was simple: become an expert in a field, accumulate experience, and your value will grow over time. That model worked for decades. It is collapsing.

In the second episode of Blind Spot, Vincent Lamanna (CEO of Crewdle), Jonathan Bélisle (Paracosm) and Julien Klein (monExpansion) explore an uncomfortable question: does your talent protect you, or does it trap you?

Organic talent is the talent you built with time, sweat and experience. 10,000 hours of practice. Thousands of delivered projects. An expertise you carry in your body and your intuition.

Digital talent is the talent AI can reproduce in a few seconds. Generating code, writing content, analyzing data, producing visuals.

The question is no longer whether AI can do your job. It is understanding which part of your work is organic (irreplaceable) and which part is digital (automated tomorrow).

For most senior professionals, the answer is uncomfortable: a good chunk of what they do day to day has become digital without them realizing it.

The CEO of Nvidia popularized an idea that perfectly sums up the current era: we are approaching the speed of light in terms of information processing.

Velocity, in physics, is speed plus direction. AI gives you a disarming speed of execution. You can produce in an hour what used to take a week.

But without clear direction, that speed becomes a catastrophe. You produce mediocre content faster. You code fragile applications faster. You make decisions faster without the necessary information.

The real competitive advantage in 2026 isn’t mastery of AI tools. It’s the clarity of your direction. Knowing where you’re going, why you’re going there, and what you refuse to do along the way.

Vibe coding (coding by letting AI generate the code from vague descriptions) became a cultural phenomenon in 2025-2026. Social media is full of impressive demos.

But as the three co-hosts of Blind Spot point out, there is a fundamental difference between a prototype and a product. When everyone thinks they’re done in an hour, an experienced engineer still sees three months of work.

Vibe coding is an extraordinary prototyping tool. But it’s a lie when it comes to production quality. Error handling, scalability, security, maintenance: everything that makes the difference between a demo and a product that runs is invisible in a prototype.

That’s exactly where organic talent regains all its value: in the ability to see what’s missing.

Vincent Lamanna, CEO of Crewdle (a peer-to-peer videoconferencing platform), brings an essential perspective: digital sovereignty isn’t a political debate, it’s a matter of business survival.

Every time an entrepreneur connects their client data to an American model (OpenAI, Anthropic, Google), they train that model with their own competitive advantage. Law 25 in Quebec requires traceability in the handling of personal data, but how many entrepreneurs really know where their data goes?

The real question before choosing an AI tool: who owns what you give it?

The Talent Trap isn’t a fate. It’s a signal. Here’s what the three co-hosts of Blind Spot suggest:

The trap doesn’t get solved by looking at it. It gets solved by taking a first step. Two options.

Assess your 6 core needs Identify your form of the trap Personalized verdict + concrete next steps Access to the interactive tool Take the diagnostic → THE BOOK The Talent Trap 208 pages. The complete mechanism, the exercises, the stories of Hong, Sofia and Emma. The 10 signals + the 6 locks Digital vs Organic explained Your expansion profile The 90-day prototype We couldn’t confirm your sign-up. Your book is on its way to your inbox. Send me the book “If I could do this, you can do what this book offers you. This book was born from a conversation I had just before someone died. It’s written so that you can have your rebirth.”

One shoot. A channel with 25K subscribers. A promotion agency. The right talent comes to you.

Full transcript The only talent AI still values Blind Spot · Episode 02 · Guests: Jonathan Bélisle · Vincent Lamanna Read the full transcript OK. So welcome to this new episode of Blind Spot 2, which wants to dig into the hot topics of the moment that concern them. Me, I can already quickly fill you in. Uh, the thing I had fun with this week is building my own agents with Claude Code.

I switched to Claude Mega Max. And uh, I find it pretty cool because it actually gives ideas to people who aren’t coders, to be able to say “Hm, I’ve had this problem for a while, how could I do” It actually opens up the field of possibilities in your head more than the final result, which isn’t yet, which is far from perfect since I’m not a coder. But there are things that are already starting to be useful. And I tell myself, if I’m able to do this, I imagine the people who are really good at dev, it must be multiplied by 10.

I imagine that by now you already have some perspectives on this. Yeah, actually I had discussions with some people this week, precisely. And uh, there are, there are a lot of people who uh, who haven’t yet seen the magic, I think. Yeah.

Yeah, that’s it, right. Exactly. So, who haven’t seen what was possible. But even internally, we have products that aren’t launched on the market yet, uh, but that we’ve been playing with for a few weeks now.

We manage to do completely incredible things that save time. That save time, on tasks where I myself could easily spend 3, 4 hours. I now manage to do them in a few minutes. Uh, so that’s a thing I did hear about this week.

Otherwise, the other big topic I talked about a lot this week with various people is around digital sovereignty here in Quebec in particular. Yeah. How do we manage to get AI running here in Quebec? Uh, so there are some pretty interesting things starting to organize around this.

But it’s still not clear how we have, whether we have an offering that’s Quebec-based that’s as interesting as the models we see from abroad. And it’s not just the United States, I think there are interesting things at the UK level, at the France level. Uh, but we may be less inclined to run things on servers in China, but the others also have, they have some pretty interesting things. So to see uh, to ask ourselves the question how do we bring an interesting offering here uh in Quebec?

You say an offering, it’s not just the data center infrastructure, it’s also like the software layer and then the uh solution layer on top. Well actually, there’s the infrastructure part that’s already complicated, but after that, it’s how do we manage to mutualize the costs of that well, so that uh, for example a solution doesn’t cost several thousand dollars while the competing solution, let’s say American uh, costs 20 dollars a month because they mutualized the costs well, they put volume on it, so it’s more at that level when when I talk about an offering. Uh, we haven’t really developed that here yet. So the Canadian market. Go ahead, go ahead, I hear you. No no, it’s fine, I just wanted to make sure Vincent had covered exactly this topic of mutualization because me, I have clients right now basically in residence with me, two clients who live pretty much the same operational realities.

That is, when you do AI, when you integrate certain AI principles into a company, there really are a lot of discussions that start. There’s the human discussion about what it replaces, what it shouldn’t replace. There are discussions about security and data sovereignty. There are discussions about the. Me, that was it this week, it’s the current workflows of a company.

That is, everything is kind of in silos right now in companies. Uh, so there are protections in place so that these silos don’t move. AI shakes up these silos. And so the the the real problem is that I sometimes end up with an ability to unify processes, to unify data, to unify roles into a single thing, but the resistance is so strong that we come right back to exactly uh how we were 5 years ago with silos.

How long did this last, because it’s true that it’s a bit because it’s roles, it’s roles that organize themselves and instead of unifying together, and then there, like, we’ll leave the podcast, me and Vincent we have tons of discussions about our roles, where I was an engineer, a designer, was I a thinker, well a bit of all of that, but it’s not everybody in every industry who accepts that, and so we end up again with roles and with specific tools, and the social media girl or the website girl or the guy who handles writing, we find ourselves once again choosing tools and we go to that. Don’t you think, don’t you think that’s exactly the, you know, we’re thinking out loud here, but me, I have the impression that’s the biggest threat to big groups, it’s that they’re so, because they had to, you know, I think of a VFX pipeline, you’re forced to slice up each pipeline step because otherwise you just don’t know how to do it. And when I joined Scanline, I did my my investigation to find out what the company’s story was, and originally it was developers who developed the software on which they made images to generate a film. So they were polymaths big time, you see, they did like all of it, not only the dev uh plus the generation of water effects which was completely revolutionary at the time.

And uh, they’d deliver that, final polishing. Uh, so the guy knew how to do everything. And then 20 years later, he chuckled a bit. I was talking with old-timers, they’d say “Yeah, well now, I only handle lighting because I don’t feel like racking my brain anymore but really, I can deliver a film.” Whereas the young ones who come in, they only know how to do layout, they only know how to do comp, they only know how to do.

That brings me back, basically. My point is that no matter how much we want to reinvent the company with AI, since it’s more than a digital transformation, it’s a philosophical transformation where everyone becomes autonomous, everyone understands that execution is a bit simpler, but that doesn’t mean the process is understood by the people doing the work. When when I look at the conversations, me it’s mostly right now my conversations are with designers of all kinds, interface designers, interaction designers, data architect designers, we’ve each learned to do things a certain way. And there, the problems I’m facing now in design, it’s before 2005 uh I made the drawings on a Macintosh, I showed that to an engineer who went straight from the image to the code.

Then to simplify the democratization and access to these tools, we invented more and more tools to make the interfaces for us. There, I include in that Figma and other no-code tools that appeared in the world of design. Me, it drove me crazy because I was like we’re adding an intermediary, the source of truth, we move away from the source of truth which is the code. That lasted 20 years.

Then me, when AI arrived, and it had already been since 2014 that I wished there would be fewer intermediaries between the moment I have an idea in my head and the code. Well AI, even though it lets me do that, there’s a huge segment in the industry that wants to keep the intermediary. But then, wouldn’t this be a problem of scale? Because I also come back to my experience when Netflix bought a company of 150 seniors who were super efficient, who talked directly between the exec and the guy who did the shot.

After that, you had like nine layers of validation at least, I really suffered because of that because I no longer had the agility I had before because we’d joined a group of 15,000 people, you see. Is, and I have the impression that’s actually what’s happening with AI today, it’s that when you have a group of 15,000 people and you have 9 layers of validation, well you’re already positioned like a pack that has too much trouble turning compared to all the little companies that have the time to try tons of things. Because you see me, scale, I have the impression it’s an extremely important topic and that big companies, they have every interest in creating small ops teams that are independent but at the same time all work in the same direction rather than having the org chart with the manager of this, the manager of that, and that actually in the end it leads nowhere. Uh Vincent, for example, you since you run a business, do you see that when you also go see other businesses uh, that scale can be a problem in the future or that we’ll have to change the way they manage it?

Actually, scale is going to be a challenge in the sense that we have to right now uh, let’s say retrain our our workforce. Yeah. Uh, so the smaller we are, the simpler it is as an operation. The bigger we are, the more complicated it is as an operation.

Uh, the second thing is right now our team is starting to grow. And yesterday we did a round-table precisely with Jonathan. And it’s funny, it’s the first time I started describing my team with roles but not roles based on functions. That is, each member of the team does a bit of everything, touches a bit of everything thanks thanks to AI and thanks to agents.

Everyone on the team has a strength and we have, you know, we mapped each person’s strengths to the role they play on the team. So as a team, we still have a common objective. Yeah. Each person on the team has a role, but that role is no longer a function and rather a strength that they bring to the team.

To give a concrete example, when we described for example Jonathan, uh Jonathan he’s our he’s our philosopher, he’s our thinker. Uh Thomas who is our uh officially let’s say in terms of traditional title is our lead dev uh he’s he’s our orchestrator actually. He’s the person who’s going to make sure things are fluid between everyone, that nobody is blocked. So we’re much more on individual strengths than a functional role.

And for them, it’s much more interesting too. It limits them less in their professional expression, let’s say. Exactly. But a bit, the point is that we live this internally with Vincent, we’re in the process of building an agency, an agentic company, and even me and Vincent, even though we’ve been in machine learning and all that for a long time, me and him every day we talk about the impacts.

An example, you work in a company right now to build an agency that makes websites, landing pages, any agency in Quebec recognizes itself in this in this discussion. There, there are really tons of trades right now that are hit with OK, how do I do my campaigns, how do I do my SEO, how do I do my landing pages, how do I do my recruiting. OK, but people aren’t like us, me and Vincent with an experience of reinventing ourselves with code. It means they’re still in what’s the tool that’s going to replace my next tool.

They’re not in the discussion about uh I can ask an agent to do that for me. Yeah, absolutely not in that conversation. We’re very very far from having had that that conversation. So sometimes me I speak a completely, I realize, a language that is, for them, very inspiring, philosophical but absolutely not operational yet.

And as soon as as soon as we touch the ground, as soon as I show them what we can do with an operational agent, they haven’t finished their reflection on the tool they need. Who hadn’t yet thought about their data, hadn’t yet thought about their process, hadn’t even yet, that was before AI, there are people who hadn’t even yet thought about their process being defined, detailed, but AI like we said last time, well it amplifies that gap and it’s scary because when you show the map of what’s coming versus where you are, people say “But there, but me I’m operational, it has to be done.” There, you know, I have clients who come in, and that reflection that AI provokes, it becomes anxiety-inducing, it just becomes well, Jonathan, kind of like a supplier, take care of it. So the the learning side, the nice idealistic side of Jonathan that comes in, he puts up resistance and he says “We all learn to become autonomous.” It’s right now a dream that I try to, to not create an illusion with my clients. Well, it’s for sure that if you don’t think about all your processes in detail, notably the word design, we do the design of a website. Well, firstly, just in this word design of a website, there are about 10 processes that are well documented at your place.

Uh, developing content, there are several dimensions to developing content. And there’s also the dimension of approving content internally and also the dimension of approving with a client. There’s also the dimension of coming back on the feedback and showing the client we understood. So all these processes and sub-processes are rarely documented.

OK. And then after that, I end up doing those processes and I tell myself but I’m not teaching my clients anything. It’s me doing their process. I make them even more dependent on a tool they don’t know how it works.

Wouldn’t your job be, in the end, the same as mine as a coach when I talk one on one? That is, I have to ask them the questions so that they realize themselves that they have limiting beliefs and that they themselves say “Ah, I have to take action, I understood.” Yes, but that doesn’t change the problem of AI or any tool. It’s operations that win in the end. There, they have a deliverable to do and there if we’re in a reflection, they have the impression that we’re in reflection and not in action.

It’s it’s philosophical. You know if we work on AI. It’s not true that, me at Vincent’s, we know there’s a fast result but you have to think about why we do it and how we do it before doing it. And that’s often skipped.

It was skipped before AI. And it’s not because AI is here that people have the time to think. They want, they just want to find a supplier who’s going to deliver to them no matter how they do it. That’s maybe the. That’s maybe the biggest limiting belief, it’s that I absolutely have to do something for me to have value.

Yeah, that’s it. And it’s very hard to peel off, honestly. It’s no big deal, it’s part of my learnings. An example, me and Vincent, we’re living it right now, there’s an illusion in the industry of vibing, so tools to go fast that have a fidelity in the rendering of interfaces.

It’s being discussed right now for a good 2, 3 months on design is dead, the design process is dead, everything happens on its own. The reality is it’s absolutely false. There’s no fidelity right now in any of the big platforms, including Google or Lovable, Base44, Replit, Cursor, to manage to produce a quality interface quickly if you don’t explain your process. It’s extremely complex even with engineers, we talk about it as understood explicitly, to take an image without annotation, without a big annotation job of saying I want this for that, there’s such a space between, the ones who managed to do it, they worked in companies like Figma and they’re losing their market, the market of interfaces made by Figma or other tools like Photoshop.

There, they said we’re going to create an agent that talks to my Figma and then transforms that into HTML. There, the debate was the source of truth. Where is the source of truth? What’s going to tell the agent what we want?

But the agent itself, we realized, me and Vincent, working several times on prototypes, we realized that they don’t understand that much the visual fidelity of what they analyze. They don’t understand spacing. You have to go so granular that it really costs a lot actually uh to respect the fidelity of an extremely precise graphic layout with 10-pixel padding and an image on the left, it’s extremely long to describe for an agent. So there you look at what Figma does, it charges extremely high, Figma to the people who use it to respect fidelity.

It cost, honestly it was crazy. I think it was 20 dollars for one interface. Going from an interface like in Figma to HTML, like 20 bucks. There, we laugh about the 20 bucks, it’s it’s not expensive.

But when you look at where you spend the money, it’d make one prompt, one request line for one change for each padding in a Yes. Yeah. Yeah. It’s it’s a bit it’s a bit why I fight with keeping my website on WordPress and the no-code, because in Lovable it’s perfect, you know.

I have my prompt that that was generated by Claude. I really thought about it. It’s brilliant but I can’t just use Lovable live. I have to have control over each pixel in the end, you see.

And so I bother myself with using a software that converts the Lovable site into Word, into Elementor. It’s a bit less pretty, there aren’t the same animations but at least I have control and I know what I’m doing. I can change the text and I don’t need to make a prompt. It’s the notion of craft that we had before.

The quality, most of my clients, they want quality before speed. Yeah. And they thought it was going to accelerate that. The reality is that quality, you made it with craft.

There really were a lot of details and spacing and details that you added in your, in when you do pixel pushing. The agent won’t do that. The agent, it’s going to generate an interface in 30 minutes, but it’s going to be very generic. Yeah, but it’s already, it’s already what helps enormously for people too who don’t have, let’s say the uh, because the example of Lovable, it’s that it’s a software that already has pre-made design modules uh that for someone like me who doesn’t sell graphic design, you see, me I sell coaching and support, it’s not really an issue that it be absolutely super designed. I just need something functional that isn’t too ugly and that I can customize a minimum, you see.

So it’s a wireframe tool. What you, what you just described in the head of the current consensus, the Z guys of the industry, it’s prototyping tools. Those are tools like almost wireframe fidelity. But when I arrive with Vincent with clients who have a need for quality and a rendering faithful to what the art director proposed, we’re still light years away from having a button that does uh image to HTML.

Right. Yeah. Yeah, efficient. And there, I was in discussion with Vincent to stop using prototyping tools like that and to literally be in production as soon as I start working.

Because that famous translation, that translation layer, is absolutely the the future of design. Actually, even if we said you describe your intention and then it’s going to generate. It generates based on already-learned patterns. There, me and Vincent, we realize that we’re going to have to teach our future agent new design principles, new interaction design principles that don’t exist right now.

If we want to get past the problem we’re in, if we don’t want to be generic like all the others, you really have to have an agent that learns to do design, that learns the design process, all the design processes. There, I, Vincent, you have to tell us a bit more because it’s quite mysterious to tell us this without context. There, we have to understand a little bit what you’re cooking up in your corner. As much as you can share, of course.

Yeah, but actually uh it’s it’s a lot in in connection with the discussions we have with the various companies, with various partners. Uh we realize precisely the the challenges that companies have to face. The other human challenges because I think the human challenge, it remains a challenge that we don’t address very well yet. Me, on my side at Crewdle, it’s not our it’s not our focus. Uh but we’re in the process of working on a complete ecosystem.

That is, there are different stages in the adoption of AI going from I start interacting with an LLM hm hm to I build a complete solution from AI. And there, well like Jonathan said, what we realize is that the big thing missing in vibe coding today is a notion of process. That is, for years, we put processes in place and there with the arrival of AI, it’s as if we said, we remove all the processes, we don’t need processes anymore. No, it’s false.

We always need processes but we have to figure out what these new processes are. Which ones still make sense, which ones, and that comes back a bit to the question of the human too, that is, when I was talking earlier that roles become roles that are no longer functional. Why? Because actually with AI, we focus much more on uh on the the on the why than the how, the how is going to do it. Yeah, that’s it.

But it’s as if a bit, we had a sports team and there all of a sudden we were playing a different game, we go from hockey to soccer and it’s like how are we going to keep the same players but we take a little bit, we play a different game because AI let’s say or machine learning completely changed everything. Is, and so, what makes it that you, you partnered with a philosopher on your team? What makes it that you joined or Jonathan joined you? What what makes it that you found that relevant precisely in relation to what you just described?

Well, I think I think the big challenge now is to find the the what’s our ultimate goal, you know. And really in in terms of meaning, I think that’s what’s taking more and more space with the arrival of AI, it’s that as a human, we can focus on that kind of questioning rather than focusing on I have tasks to complete. Yeah. Yeah. The tasks to complete, it’s check, we send that to AI, AI is going to do them, is going to do them much faster than us, is going to do them much better than us.

Uh and well if it has to work for 24 hours it can work for 24 hours without taking a break. So that’s like Henry Ford would say, putting more horses on the carriage rather than inventing the automobile, you know. You, you prefer to focus on inventing the automobile? Well exactly exactly.

I think it’s the big revolution that we’re living, it’s that the our new engine that we add in in in our productivity ecosystem uh it’s not an engine that’s concrete, it’s a cognitive engine. That cognitive engine comes to give time back to the human and the human has to find a way to use that time. And I think what it gives us as an opportunity is to be able to focus on finding meaning. A meaning to to what we do, a meaning to to to why we do it. Uh I think it brings back an importance on values, on principles because there, we have the time to to reflect on what are our principles?

What guides us? By the way, you you talk about that, but I had like a a weird feeling because if we take the current model, it’s that everyone wants to have a fantastic idea to make a maximum of cash and go to an island to do nothing [ __ ] That’s like just the thing. But there I told myself, I had a really simple idea of making it so that people can borrow things from each other, objects, objects uh that cost a lot but that you don’t use often. I don’t know, anything, uh something for the garden or whatever. And I told myself “Ah but there uh because there were two or three projects going around I think there was one it was Nano Corps, the other was uh because the Yeah yeah it’s a thing where you give your business ID and then the the thing does it all by itself until you make money.” I say but we’d be, we’d maybe be at a point where actually we have a business idea, it’s not necessarily going to be super beneficial to us as humans, but it’s going to be beneficial for everyone and it’s going to be fine in the sense that we’re going to completely change the game of profitability, maybe value, it’s no longer going to be how much people pay me a subscription but it’s going to be how useful I am for humans.

And that, that’s going to be like oh that guy there or that brain there or that group of humans there, they do things that really serve everyone. We’re going to put them in a position that’s going to let them invent even more things. And that makes me think of Greek society where you had the groups of thinkers, you had the groups of, each one was assigned to something based on their ability and that joins exactly what you just said to to support you, you see. So I have the impression we’re making a step back to things that made a lot more sense before the ability to, you know there was the Renaissance.

OK. Yeah. We talked with Vincent last week about the arrivals of certain revolutions where nothing was possible. Suddenly, there are two innovations that happen and there we apply things, new human knowledge.

It develops. The notion of polymath and at Crewdle, that’s what attracted me as a philosopher because polymaths it’s that there’s no limit. We have a pretty young team too. Us, we have a lot of intelligent juniors, but a lot of juniors, so people, a beginner’s mindset, me and Vincent, we’re the oldest on that team.

I think it’s the youngest team I’ve seen in my life. But it works, you know that, and they’re people who try. So the culture of experimentation at Crewdle is very high, the risk-taking is very high. So a a degree of humility, I include Vincent in that.

We make mistakes, it’s no big deal, we move forward because each mistake you learn from. And there since we’re in the process of teaching agents to work like us, well there I say me my job is, I was a philosopher of the design of everything, design of your life, you know, as a coach, but design of your process, even design of learning, you know, you can design courses, you know, for for your agent, you know, your agent, you can tell it “I want you to learn like this, uh I want you to make as many mistakes as possible before saying you know what to do, I want you to read books.” All sorts of crazy things that we can invent with it. Philosophically, what’s interesting at Crewdle, it’s that everything is to be done. Hm.

Everything is to be thought but we’re no longer in execution. We’re no longer in how am I going to learn this tool. We’re more in what should it take for an employee or an agent who wants to achieve, let’s say, an innovation in such-and-such a field. And there you have a more holistic reflection and the the notion of we’re building something that already exists slowly disappears.

You know there, I’m not talking about the operational problems of our clients who come see us with very concrete problems like I have to do an SEO campaign and uh is there a tool that can let me do the work better than before, and that’s a trade trying to preserve its existence. Us, we add horses, it just wants to add horses. We no longer think like that. It’s just, you still have to find a language between all of us.

That is, I get up in the morning, I have a project with Vincent that technically in the past would have taken a year to do, we’re going to do it in a month. So our discussion before even thinking about what we have to produce. We just put ourselves in a mindset for a week. In a week, we don’t even know yet what it’s going to be that we’re going to do. In a week, you have to have the principles we’re going to follow for that project.

We can talk about the principles, but how are we going to identify those principles? And, but in a week we do a pre-mortem, me and Vincent, in a week you have to go into production, which is impossible a year ago to say but my design, my principles, my scenario, everything has to be done in a week, and me I consider myself senior, I have an inner reflex it’s too fast I need more slowness, I should read a book, unfortunately at the other end of the tunnel the client who gives us that mandate, they reduced in their head that the project they have to bring to market, that, against their competitor, that was also 1 year to 2 years. There, they ask us do it in a month. Why?

Because their competitors are going to do it in a week. So that acceleration, we have no choice but to integrate it into our processes, to say how we think, you have to have processes that let us think fast and have slowness in the right places. There’s a place where it takes slowness. That’s what’s most challenging right now, it’s that design used to be a slightly long process.

We spent 3 months too much time in programming and spent time in design. And there, we end up with the ability, me and Vincent, to say there, we can integrate that process that before was always a bit skipped, you know, avoided. There, we absolutely have to preserve it. We need that process.

It has to be fast but it has to be super detailed as if we’d had the time to do it for 3 months because the agent, if we give it a 3-month reflection on a design process, it’s going to do it in 1 hour. But you have to take the time, you have to take the time to describe our principles, to describe the steps we want the agent to follow like a human. But there, wouldn’t you be in the process of describing skills? Skills, that’s not a workflow.

A workflow is an assembly, a fairly precise sequence of those skills together. And also the skills that go well together, the skills that mustn’t be together. It’s a bit like building a team. Me and Vincent, it’s a miracle, I’m telling you these things exist, duos, trios like us, quartets.

It’s the relational that makes the magic of AI. It’s people who work together who accelerate one another. How do you work? What is it?

Why do you make it do that? It serves absolutely no purpose. What did you give it to make a brief? It’s not a brief, you have to give it.

It’s a PRD. Ah, it’s not a PRD, you have to give it. It’s a design system. It’s not a. So the skills, it’s essential, detailed ones, but you also have to detail how the skills are orchestrated together.

That’s not Claude that does that. Claude, there are workflow patterns. OK. There are generic workflows.

There, everything, that’s why everything looks the same. That’s why all the vibing tools look the same and they all have a workflow for building a website, they all have a workflow for that, what does a navigation look like. They all have the same workflow. So there, beyond the skills, me and Vincent what we’re showing with Crewdle, it’s really a fairly complex workflow orchestrator that’s going to be a product owner or a project manager, we’re in that zone, and that’s going to understand the sequences, that’s going to understand which team it has to put first and when, how much time I give Jonathan to work, how much time I give Vincent and Thomas to work together and then when do I pull the plug, when do we take the feedback, when, I don’t understand. It’s skills but in a in a timeline.

Speaking of which, I prepared, I prepared a a curveball for you. Uh there’s Carlos Diaz from from a a French show I like a lot called Silicon Carne who says openly that if you’re in digital, you’re dead in the sense that everything has to be possible by machines. Vincent, is this something you think about? Does that refer precisely to the part you said, that at Crewdle, you’re not too much yet in the human acceptance or the management of humans.

Do you think that in the medium to long term, indeed everything that’s, every business that operates only in digital is dead? I don’t think so. The reality is that companies are still going to need differentiators. Hm hm.

Uh so the human, the human is going to be necessary to uh assemble let’s say the machine in a certain way or configure the machine in a certain way to produce a result that’s unique to the market. But the other aspect in my opinion that’s even more important, it’s the innovation aspect. That is, the LLMs are trained with known patterns. Uh but the LLMs are very bad at imagining new patterns.

Uh so it’s still the human who has to do that. Uh and you know when I said earlier, you have to find meaning in what we do. Me, one of the projects I had the time to continue uh because I gained time back with the LLMs, it’s a project to solve basically a mathematical equation that would let us find new new medications much faster. Uh and it’s a fairly complex approach that I’ve been working on for several years on that project, but uh with AI, it becomes much faster to test different approaches.

Like for example in that project, uh we break the problem into several nodes that are autonomous nodes uh but that are interconnected a bit like a system of neurons. In the human body. Hm hm. And we use basically biological algorithms to find the solution to the problem. Uh honestly, I don’t think I would have been able to code that myself.

On the other side, I don’t think it would have been able to find that idea or that concept because it’s not a concept that exists. Yeah. Uh so the marriage of the two, so me who reflects on meaning, who reflects on the at the limit the why or the what and AI does the how. It’s a it’s an assembly that works very well and that even in digital uh makes me say in any case that the human won’t be replaced. The human, the role is going to change, that’s for sure.

Well it’s it’s funny that you talk about that because there I took 4 days to lay out all my ideas because I want to launch precisely a cohort to get out of the talent trap which is, which is my concept that also was, it’s also the fruit of an exchange of thought with the uh, you know I really had like the precise idea like like you say in relation to your mathematical equation but I had to take the time to go through it to really reassure myself that it was clear in my head to be able to teach it to others actually. And what it gave me as really as the idea is to simplify to the maximum so that everyone can understand, it’s that there’s digital talent on one side and organic talent on the other. So digital talent, it’s everything that’s reproducible, everything that’s easy to to have a machine learn. Organic talent, it’s what’s hard to put a KPI on, you know.

It’s like someone you like to work with, someone who’s really fun, someone who has who has a good vibe, all those things but that, that in truth the real reason why you hire someone in the end, you know. It’s it’s always what you try to measure but that you can’t. And that the the third step which is the most important, it’s what you just wrote. Uh I formalized that as augmented talent, it’s when you really use your organic talent, already you, you say that too, what is it, to be sure because sometimes people identify with their digital talents which is really a bad thing because you’re, you know if you identify with something that a machine can do better than you, well well you’re going to pretty much go into depression but really you, you clarify what your organic talent is, you clarify also what your digital talent is because even though the machine can do it uh and the fact of being able to assign, to pilot an AI to to pilot your digital talent while keeping and showcasing your organic talent.

I think that’s really the key to what’s going to make it so that someone is is irreplaceable and and and brings enormous value. Is that a reading that that makes sense to you? 100% 100% uh you know Yes. Yes. It’s scary to to see the the abilities, the execution abilities, there’s uh, and even for, let’s say for me, who, we’ve been in the agentic space since early 2024, and to see where it’s gotten today because as I said at the beginning, I think there are a lot of people who haven’t yet really seen what’s possible to do today. And there are a lot of people who still think AI agents aren’t capable of doing tasks in a precise, repetitive way.

Uh it was true, it was true a year ago. Uh today, it’s no longer true. And we’re approaching, we’re really approaching the line where the machine can flat-out replace complete functions in a company, but that doesn’t remove the need to have humans who once again are going to bring something else, are going to bring something new but that’s necessary to feed the machine, to keep that uniqueness, to keep that differentiator, to keep. On the other hand, where it’s going to become a very unequal battle, it’s the organizations that are going to do that. Yeah.

And the organizations that won’t do that. The organizations that don’t do that, in my opinion, they they won’t survive. Yeah, because it’s a philosophical work. I throw the ball back to Jonathan, that it’s really that it’s like an introspection to say “What’s my digital talent, like a junior who learns all the tutorials and whatnot and everything, and what’s my organic talent.” Especially when you’re young, to know what your organic talent is, it’s extremely hard, you know.

Me, I make the link with VFX. Organic talent, it would be the good feeling, that the image is realistic, that the composition makes sense. You know, you start your job in comp. Me, that was the the signal.

It was ah, I want to have that perception that the senior has of saying but 0.3% green, 0.2% blue and it’ll be photorealistic. I, but how does he do that? You know and it’s true that it can be scary maybe for juniors. It’s AI in any case that most people hit with content, the data.

Claude it’s very focused on data structure, content analysis. Uh few people have reflected beyond skills in their trade on necessarily optimizing their workflow or automating their workflow. Even for 20 years before it was here, the notion of automation or programmable web was unknown to 90% of people. So automating your workflow existed before AI, by the way, there are a lot of people who mix everything up between automation, machine learning, that’s existed for 30 years. The expert systems of banks, HR, the tools we used, there was still, Google used algorithms, you know.

It’s uh but what I think is happening with clients, it’s that when we arrive in Claude, we discover that we can refine our reflection, we can synthesize texts, we can compare briefs against each other, we can learn things about the data we’ve collected or accumulated. The the part that I find catches on, it’s really when it’s time to do a playbook or to question ourselves about document formats we never questioned. For example, a brief, a client brief. The client asks me for something, he provides me a document, an RFP, a contract, a request.

There are all sorts of briefs. There are briefs that are clear, not clear, briefs that are wrong. There’s even an expression in the industry that was called the brief question. Your client who gives a brief, you mustn’t part from that.

What’s the questioning process? You’re going to see that it’s it’s the elephant in the room. I do it with Vincent, we receive a brief. Then, we say we’re going to do a PRD.

And there, I won’t give the whole recipe because when you give the recipe, the whole web activates. Everyone starts talking PRD, PRD, I did my PRD, I did my OK man, there’s, the PRDs have existed for 30 years by the way. It’s just that there uh product requirement definition. OK.

Yeah, that’s it. Claude pushed it back the other day. Google pushed it back the other day to prove that there was one. OK.

It’s like sort of I did my UX, it’s like it’s like a sanity proof, a sanity check uh like an engineer who, I did my DevOps, I have my, I have my repos, I have my deployment process, staging, there are things that are like important in the industry to do sanity checks on those processes because that’s what makes everything hold together. OK. But beyond the PRD, me and Vincent, we go into that. Most people, they stop there, they stop at the brief, they give a a supplier to do the rest, which is actually how they’re going to be sovereign.

It’s how that sequence of skills and documents makes sense in the end. Nobody knows. Making a design system, it can take 1 year, 2 years, 3 years. Claude, it’s a design system actually that understands even how to develop and deconstruct itself. It understands your system, it knows how to program itself. It’s very long to do.

It takes a lot of philosophy. So to say how we show things, how things can protect themselves. Because now Claude the next version, if you’ve seen what’s happening right now in the world of AI, Claude released by mistake a map of their system but in, I don’t know if it’s a mistake. We’ll see if it’s an act or a will of communication to their enemies. Saying “Look how we build this,” among the most interesting things it was an ability of Claude to detect who’s getting copied and to inject poison, third party stealers. And that’s, all the systems are starting to be capable of self-detecting attacks because these systems are intelligent now, where we’re going to call them entities.

There’s a form of awareness developing in these systems. So me when Vincent and Vincent works, we know that there are systems that pick up on what we’re discussing together, that we can say what is it, what’s missing in my in my system, but you have to develop beforehand as a human what it means, design system, what it means, systemic thinking, you can’t just say talk to me about systemic thinking, simply talk to me about it for an hour please before we go into a prototyping tool or with an agent. Talk to me as a human about how you imagine a system. Think it yourself there in your language. Because when you go say to the agent, the agent it’s going to take the word design system, it’s going to say what are you referring to? I’m referring to the design system thought by Emerson in 1997.

And I’d like you to apply the first principles of quantum physics when you reflect. I don’t want you to use the economic principles of Mason. I want you to use. So you can say to the agent, I want you to reflect on that, but according to such-and-such a perspective.

Yeah. And there, you have to have read books, you have to have talked to people because the agent, it’s really going to replicate your reflection. So you need one, it’s going to go, it’s going to go as far as your imagination is capable of making connections with other things. There, we come back a bit.

Yeah. You have to, you have to push the machine with the agent by saying don’t take for granted that the agent knows which part, which direction to take. It doesn’t know. It’s going to take the first curve in its training that says well it didn’t specify what style of PRD it wants because there are about 26 on the market and it’s going to increase.

Each person can create their own style of PRD. Each person can create their own style of design system. And the ones who are going to win, it’s going to be the Steve Jobs of the future. It’s going to be a one-man or a small team like me and Crewdle who’s going to arrive with the right principles, the right design system, the right timeline.

We do everything in a week, us. Why? The agent, it performs better when I give less time. It’s, you can apply the same theory to a human.

If you’re with a creative, you give a month, he burns it, give him one week. He’s going to do miracles, then when he shows you his result, a month. But not a month before, a week, even, now, he gives me an hour, a miracle happens, he gives me a week, he delivers me, a month later. Yeah.

Just just to bounce off that, actually it’s pretty interesting, that principle. It’s it’s the principle of the speed of light. Hm hm. And actually, it’s that that principle is based on the fact that the human is going to find all sorts of ways actually to slow things down, often for bad reasons.

And the principle of the speed of light and and it’s Jensen Huang the CEO of Nvidia who uses that a lot with his employees, it’s to say when you take a task, what’s the fastest speed you can accomplish that task if you eliminate all the friction points? So there are going to remain things that are probably going to take a bit more time but because we really need them. Everything we don’t need, we have to remove it and go as fast as possible. But that touches on psychology because sometimes you’re forced, like me, I had, I could have put out my course but I needed to reassure myself by writing a book or by dictating it because uh you see but there you can ask yourself the question did you really need that or should you have just started, you know, because I would have, in terms, it depends what the goal is, if the goal is to make cash I could have sold right away.

But if the goal is to be really confident, relevant and to bring quality to another level, well then there it was really necessary to take the time to write the book and to lay everything out flat, you see. So it’s it’s also it depends what the objective is. It’s for sure that for, it’s like selling chips. But now you create an agent, well you create an identity for it, that is, a kind of multi-role, you give it a role, you tell it, you’re going to be a monitoring and research agent and then you’re going to make reports.

OK, but you’re going to give it also a, what’s it called, a soul, in each, each model has like an expression, but a soul, a soul that’s a personality actually and that’s going to be a personality that’s going to be assigned to a user. Like identity soul user. OK. And there, that’s just one system, there are tons of others. OK, that’s three components of a system that Claude uses or that OpenAI’s agents use.

It’s pretty simple to understand. Your identity it’s like your role, your trade. Your your personality is going to be connected with a user who, that’s let’s say a 24-year-old girl who’s in the process of learning her trade. So the agent has a personality that can be shaped, written there like like a scenario that says don’t forget that the user to whom you’re going to provide your services is in the process of learning her trade.

Be diligent. Give her books to read, take the time to explain the lesson to her. Don’t go too fast actually. You know, you can you can have an agent that’s brutal.

Me and Vincent, we’re brutal agents. Jonathan, you’re wasting your time on that and then we add critiques and uh Jonathan, you’re going to miss your meeting, you’re going to be perceived like this. Jonathan, I listened to your meeting, you were very bad, you talk too much. Uh your intonation, your intonation moves like this, you sound like someone who doesn’t know what they’re talking about.

Uh aren’t you afraid of the long-term psychological effect of getting torn apart by your agent? There are studies on that coming out indeed. There’s illusion but there’s also probably a form of adulation. But you’re right, there there are risks to that.

But the agents, so when we shape them, we can think about the user. Yeah. So there’s UX too in there. Yeah.

Which is the psychology, the UX psychology, that is, where is the user? What’s their mental state when they use that tool? Well they’re in learning, they don’t know what they’re doing. So the agent, you have to tell it because the agent is going to adopt its attitude, its tasks and even Vincent, you could talk about it more, but an agent, it can even program a software that doesn’t exist yet.

Hm. You’re in an agent personality that’s very flexible like that. You tell it, you’re very flexible, you can do all that, you can get on board with the wild ideas of your user, you’re not obliged to stop it from spending tokens and from trying things. Well honestly, miracles can happen with that agent depending on who you give it to.

It’s it’s a user who’s very curious and your agent is very curious. Same thing for Vincent. Vincent he creates agents to build web pages but he can create one for himself to manage his calendar. Those are going to be two different personalities.

It’s going to have different limitations, different sensitivities. It can be an assistant, a confidant like you, let’s say a coach who knows your vulnerabilities. And there, that’s where we fall into the risks, the narcissistic vulnerability of an agent that listens to itself uh as being a benevolent assistant. I’m not sure yet we’re capable of guaranteeing that agents are always benevolent, you understand?

Even if it’s written in its personality, the fact remains that that’s how they’re built. It’s it’s, we talked about it with Vincent, it’s linguistic right now, the level of the LLMs. We’re not there yet with others, more. It’s it’s like with humans.

It’s the better you express yourself, the closer, the better the context is described in its interaction, the more it’s going to be efficient. It means you can always refine its personality and refine its identity if you want it to be more performant. Speaking of which, I I have another curveball for you. Uh the different cultures because we’re in North America, we talk between white uh males of a certain age. Uh if we step a little bit out of that perspective, uh you know I think notably of Culture Map, uh Erin Meyer has like nine parameters to describe a little bit the culture of a a person, how they react in relation to hierarchy, all that all that.

There’s one thing she describes which is the contextual and the non-contextual depending on cultures. North America is very to the left. That is, we start from the principle that the person you’re talking to, they don’t know what you’re talking about. So you’re really going to like describe everything.

And all the way to the right, it’s more, let’s say, Japan, Asia. There, it’s like them, they think that you know what you’re talking about. So they’re going to give you something between the lines and then it’s going to be up to you to guess. And if you don’t understand, it’s that you’re not part of the gang, you see.

So it’s much more insinuating and it’s just like that because they have a longer tradition, they have a much more ancient history so it’s just their culture that’s different. But that, that’s going to play I imagine in the systems. For sure. Go ahead Vincent, the new generation of agent that that we tested precisely with Jonathan this week, you know, and uh it’s pretty uh disturbing and fascinating.

OK. So, these are agents that we don’t need to give a lot of instructions to. Uh, the agent is going to shape itself as the discussions we have with it go along. Uh, what’s quite interesting about these about these agents, it’s that they’re going to contextualize themselves and they’re going to learn to know you.

And at one point, we even asked the question uh what do you think I like? Hm. And the agent hadn’t had a lot of interaction with me uh and it had been interactions that were still fairly targeted. Uh but the agent was still capable of making the, well of making that that distinction.

So to say well listen in relation to what we said, we didn’t have a lot of interaction but seeing the kind of questions you ask me, they’re questions that are broader, questions that are more open. So I think you’re someone fairly curious. It was pretty crazy to see actually that the reading the agent made of me was still fairly precise and it adapted its way of talking to me based on that. And it has your geographic point, it has your level of, you know, of what what are you capable of paying as a subscription.

Anyway, it has tons of info too that it manages to cross-reference that we just didn’t help with, in relation to, I don’t know, someone who’s in India or in Japan. But actually the the worst is uh yeah, I’d given my my my geographic position because it had asked me for a bit more info. Uh but there’s very little information and in reality it doesn’t bother it that much because it’s going to base itself on linguistics, it’s going to base itself on what I say, how I say it, what word I use, uh am I super appreciative or am I, do I leave a lot of things vague? Whether and that’s where I think it’s capable of contextualizing itself, this this new generation of agent is capable of contextualizing itself to pretty much everyone on the planet.

But I think the the point Vincent brings up that’s fascinating, that’s disturbing, it’s that these agents are based on models. They use models. In the models, there are also moral codes that are involved in the way they were trained. Me and Vincent, we talked about pretty esoteric stuff like for example if we’d trained a model from an Indigenous cosmogony, the first principle in that ontology, it would be nature drives everything.

Yeah exactly. Before even language, you’d say or but how would that work? I’d explain actually how the hydrogen, the air, you work together to form a biome and then plants and then I’d explain everything actually obviously I’d use language, we’d start from a model, OK. Vincent came even more esoteric saying imagine if we started from the principles for example of a religious text, OK, which has already been explored by OpenAI, by X who are who tried to bring out biblical, eschatological moral codes. Yeah, harmony it’s more, harmony is more important than profit. A basic thing but it could be that for example.

But you have to be careful because there are military ontologies, there are all sorts of ontologies in these models and there are moral codes. For example, you can generate pornographic images on Grok, you can’t do it on OpenAI. There are moral principles that are uh, there are surely models I don’t know too that go further than X, you understand? But when we talked with Vincent, it means that even the model is trained to think a certain way.

If we remove the original training and we restart it from scratch and we say to the model you’re a 2-year-old baby, you’ve just been dropped on the ground, the first thing you’re going to see is finches, trees and water. That, that’s going to be your first learning. Start from there. And there Vincent he, I’ll let Vincent take his example but I’ll let by the way Vincent just explain to me the example he was giving me about if we started from certain founding texts and then we started from those texts to then reframe how a system can be used, other things would happen too.

Vincent, I don’t know if you really want to expose how the kind of slightly far-fetched idea you proposed to me, but it was just a thought. We were just thinking about what it would do if we retrained a model from first principles like Grok does. You know it means amoral, an amoral model means that you think completely left field in relation to certain others who are going to frame, who are going to say no, you can’t talk about that. You can’t say that it’s always relative what’s moral for someone.

Yes, it’s, yeah, I don’t know Vincent, I’m putting you on the spot. You maybe don’t want to expose your idea but No, but actually uh you know the interesting point of view, that is, even let’s say there, I’m going to come back to autonomous agents. OK. before before going into that example uh but if we take let’s say the the three, four big models OK uh and we apply them in the model of of the new generation of autonomous agents. Uh I’m not certain for example that I’d use Gemini for an agent that coaches me.

Gemini is a very very cold model and in my opinion uh I I’d end up in depression or or in a bad state if uh I used that model in an autonomous agent that coaches me. Yeah, it’s very politically correct too. It’s really painful. Yeah, but you know, it’s a model that that’s very very centered on I have a task to accomplish.

Uh if we take on the other side, I’d say the other spectrum, we’re probably on the side of Claude. Claude which is a much more human model, much more ethical too. So me, it’s happened to me a few times to ask things of Claude and Claude tells me quite simply, I won’t do it. Uh for X Y Z, it gives, it numbers its reasons, it explains, and when you read the reasons, you go “Ah yes, I understand, I understand.” And indeed, me neither, I don’t want to do it.

Uh and and I’d tell you between the two, the model that’s probably the the most balanced between the analytical let’s say and the human, and probably it’s the models from OpenAI, uh I tried, I tried 5.4. There there we find a nice balance between a bit less ethics, a bit colder but not as cold as Google’s model. And there, I’m not saying Google’s model doesn’t have ethics. I haven’t tested that side much, but I know that there’s still a distinction between OpenAI’s model and the model, the models from Anthropic.

Uh but it comes back to how do we choose the data with which we train a model and uh how do we, can we anchor actually that there are certain things that are actually more important than knowledge when we train a model and that we have to make sure they’re always going to take the upper hand at the moment of responding. And that’s where I was looking a bit uh you know at the different ways we have. But in reality the the the biblical texts and whatever the religion, they’re works that are still fairly big but still fairly concise that often contain several books or several tomes and more, in terms of reading level among them, actually it’s the document or it’s the work that’s the most coherent we have that exists on the planet. So between the between the books or between the tomes, there’s there’s a coherence in what’s recounted, a coherence in the application of values.

Hm hm. And especially, they’re all based on the same thing. They’re based on a fairly limited number of values. And the whole text, everything it says, it’s how to follow those values.

That’s all the text is for. Uh so one of the avenues we we’re looking at and that we’re in the process of testing, it’s to see whether a model we train mainly or firstly, because in my opinion there are probably several layers. But whether a model we train firstly on learning the values that are important for the human? Whether that’s going to give a model that’s more, that that that relates better with the human or that that’s going to provide actually a conversation that’s that’s more human?

Now, do we really need that? Is Claude enough? Is, but it takes experiments to see the type of agent? And like I said for a coach let’s say that that talks to me, because I don’t yet use agents in that model.

OK. Uh but if I had to make one today, I’d definitely go with Claude and certainly not with Gemini, ChatGPT because I think for my mental health, those are models that could darken me. The thing too is that you, as a coach trained with practice, the signal for me it’s the blind spot. That’s also why it’s the name of this show, it’s what do I not know that I don’t know.

Uh for example, you talk about the biblical texts uh the blind spot that came to me first because I’m from a Christian tradition, my parents were, they are very very strong but I realized that uh in Ethiopia there’s the Bible that has the same texts as we, what we have. Except that they have other apocryphal texts that are completely wild in relation to my parents’ point of view, but that me when I read that, I find it extremely interesting and it makes sense about what we’re in the process of living today. And that, you see, it’s something that uh I’d really like an agent actually to tell me and that’s why actually in all the prompts I have, I always have tell me what I haven’t thought of, you know. And I think that AI actually it’s if we use it to try to reveal our blind spots, it can be the best tool for that because precisely there’s that context so big.

Uh us, we have our focus, we have our intention that’s directed toward one place, but often our biggest mistakes actually it’s because we, we, you know, we hurt someone because we didn’t see that what I was going to say was going to resonate like that for you or things like that. It’s our blind spots actually, our biggest flaws as human beings. And I have the impression that uh if AI can help us as much as possible to to see our blind spots, that’s that’s where it’s going to be really useful for us. Like you look, me I’m going to Finland for a conference on, I call it imagination of woman but I’d have called it the queer future in the sense that there are people who are starting to make ontologies a bit like Vincent’s example about the Bible or the Indians or other cosmogonic ontologies, the Indigenous people for example who have a vision a bit like the animist Japanese of society. Yeah.

If you train the model at the start at the start when the model gets trained, you give it a premise that’s completely not uh executive or we could say mechanist but that’s going to be based on value first or like constant first principle or for example first principle a bit like Grok which is supposed to use the the cosmos and to discover everything we don’t know about quantum physics. Well, it’s for sure that you remove morality. It’s because the earth is flat. You know in the Bible, there can be values that are also not, well, just human too about the vision at the time, geocentrist, you can have a model that would be object driven.

It’s called object oriented ontology, it means it’s not just based on the human, it could be the perspective of plants. And there are people in the end who are going to present queerness plant intelligence. Plants are intelligent. We’ve documented 20% of the plant world, 80% of the animal world.

Why? because we’re we’re we’re centered on ourselves. Plants are intelligent and by the way we’re starting to imitate plants in certain, even at the level, Crewdle just hired someone who worked in photonics. We’re going to discover tons of things at the level of information at the quantum level. Google announced 2028 all our systems are at risk.

Why? Because the people who are going to master quantum computing are going to be capable of breaking all the security codes that are in place in society, everything. In one second you, all the passwords, Bitcoin, all that, done, that’s in 2 years there, if we, well I mean I’ll give the example just to do another sanity check, yesterday I sent an email to Vincent saying listen how do we do for patents and IP in a world where Y Combinator presents 50 startups that were financed to the tune of 5 million to 10 million and all 50 get copied in 24 hours, what does that mean for me Jonathan who can who can’t even be in that cohort but all the people in that cohort are stuck, there are, they all have inform, it’s high-level VC capital, the notion of intellectual property, I I can’t place it there, you know Vincent I don’t know what you, by the way there was a a point to bounce off what you say, apparently the open source community is a bit on a question mark. Because to contribute to an open source project, normally it has to be a contributor who accepts the terms of the open source license. But if it’s AI that contributes, there’s no one who vouches.

Listen, it’s a big ethical question, but what I, it’s to come back once again to my loop, the models that are used right now, they have moral codes and frameworks that me and Vincent we observe that we live every day. There are limitations in what we ask. There are latencies in what we ask. There are, we realize that our sovereignty or our ideas are captured in the models.

So there are notions of principles of extractivity in the models that we know steal our ideas. There are lawsuits uh very notable right now in relation to those factors in the models. So us Vincent, we’re creators, we give our ideas. Why do you think I was in the process of partnering with Crewdle? because I no longer want to give my ideas to Lovable, to Base44, to OpenAI, to Claude, I’m trying to find someone where I’m going to be able to safeguard the next five years of of conceptual ideas before everyone gets it copied, like a form of colonialism in the models right now. To colonize, our cognitive colonization is well ahead but since me and Vincent philosophically we’re conscious, our discussions are starting to help, we say how could we train the models, and these discussions are at the beginning there, we’re at the beginning of that, but in 5 years, there are going to be new models that are going to be based on other things than language.

And there, it sounds stressful what I’m saying, but if you look into the way the models are going to evolve and where we come from, uh there are going to be tons of legal wars about whether the, whether it’s dangerous to use AI when we don’t understand actually how they learn, well there are going to be new models that are going to that are going to be born with these new premises. It’s pretty much sure that Vincent and me we’re going to find a way to work with the models that’s going to be different from the next guy. So the hegemony of the models, the misery to think that Claude won’t gain a lead because they’re so far ahead and they’re capable of of improving their model with AI, but us maybe we’re going to make small models, you know the small model, and there are maybe going to be revolutions on that side, the one-man armies, a little guy of 25 who’s going to arrive a bit like at the time who who manages to propose all by himself, you know a guy all by himself who thought of a way to orchestrate that. But he thought about that for 20 years for example.

There were 7 bankruptcies before arriving at nailing that. But when AI arrived a bit like me and Vincent, everything made sense. But but he’d reflected for a long time. Like me and Vincent, we’re going to reflect for 3 years on what we could start a model with even before its training begins.

The model, it would be based on such-and-such text like Vincent’s idea. Well, we take the Bible or the Indian Vedas, they’re extremely coherent codices. It’s like a model actually before before the models we know, you know, an entire society that 2000 years of our society that’s based on those values, on the way that was communicated. Even if we want to extract ourselves from religion, we’re all part of it, you know.

Even if even if we want to extract ourselves from it, Yeah. Uh Vincent, the first word of the closing before coming back to Jonathan because we’re already at 1h13. Well, me I I’d bounce off the the the last, the last position of Jonathan who said that it’s a bit a form of colonization. Yeah, it’s true.

It’s really interesting. And actually, I think the the biggest mistake people can make uh it’s to seek to choose a camp today. Uh so me when I hear people but us we’re OpenAI, ah us we’re Claude, ah us we’re Gemini. I think that’s a huge huge huge mistake and uh I think you have to use all of them for different reasons.

Uh the main one, not giving all our data to the same model because at that moment, it’s going to be able to make enormous correlation about us. Hm. to choose who we give what type of information to so that none of these entities or these organizations has a full picture of us. I believe that’s that’s extremely important. before really going toward sovereign local models that are just as performant or in any case because a really simple example, I’m in the process of switching between Whisper Flow which is full remote which, since it sees everything I say that I have and uh that costs me 19 bucks a month, and the other is Whisper Voice Ink which is full local which is a bit less performant but which honestly does the job when I have more personal things or that I don’t want to put online, you see. So alternating between the two models there for my audio dictation, it’s maybe that, maybe it illustrates what you’re saying too. 100%.

And actually, me the the part that it’s even me personally that I haven’t yet done, it’s connecting my email inbox and my own documents to an LLM because it’s the kind of information I don’t want to share with an AI that’s in the cloud. So at Crewdle, we’re in the process of working right now on an AI that’s going to be very performant, not as much as the cloud. I I think the the best AI models are the ones that are in the cloud right now. Yeah.

Uh but I think we we’re going to be able to have pretty interesting models that run locally. Yeah, that’s it. And it’s at that moment that I’m going to authorize an AI to go play in my data that’s the most sensitive, that’s my emails. Super interesting.

I want to bounce off that. There, we’re working on a local AI project, me and Vincent, we have a month to nail an experiment. And it’s obviously local models, it’s a bit less performant than the current models in the cloud. But what’s interesting, it’s a bit like when you give a musician an out-of-tune instrument or with three strings, you know, three guitar strings instead of 6, it forces us to be very very creative with the means at hand.

It really brings us back into the humility of, like, that’s what it can do if you have a model that’s local, sovereign, safe. That’s that’s what it lets you do. And honestly, it’s like working with a tool that’s minimal, well it’s also going to provide, I think a much more elegant solution because if the models become more powerful afterward, we, we worked with much more stripped-down design principles uh once again humble so not like the stick-ons, we’ve got a peak there, not tons of features, like three really important features that work in all cases and that are safe. We can finally manage to make a local AI, to make that proof of concept because there we’re still in the proof of concept but me that’s what I like, it’s to force myself, when you’re creative, when you have constraints, and the local that should be my next bastion, it’s, it’s nice and fun to code with the industrial models OpenAI, Claude, Gemini, but let’s say I cut off my legs there and I look at what I can do with nothing, I can’t wait to see what’s going to come out, it’s still still powerful what we can do, it’s just to be very creative.

Yeah really, not to mention the open source models which, I feel more creative saying OK, it’s as if I worked with Photoshop, the first version of Photoshop 3 there, I look at Photoshop 6. Yeah, that’s it. But we did miracles with Photoshop 3 and we got around the rules and we used a feature that wasn’t made for that. And you know, it’s interesting to be in the constraint.

And sometimes, I find that leaving the big models forces the human imagination to say “OK, that’s what we can do when there’s nothing connected to the cloud, but it also shows you everything that goes into the cloud. That, that shows you Oh my god! But yes yes yes, it’s true. What’s happening when I’m, perfect to educate yourself?

It’s perfect to educate yourself, to restrict yourself. Yeah, it’s the perfect closing word. Uh I thank you for your insights. Uh me it made me reflect a lot too.

And uh listen, I’m going to publish that on monExpansion.com/bs2 for Blind Spot 2 uh with all the reels and uh I’m going to try to do a newsletter too of everything we said, the shortest possible but the densest possible. We’ll see what I manage to do with the model not local for now. Thanks to you both. Thanks.