The future of engineering in the age of AI with Youssef Rizk

Youssef Rizk: It felt like huge FOMO to not be trying to build an AI. So kind of… without really thinking about it twice, we were like, we have to build an AI. And at that time it felt like… obviously the LLMs were very text-heavy. There was still a lot of work being done on images and so on, but not so much on audio.

———

Synthetic Salma: In a courtroom, not a war room for a change — Anthropic just settled a big one. 

They're mandated to pay USD 1.5 billion to authors who accuse them of breaching copyright laws by using their books to train Claude. 

It's the largest copyright settlement in U.S. history, and it's one of dozens brought by authors and publishers against mega-tech companies. 

[

Whether or not we think it's a real warning shot, some of the plaintiffs opted out of the settlement to file separate lawsuits.

Patrick FitzPatrick: Thousands of listeners trust that voice every weekday morning to tell them what broke overnight and what matters before their first meeting.

But here's the thing: The host of Morning Drive? She's not human. She's Synthetic Salma — an AI voice clone of our real executive editor.

Wondercraft makes AI that turns scripts written by our journalists into finished podcasts and videos. Its customers include Spotify, Amazon, and the World Bank.

Today on the show, I'm talking to Youssef Rizk — the co-founder and CTO of Wondercraft.

Wondercraft isn't a large language model. It runs on top of those things, so he doesn't build the models. He builds the apps that you actually use. And he's clear-eyed about the danger in that.

Youssef and I go deep into what it takes to build a tech product today — but we also tackle the question that every CEO wants answered: Does AI replace people, or does it just pile on more work?

Youssef's line is that one great coder and a bunch of AI agents can now do what used to take a team of software developers to do. In other words, AI is going to mean fewer jobs.

Here's what really stood out for me in this conversation: Youssef builds the future for a living. But when I ask him where we're going and where it all ends, he slows right down.

"No one knows," he says.

He's super-sharp on how to build this stuff — and on how a CEO should be thinking right now. And that's the conversation.

Patrick FitzPatrick: Youssef, welcome to Making It.

Youssef Rizk: Thanks, Patrick. Happy to be here.

Patrick FitzPatrick: Everything in tech, everything on the internet, is tribal. And so I've got to start by asking: what tribe are you? Are you a Claude guy? A ChatGPT guy? A Gemini guy?

Youssef Rizk: The world of AI started with ChatGPT, and that was sort of end of '22, early '23. And Claude was this new kid on the block. It looked like a knockoff version.

And then around, let's say, summer of '24, someone realized that Claude is actually really good for coding. There's a whole industry that's spawned because of Claude, which has all these coding agents.

Gemini is also great, but Gemini is great for a variety of different use cases. It's great as a chatbot, but it is the only one that has some capabilities that the others don't. 

Patrick FitzPatrick: What’s that? 

Youssef Rizk: The one that's useful for me is that it's the only one that understands video natively — none of the other models can watch a video. Gemini is the only one that can watch a video. So from a multimodal perspective, it's by far the best.

And if you're coding, you want to use Claude, but sometimes you use Codex, which is the OpenAI one. So it's a little bit of a depends-on-the-day.

Patrick FitzPatrick: Within the tribalism of models, if you scroll on X there really is, you know, a group of people for whom Claude is everything. There's a group of people for whom ChatGPT is everything. It's a constant pitting of model against model, and people make that part of their personality.

You're talking to a coder that you want to bring on board — their choice of Codex versus Claude Code. What does that tell you about them, if anything?

Youssef Rizk: It doesn't tell me too much. 

Patrick FitzPatrick: No?

Youssef Rizk: I think the only weird thing would be if you use Gemini for coding.

Patrick FitzPatrick: That would be a bit odd.

Youssef Rizk: That's a bit of a weird one. But for the other two, it doesn't really tell me much. To be honest, there are a lot of practical reasons that we can get into as to why you might think Claude is better versus Codex is better.

I do think that the industry standard now is Claude for anything that's technical — and I think even veering into the non-technical now.

The first thing is, even though it's been so prevalent, and for the last three years, AI is the name — it was weird to not use AI — there's still a bit of a stigma around it. People still think…

Patrick FitzPatrick: Even among coders?

Youssef Rizk: Even among coders, in the sense that people, I think, are a little insecure about what it means for them.... It becomes a little harder to differentiate where you add value versus where the model adds value. So I find that that's a little bit the case.

But to be honest, when we're interviewing someone, we've always taken a slightly different approach, particularly because we're a startup.

And I think if you go and interview somewhere like Google, the question is always as follows: build this thing in an interview. And then for the last ten minutes you talk about how do we make this scalable, production-ready? How does this serve a million users? All these questions about how you deploy it in production.

The reality is, if I'm interviewing someone and they give me that answer, that's an immediate no-go. I need you to tell me the least scalable version that you can get today.

Patrick FitzPatrick: Okay.

Youssef Rizk: The thing I'm trying to highlight here is that the only thing I really care about is function. So if you're achieving that function with Claude, Codex — you've set up workflows and skills and whatever it is on Claude, go for it; on Codex, go for it. We have people on the team who use basically everything. You have your individual licences. We don't top-down declare what you use.

Patrick FitzPatrick: Okay.

Youssef Rizk: It's up to you. Whatever makes you more productive, you can use.

Patrick FitzPatrick: So let's go back. 

Three years ago you were 27, sitting in Y Combinator trying to crowdfund tennis players in a startup called Moonshot. And today you've got a quarter of a million users. You've got Spotify, Amazon and the World Bank on your customer list. ElevenLabs and Steven Bartlett are on your cap table.

How did you get there? What's the story?

Youssef Rizk: Every ten years someone has this idea and they try it and they fail. And so, just like clockwork, it was our time to try it.

The motivation for this idea behind Moonshot wasn't so much crowdfunding as much as it was — really, the thesis was: how do you incentivise people in the world to back different people? 

And so our thesis was that if you incentivise people to back people who have the potential to become celebrities, that could really incentivise you, because you don't just make a financial return, you also get the emotional dividend, as we call it — particularly in sports, where, I mean, you know how people are about sports.

Patrick FitzPatrick: Sports. Yeah. Fanatics.

Youssef Rizk: Tennis was one of those where the economics made sense to start, in that it's super expensive to get started. It's also an individual sport, so you don't have this team dynamic where the team sponsors everything. So it was a good place to start.

We basically tried this. We ran the math: if you had invested in Federer under what we were saying, you would have 20x your money now. 20x is maybe a lot — but that's a once-in-a-generation player.

So long story short, that didn't really work. And that was around March 2023, when we came to that realisation. At that time, ChatGPT had been out for five or so months. And obviously, AI was just starting, and everyone was like, wow, this is a game-changer.

It felt like huge FOMO to not be trying to build an AI. So kind of without really thinking about it twice, we were like, we have to build an AI. And at that time, obviously the LLMs were very text-heavy. There was still a lot of work being done on images and so on, but not so much on audio.

So we started with audio, just because that seemed a little bit underserved. And we kind of got known for doing podcasts, but at the same time we started doing ads for some of our other customers. And then more recently into video — that was by popular demand. Audio is a little harder to distribute. I mean, think about it.

Patrick FitzPatrick: Harder to distribute, or harder to discover?

Youssef Rizk: Both. So podcasts are notoriously hard to discover, but also sometimes to consume. And they're also notoriously hard to keep going — as I'm sure you guys have experienced in the past.

Patrick FitzPatrick: Youssef Rizk: It can be a challenge, yeah. 

Youssef Rizk: Yeah. 

With AI, you could have looked at anything, and you looked at: okay, how do we turn text-based ideas into videos, into audio? Why that and not medicine? Why that and not some other aspect of technology?

Youssef Rizk: I think the thing that really motivated us was this idea that AI would be an enabler of expression rather than something to automate it. So from the get-go we were always pretty focused on making sure that you are in the driver's seat.

Which if you really think about it from today's lens, everything has veered agentic. There were two patterns. There was the copilot pattern, and that seemed to be where it started. If you think of, let's say, Cursor, which is the coding tool,  it really started as their main wedge was that they had a really good autocomplete. So you type, it autocompletes — but that wasn't novel. But they also integrated all these models into the same thing, and it could make some edits but fundamentally the edits were quite small, and you were still in control.

Patrick FitzPatrick: Very localised. Yeah.

Youssef Rizk: And I don't think — at least I didn't realise — how quickly the evolution would happen to that becoming agentic, which seems to be what people are still interested in now.

Patrick FitzPatrick: For listeners who don't understand "agentic," can you define that for us?

Youssef Rizk: Yeah. If you ask ChatGPT what the weather is today, that's not something that's in its training set. It doesn't have that knowledge codified in its weights. So what it needs to do instead is call a tool — let's call it the "get weather" tool — and specify the location, and that returns it.

The reason this is so important is because it scales out what an LLM can achieve, by incorporating it with different tools and integrations. That's why “agenticness” is quite useful: it's this loop that just keeps running, where the agent decides, I need to call this tool, and based on the result of that, I can call this tool, or do a variety of different things. So it gives it a kind of a … It is the brain, but this gives it the actuators that can interact with the real world.

There's only something that happened only in the last eight or so months that they suddenly became a real contributing factor to the economy. That wasn't really the case before, and you can attribute that to a variety of different things.

But I think just sticking to the question at hand: one thing that really motivated us was this idea that in a creative world, you still want the person to be creative. I say that from two perspectives.

One, because it's far more interesting to see human creativity being expressed. But the other thing is, AIs are not that creative.

I'm not an AI researcher, but intuitively, the way I think about it is: if you're trained on the amalgamation of all human knowledge, you tend to converge to some average. That average is probably 80th or 90th percentile in any given field — but it is an average.

If you try this and you ask Claude to produce a story for you, it will be a very similar narrative arc most times. It won't come up with something that is completely groundbreaking unless you prompt it in a specific way. 

So I think it's still super useful in the sense that it's great for inspiration, but it isn't necessarily the thing that's going to own creativity. So we kind of believe that in this world where AI proliferates, this expression of human creativity still remains super fundamental.

Patrick FitzPatrick: So where are you taking the company now?

Youssef Rizk: Basically moving a lot more into video. And I find it incredibly exciting for a variety of reasons.

I think as we do that, I'm recognising that video in its current form — which is the static interaction, or consumption, where you just sit down and watch something…

Patrick FitzPatrick: Passively watch the screen.

Youssef Rizk: …is not it. And I do think that we're probably veering towards a world where the way we consume video isn't necessarily going to remain that static thing.

So one of the use cases we work on is videos for L&D or internal comms at big enterprises.

Patrick FitzPatrick: And that's learning and development. 

Youssef Rizk: Yes. 

Patrick FitzPatrick: Okay.

Youssef Rizk: One thing that we're finding is that a lot of people are asking for interactive videos.

What does that mean? It basically means L&D I need to teach you about this training safety video. I’d like to make sure that after every section in the video there's a little quiz that you have to interact with. 

So it's still a video in the sense that you're streaming both audio and video content, but what it's doing is quite different. It's not a video telling you what to do; it's this interactive avatar that's coaching you on how to be a better sales rep, or do whatever your duty better is at the company, better.

This interaction has shown that it's a lot more effective, at least in L&D. But I do think that you can apply something very similar to entertainment, or to any other use case that you might want to have created a video for.

Patrick FitzPatrick: So you're interacting with the company's sales coach. You're interacting with the company's HR department to learn something, do something.

Youssef Rizk: Yeah. 

Patrick FitzPatrick: All right. I'm going to zoom out now and ask about the future of work, man. It's a topic on every CEO's mind, and my thoughts are really informed by Casey Newton from Platformer. He recently had Aaron Levie on the show — Aaron is CEO of Box, a piece of software that we used a generation ago.

And when Casey had him on, he made this argument that really matches up with my personal experience, which is that AI doesn't make us work less as people. It helps us get to that endless list of things out there, of ideas that just never get translated into reality.

And he ultimately thinks that AI creates more jobs for engineers. It's just that they're going to work in places that never would have hired them before. They're not going to work for Google or Meta, or even necessarily you guys. They're going to work at news organisations and hospitals and law firms and food manufacturers.

You've made the opposite argument. You told our producer last month that AI is actually replacing your engineers today. So if, as you said, you can have one good person with a couple of AI agents, and then you and Claude can figure out the whole thing — does that mean that the CEO can finally fire half their developers? Or that they can get rid of people on their marketing team because agents will take over the work for them?

How do you view that question?

Youssef Rizk: So I'll start by saying that no one really knows the answer. Everyone can make a hypothesis. And I do think someone like Aaron has a very interesting perspective.

What I can share is what my perspective is as a relatively smaller company.

Patrick FitzPatrick: You guys are 14, 15 people now.

Youssef Rizk: We're around 11, 12.

So when you're a smaller company, before you've really fully figured out what it is that you're selling and how you're scaling that, the thing that isn't super relevant is building more. It's building… or at least increasing the surface area of discovery and figuring out what is the valuable thing, and how do I double down on that?

That problem, when you kind of reorient the problem around this perspective, it is solved by being very close to the customers — kind of extrapolating what is happening and how people are using it. And it's just deep strategic bets, sometimes, of what is the value creation that we want to focus on.

Now that's very different if you're in a big company, where the backlog is massive, you've figured out what the value prop is, and you're just pushing forward with it.

Patrick FitzPatrick: But you guys have product in the field, right? We use it every morning. You're pushing deeper into video. That means you're building out pipelines of features and ideas. You've got the need to be responsive to market, to come up with what is next for you.

How does having fewer people help you there?

Youssef Rizk: The way we organised our team. Again, we started kind of early '23, and AI was still a copilot then. It was definitely an accelerant, but not where we are today.

The way that used to work is that we'd have, you know, a big team of engineers. Someone sits at the top, makes all the decisions, and then you federate out to the smaller engineering teams. 

Where we used to have a team of, say, three people doing one big feature. That is now collapsed into one person.

If you have a smaller team — very high agency, very high autonomy, people who can just figure out something nebulous and translate it into a concrete feature or product or something — that kind of profile, aided with the AI tooling that we have available today, is far more efficient than just staffing five more engineers on the project.

It leads me to a bit of a meta point here, which is: in an age where building anything has become basically infinitely cheap — it might not be cheap now as a token cost, we all know the issue, but just take it to the limit; as time moves to, you know, the year 2100, where that's infinitely cheap — what is actually valuable? What is valuable is figuring out what to build.

I completely agree with the statement that AI hasn't made anyone work less. It's made, I think, the top performers work more. It is incredibly satisfying to just churn through your backlog with an AI agent that does things while you sleep. So it created this kind of gamified sense of: how many things can I just do in a week?

But the reality is that for someone like me, who is making decisions about what the structure of our team is and what the different kinds of profiles are that we need, the more junior and intermediate roles are not that necessary anymore. If I'm hiring someone, fundamentally I'm hiring the same person that OpenAI, Anthropic, all the big companies are competing for — which is a very high-agency individual who's very good and understands software deeply and can do that.

That has a hugely negative ramification in the long term, which is that it's quite hard, in the conventional sense, to create a workforce in the future that has the same skill set as today's workforce.

Now, is that a problem? Time will tell.

Patrick FitzPatrick: Sure. But if this is happening — if you're only hiring high-agency professionals today, high-agency engineers, and you hire them today because they're available to you — what's creating the next generation of high-agency engineers? Where do they come from? What's the restocking mechanism?

Youssef Rizk: I think a lot of these people will not necessarily be engineers by training, in the conventional sense that we've had, but a lot more so people who are just able to think about how to design a system.

And again, if you really think about current software engineers, where do they come from? A lot of them come from not a computer science background. They come from any other kind of engineering — whether that's mechanical, which is pretty far removed from coding and what software engineering is like. But the reality is the same: it's that thinking that's valuable, and being high agency, which is not necessarily something you learn at university or at a job. It's kind of a personality thing.

The way to historically learn computer science — and why I think software has been such a great leveller for socio-economic mobility, in that you get a lot of people who aren't necessarily super privileged being able to work at the best companies out there — is because it was very prone to self-study, if you will. You just need to pick up a project and do it.

Patrick FitzPatrick: There's no licence to be a software engineer at all.

Youssef Rizk: Exactly. The barrier to entry to doing it was a lot higher, and now that's not the case anymore. You can actually simply just say, “hey, create a game for me,” and Claude will do it.

I think if you're just asking it to do the thing — which I often do, for expedience — you're going to have a great time, but you are not learning necessarily how the thing is done. But if you're using it as a learning tool, it's a fantastic tool.

Patrick FitzPatrick: So what is the entry-level job of the future in engineering?

Youssef Rizk: I don't know if I have a great answer to that. It's very, very unclear.

But I just think that there's such a democratisation happening of software that formal employment in one of these big companies isn't necessarily the normal entry path for an aspiring software engineer anymore.

I think it’s a lot …

Patrick FitzPatrick: Sure I mean Google and Facebook — Meta — are engaged in layoffs, right? There's a pretty good argument that a lot of those layoffs are not AI-driven as much as it is, you know, I over-hired in 2021, 2022, money was cheap.

And now it's the herd. It's like six-year-olds on a football pitch, right? They're all chasing after the ball saying, yeah, I'm an AI company. Look, I'm doing layoffs because I love AI.

I think the more interesting question is: where do those engineers find jobs in the future? What does that entry point look like? What qualifies me to go work for you in the future?

Youssef Rizk: If I were just speaking anecdotally about what I've seen, there has never been more demand for a good software engineer. You fundamentally have to outcompete Claude to be someone of interest, which is a pretty high bar for most people.

The group that is below that baseline will either go work in a position where tech is a supporting function — the baseline there is a lot lower, and you just need minimal tech work; it's not really the core of what you do. But what I've seen more often is that necessity is the mother of innovation, and you end up having a lot of these people just try to start something by themselves.

It's an amplification of a trend that has existed for a while, which is that generation after generation feels less and less like conventional employment is a way to have financial security in the future.

I do think that this is part of the reason you've seen such an explosion in startups, and that they have become kind of this glamorous thing. I actually don't think people view working at these [big] companies extremely favourably right now. You do it kind of if you just want stability. But it feels like a lot of what's glamorous now is starting your own thing. Everyone tells me that they want to start their own thing.

Patrick FitzPatrick: So if one of your friends — let's say you have a 35-year-old friend with a really bright kid going into high school, and it's time to start thinking, what do I want to be when I grow up? What do you tell that person to tell their kid to study? What's your advice?

Youssef Rizk: What I maintain is that if time were to go back, and I understand where things are going, I would have done the same thing. I studied electrical engineering. I would study some sort of engineering.

I don't remember anything about electromagnetism. I don't remember the physics we learned, or circuits. I don't remember this stuff. You don't need to, though.

All you really needed to get out of it is an environment in which you have the opportunity to build a bit of muscle memory and projects and expertise, and train your brain in the way of thinking that you need for the general world.

Patrick FitzPatrick: What if your friend says, okay, but you know, my daughter likes to draw. My daughter likes to write.

Youssef Rizk: Let's assume they do. Let's assume we live in a future where basically AI can do everything for you.

The fundamental question to answer is: do the current things that hold value maintain value then, or not? And there's an argument to be made that they won't. Like yes when [AI] can do everything, aggregate share of labour in the market will decrease, because you just don't need people to do as much.

There's another argument to be made that the things that people find interesting shift, and then we start shifting more and more to the thing that isn't abundant. And the thing that isn't abundant, if I really had to put my finger on it, is anything where the human experience is part of the value proposition. 

Patrick FitzPatrick: Like what?

Youssef Rizk: That is things like art, that is things like writing, that is things like the hospitality industry — restaurants, clubs, nursing, doctors, lawyers... Anything where you actually need to know the person in front of you.

So if someone is telling me, you know, I really like writing, or I really like drawing — I do think this is as viable a path as it has been in the past to create economic value, provided you're good enough at it and the aspect of somewhat…. The need to hustle has never changed. So provided you're doing all these things, I don't think this is necessarily a “bad thing” to study.

Patrick FitzPatrick: So what human labour, if any, does AI eliminate, in your view, as somebody who's building in the field?

Youssef Rizk: I think it depends a little bit on the industry and some of the mechanics in the industries.

In coding, coding is a solved problem as far as most people are concerned. You can easily do something by yourself.

If you consider other industries — something like legal, they obviously have agentic workflows and so on, but you fundamentally are a person. Like, I contract a lawyer, not the AI. The AI is a great accelerant for their job. So this is one of those where it replaces the easy work. And I think where AI replaces the easy work, that industry stands to benefit, or is generally an interesting place to be, and you will end up with a lot of situations, like we said earlier, where people are working more, not less.

For the other things, where it replaces the hard bit — that's a lot more interesting, because it's unclear what then happens. I don't have the confidence to point at a specific thing or industry and say that this is where it's happened, because even, you know, coding has always been hard, but it was never the bottleneck. It was always architecting and scoping and figuring out what to build, and for that you still need a human.

And that's why I'm saying that the top 1% is like a thousand times more valuable than it used to be in the past.

Patrick FitzPatrick: The guys at ITIDA here in Egypt, the Information Technology [Industry] Development Agency, have been making a really big deal about business process outsourcing, or global centres of competence — you know, you can call them whatever you want to call them. 

We exported services worth 7.4 billion, something like that, last year, 4.8 of which was BPO.

When I look at that stuff, and I look at what Google's been rolling out lately with voice models, I start thinking: how durable are those jobs, where they're doing high-end KYC stuff, high-end risk management stuff, right? But it's all fundamentally stuff that is at least partially done — could be done — by AI.

What becomes of this industry over the next five years?

Youssef Rizk: There is a precedent with all you've said, in that clearly, if a big part of our economy is driven by outsourcing, there is precedent that whoever is contracting is clearly on the lookout for cost reduction — and that there is no loyalty or moat that you may have as the outsourcing partner right now that AI won't replace the second it becomes cheaper.

You still have a bit of a buffer. And I'll take the example of customer support voice agents: it's still cheaper to hire an Egyptian outsourced company to do that than it is to hire a voice agent.

Patrick FitzPatrick: You see the costs coming down.

Youssef Rizk: The cost will drastically come down.

And so in that case, I do think Western worlds have a brighter future ahead in productivity gains, given that they've already let go of a lot of the easy stuff that AI will automate. And unfortunately, I do think that the developing world will feel this bite the hardest, where you fundamentally do the thing that AI promises to optimise.

So I do not think it's a very bright future for this kind of labour.

Patrick FitzPatrick: What is the future for us here, then? These were jobs that were on-ramps for lower-middle-income kids to a genuine middle-income future. 

What's the future for them?

Youssef Rizk: It's a tough one, and I don't know if anyone has a good answer to this.

Patrick FitzPatrick: Let's say that I'm running a 500-person company. I've built up a workforce over a period of 20 years. I'm a 53-year-old CEO. 

What is the wrong way to approach the question of AI and labour substitution right now?

Youssef Rizk: I wouldn't necessarily try to make a decision immediately on whether labour will be substituted or not. I don't think it's a top-down thing where you can just say, now I can substitute labour — because the reality is that maybe you can, maybe you can't, and only implementation will tell you that.

The belief I fundamentally hold is that probably 80% of what your company currently does can currently be automated with AI. 

Patrick Fitzpatrick 80%?

Youssef Rizk: It is a bit of a strong number, but just anecdotally speaking.

Patrick FitzPatrick: In the services industry?

Youssef Rizk: In the services industry, a lot of what you do can be automated with a little bit of human supervision at the top.

The things that really picked up for agents were customer support and coding. Those two things were huge. So customer support is something I would focus on — see if we can improve that.

And you can maybe try to drive AI from a perspective of revenue maximisation rather than cost reduction, and try to figure out how you can not play the zero-sum game, but rather a positive one.

Patrick FitzPatrick: How do I make more money, not cut—

Youssef Rizk: How do we all get better with this, rather than how do we reduce cost? Because the reality is that you can probably do more with more people, if each of them is more productive.

Patrick FitzPatrick: All right. Before we go to break, I'm going to tell you my worry about this in the near term.

We all go down this path of maximising revenue and/or reducing expenses through the power of AI. Life is great. 

And Sam Altman goes bankrupt. Anthropic goes bye-bye. The ability to rely on these agents, these robotic employees, suddenly disappears.

Is this something that keeps you awake at night, as a guy who builds on somebody else's tech stack?

Youssef Rizk: No.

Patrick FitzPatrick: Why not?

Youssef Rizk: I think we have tasted too much of this for it to not be a thing in the future.

It's hard to imagine OpenAI will not exist. It's hard to imagine Anthropic won't exist. When you're a USD 1 trillion company that's still private, with that much money invested, it's very, very difficult to imagine that you will not exist.

Patrick FitzPatrick: The delta between money in and money out on the revenue side is pretty scary.

Youssef Rizk: It's pretty scary. But you're also factoring in a crazy growth, and maybe justifiably so in some cases, where this is a literal life-changing technology and bigger than the internet. I mean, as all the executives are saying, the biggest worry is under-investing, not over-investing.

Genuinely, people will not know how to be productive anymore without AI. When I am on a plane, it is a better time spent for me to just relax on the plane than code something manually without internet — because the second I land, I can ask Claude to do it in five minutes, and I can spend my time thinking about something else.

Now, whether we're talking about the specific providers and companies — who knows who will exist in 20 years' time? I mean, Bell Labs used to be a huge company, right?

Patrick FitzPatrick: Fair point.

Youssef Rizk: But there is a new Bell Labs.

The other thing I want to say is that these big companies are not the only players. You have a huge investment into the open source community that keeps happening. Something like DeepSeek is open source. Admittedly, you can't run it on your laptop — it's a bit too big — but you can go rent a server and run it there.

We reached a point of independence earlier this year when — what was it? — OpenClaw came out. Right? Mac mini sales—

Patrick FitzPatrick: Exploded.

Youssef Rizk: Exploded, because everyone's running that. And actually, internally, we run a lot of our agents on a Mac mini in the office.

So a lot of the world is also shifting more to: how do I maintain my own independence? Which is generally a theme that I think is happening in the world.

So I don't worry too much about, you know, maybe a specific model provider shutting down. And that happens a lot, with a lot of the consolidation that's happening now — you'll find a lot of labs being absorbed by the bigger companies. If you were historically relying on one of them, you might be inconvenienced. But I don't think that as a whole, the AI industry is at risk of not existing.

Patrick FitzPatrick: All right. We're going to take a quick break. And when we come back: the value question — where in this whole AI economy does the money actually end up, and the constant risk of being sherlocked.

———

Patrick FitzPatrick: Okay, I want you to challenge me on a view that I hold pretty heavily, and it's that I see, over time, the foundation model sort of becoming commoditised. That Claude, that Gemini, that ChatGPT sort of become the technology upon which real value gets created at the application layer — the stuff that you're doing, the stuff that other people aspire to. That the model enables you to build something cool, and that all the money accrues there.

Sam Altman will have you believe that all money accrues to him. There's no other way to justify his valuations.

True or false: the value is at the application layer more than it is at the foundation model layer.

Youssef Rizk: It's true. So I'm not going to challenge you too much on that.

Patrick FitzPatrick: I'm going to challenge you on that. Okay.

Youssef Rizk: I mean, it's an interesting one. There are a lot of analogs you can draw. And I think it's probably a bit of both.

I think it's hard to imagine that the amount of money it takes to train a model, especially a frontier model, isn't prohibitive for anyone to really do it. So you're always going to have a few players left in the field. It's not a thing that you can go do now.

The open source models — let's see. There will be major innovations that make a Claude-level model run on your computer. Obviously you'll need a bigger computer, but it'll be runnable.

Patrick FitzPatrick: You're going to have 256 gigs of RAM or whatever running locally.

Youssef Rizk: So you'll be able to probably run these things. 

Interesting point for Apple, which is: I do think they're going to be very well positioned for this.

A model by itself doesn't do much. As it used to be, it was just a chatbot, and it was hard to justify valuations at that point. Now it really integrates into the rest of everything completely. And so if that's the case, then inevitably the value lies at what it can do, which is the application layer.

Patrick FitzPatrick: Yeah. I mean, Claude was pretty smart before. It's the product thinking that went into the harness, and all the other knowledge that they built on top of it, that enabled the revolution that you've seen in the last eight or nine months, I think.

Youssef Rizk: Yeah. The consensus is that something changed in November 2025, which coincides with the release of Opus 4.5.

Patrick FitzPatrick: 4.6 - 4.5? 

Youssef Rizk: 2025 was a year where they did a lot of reinforcement learning on — what was the word — “verifiable rewards.” Basically, they were giving it problems where the solution was verifiable, and that in some capacity improved the model's ability to reason. So these models got really good.

But what also got really good was the harnesses. The engine was good, but they also built a really nice car around it.

The interesting point here, always, about application versus model is: which view do we hold? Is it the one where the model will eat up all the applications, or do the applications have value?

Patrick FitzPatrick: Yeah, do they sherlock you?

Youssef Rizk: A couple of months ago… My opinion has always been, yeah, I mean, the models will eat you. And I think they've actually historically done that. Whoever was worried about that early on has probably been right, unless there was a specific reason why they wouldn't eat you specifically. And it's not immediately obvious to me which one.

Maybe in the case of vibe coding you could give more bias to Anthropic winning, in this case.

Patrick FitzPatrick: Probably. 

Youssef Rizk: But I don't think it's as clear-cut.

The reason is: the model may become commoditised to a point, right? Open source models are around six months behind the frontier models. And in a world where some of the tasks that your program solves are not frontier-level complexity, they do not need frontier models.

I think the dynamic is always risky when you are competing with your supplier — your supplier keeps moving up the stack and you start competing, because then they have all the leverage over you.

Patrick FitzPatrick: How does this challenge apply to you guys, in a world where NotebookLM can now create a video or create a podcast for you? Why would a brand use you and not use NotebookLM to create that asset that they want to create?

Youssef Rizk: So I'll go from the start and from first principles on how I think about this.

The reality is that you are in the line of fire of all these model companies if you are fundamentally competing with them in a way where, if the model gets better in a year, you get hurt. So you always want to be riding the improvement of the model. And the way you do that is with a very deep understanding and knowledge of the customer workflow that you're trying to solve.

Patrick FitzPatrick: So Specialist domain knowledge.

Youssef Rizk: Specialist domain knowledge and integrations.

I think we have a historic precedent here, which is Google. Google has been around for 20 years, and since the dawn of Google, anytime you want to start a company that is vaguely adjacent to something that Google does, the investors will always ask: what if Google does this? And in all this time, Google hasn't really done this. Google is, I mean, known at this point as a graveyard of the products that they create.

Patrick FitzPatrick: The Google Graveyard page just scrolls on infinitely.

Youssef Rizk: And you ask about NotebookLM — NotebookLM will probably have the same fate.

Google makes money on one thing, and they will fight to the death on that thing. Everything else is a little bit of brand building. Maybe a project there takes off — you have a few projects that really kicked off, whether that was Gmail or Chrome or Google Maps. These things, that's their bread and butter. And all of these things serve to get an ecosystem of products where you exist, and they can kind of—

Patrick FitzPatrick: They can serve you ads.

Youssef Rizk: It’s all of that fundamentally. NotebookLM does not serve that purpose. NotebookLM is a fun research project for someone in a fun product team.

So I do think — like, will Anthropic… you know, they've got Claude Design. Will that be usable by most designers? I don't know. I compare Figma, I compare Claude Design. I asked our designers to try them. We still use Figma. All your components are there. Your entire design system is there. Will Claude be easily able to replace that — slash, is it worth it for Anthropic to—

Patrick FitzPatrick: To go down that route.

Youssef Rizk: TBD. I don't think so.

I also think there's a different thing here that really helps startups, which is that startups are this bundled set of risk that exists far away from a parent company. If you really think about it, all these labs really stem from DeepMind. DeepMind kind of branched all of them out.

And the reason is, you have someone like Ilya, who co-founded OpenAI. Is that the kind of thing that he wanted to do? You couldn't really get past the Google lawyers to do it. You can't simply tell Google lawyers, I'm going to go scrape the internet. Whereas some of the labs may have done certain things that weren't super aboveboard with a lot of the legal.

But that's how startups worked. Uber is technically illegal when it started, the kind of thing that they're doing. But the economic value that they produce outweighs the initial hurdle that they have to pass. The economic value gets bigger, and then we create legislation that makes this all fine retroactively.

So, I mean, there are a lot of copyright issues with AI, and of course they're pretty serious things to take into account. But from a slightly more practical, pragmatic capitalist perspective about the world: the value that AI has created far outweighs the severity of any copyright claims that the world may have against them.

Patrick FitzPatrick: Big statement to make.

Youssef Rizk: It is a big statement to make. I just think, practically speaking—

Patrick FitzPatrick: I created nebulous value for humanity by stealing the corpus of human knowledge.

Youssef Rizk: You're probably right in framing it that way. And there was definitely that.

I just think, from a very practical perspective about the world: will any government now put sanctions on or penalise any of the big labs for doing that? I doubt it.

Patrick FitzPatrick: So let's say I'm running a 200-person Egyptian business. I've got a ten-year-old CRM or an ERP that everybody hates, and some 21-year-old kid comes up to me and says, you know, I can vibe-code a replacement. Do I let them do it?

Youssef Rizk: No.

Patrick FitzPatrick: Why?

Youssef Rizk: The reality is, every line of code you write is a liability. And unless you are willing to dedicate staff and engineering and other resources to maintaining it, I don't see a point in doing that.

That's not to say that there's no point in coding. I do think there are a lot of SaaS companies whose value prop is fundamentally flawed.

Patrick FitzPatrick: Like who? 

Youssef Rizk: Basically, in the last six months or so, the threat of Anthropic and other AI companies going up the stack into territory that has historically been occupied by a lot of the SaaS players has caused a lot of these SaaS stocks to plummet — I think 50, 60% down. Figma is actually one of the worst affected. I think it went from highs of USD 140 per share to now it's at 22.

If you ask me, that is incredibly shortsighted — in some cases. And in something like Figma's case, maybe. But it's kind of on a case-by-case basis.

The interesting thing is, I think it's somewhat naive to say that this SaaS apocalypse is justified. But I also think it's somewhat naive to say it's not justified and it’s completely reboundable. 

I have a firm view that these companies actually got hit a lot harder — the worries are a lot worse than the reality suggests. These companies are still growing tremendously.

So if I were Salesforce, I'm not worried about a vibe-coded CRM. I mean, you tell me: would you code a CRM? You'd have to maintain it.

Patrick FitzPatrick: Buy anything you can buy, and build only the things that you can't buy.

Youssef Rizk: Build the things that are your bread and butter…

Patrick FitzPatrick: 100%. That you cannot buy at a reasonable price somewhere else.

Youssef Rizk: I know that doesn't necessarily answer your question, but the first question was: would I code my own thing? No.

If it was for the purposes of saving money or creating something better — there are almost guaranteed to be far better solutions that will emerge for a far cheaper price, because of the same thing that made you consider vibe coding.

Patrick FitzPatrick: That they're going to face competition.

If you were a 30-something right now — smart, an engineer — and you wanted to build a great startup that played on a theme in the Middle East, in our part of the world, what would you build?

Youssef Rizk: I do think that the Middle East has been historically underserved in terms of its representation in how these models behave.

A classic one — I'll say it from the audio world — is if you look at text-to-speech providers that offer Arabic, it's a disaster. 

The dialects are incredibly different. The words are completely different. And so it's a lot harder to train a model to do that.

The models are very adapted to a Western way of thinking and a Western way of communication. If you try to deploy a customer support agent, they're so biased to be nice and friendly and "can I do anything else?" and so on. That is not necessarily going to work well in an Egyptian context, where the cultural dynamics are a little different.

So there is maybe a world to be built here around that kind of angle, which is trying to create these purpose-built models that serve this specific demographic — though I don't know what the ROI here is. There is a reason—

Patrick FitzPatrick: I mean, you see people like TII in Abu Dhabi building models, you see the Saudis building models. But what would you build on those models? What product would you build today?

Youssef Rizk: Weirdly enough, a big area of disruption we have relative to the Western world is: we are not that streamlined.

Patrick FitzPatrick: Okay.

Youssef Rizk: If you really think about it, most business here is done on WhatsApp. Email — sure, we use it. But look at the WhatsApp penetration in business here relative to somewhere like Europe or America.

And I think with the advent of AI, you're now able to meet that slightly more chaotic market dynamic, and customer, where they are, while offering them the same solution.

Patrick FitzPatrick: I want to spend the next couple of minutes talking about a really specific type of listener. 

She's 50-something. She's a CEO. She's super accomplished in her field. She, however, is completely unaware of how to start, or where to start, when it comes to AI. She probably thinks ChatGPT at best is a replacement for a Google search box.

Where does she start?

Youssef Rizk: I think there are two axioms I would start with.

The first one is that a lot of AI buzzwords are flying around — they change every couple of months. I think one thing to figure out — or to start with — is to completely ignore all these buzzwords. 

Just start from: what is the thing I want to do? And you'll discover, kind of inevitably, all the different things that are there to help you get that done.

Patrick FitzPatrick: What's the first thing she does, then?

Youssef Rizk: We are now at a point where AI can do pretty much anything you ask it to, or there is a way to implement that.

For instance, if your job involves going on LinkedIn and messaging people for, I don't know, outreach — AI can go and do that. Whether it's through a tool, or whether by literally going on your browser and clicking through LinkedIn to get to what you want. 

You can basically go to ChatGPT, Claude, Gemini and say, how would I do this? And they'll walk you through it, including actually setting things up for you automatically to get things done.

I think AI can be super daunting, as is any new technology, and you will hear people talking about things that you don't fully understand. That's all irrelevant. 

The main thing is: what is a problem that you think a business can benefit from, that you know about, that you can spearhead? You cannot delegate this.

Patrick FitzPatrick: Does she have to grab hold of Claude herself, and graduate from Chat to Cowork to Code, and learn what a CLAUDE.md file is?

Youssef Rizk: Go straight to Code.

Patrick FitzPatrick: Straight to Code.

Youssef Rizk: These days, in some managerial capacity, you get bogged down in the motions of making the business run. And I do think it's actually pretty fun to sit back down and be like, let me implement this.

A lot of the tech founders turned investors who have basically stopped coding for 15 or 20 years are notoriously getting back into coding.

I think it's just, again, don't put the carriage before the horse. Just figure out what it is, and make sure you're comfortable enough understanding what it is that can be automated, and then push for that automation.

Patrick FitzPatrick: So let's say six months in, she has learned Claude. She's picked up one great project. She's implemented that project. People are starting to adopt the technology.

At some point in time she kind of has to zoom back out and become CEO again — you know, in the classical sense. Maybe an AI-enabled CEO, but she's got a business to lead.

Who runs the rest of her AI transformation for her? Is it a classical CTO, or is it a different skill set? What type of person or mix of skills does she need to carry that forward for her?

Youssef Rizk: I don't think this function sits within a predefined role. I don't think it's a CTO or a CFO. I don't think it falls with it.

In our company it came from the tech team — we obviously adopted AI way quicker; that's what the software world has done. But we have very much forced the rest of the company to operationalise AI in their workflows as well.

I think it needs to come from someone within the company. Who that person is depends probably on a case-by-case basis, but it needs to be someone with enough remit to actually exercise some sort of influence and mandate over the remainder of the company.

I'll just say that if anyone wants to do this, it's not a bad credential to have.

Patrick FitzPatrick: How much do you use Claude as a thinking partner about the business, versus about problems?

Youssef Rizk: I use it a lot more for the problems than I do as a general thinking partner.

Patrick FitzPatrick: You're not one of these guys who's trying to replace himself with an agent as a boss.

Youssef Rizk: I'm trying to replace all the grunt work that I do. I would like to unlock more thinking, basically more observation.

Patrick FitzPatrick: But you don't want to create the future-you CTO, the future-you CEO version.

Youssef Rizk: No, I think that should still be—

Patrick FitzPatrick: A human you.

Youssef Rizk: That should still be me.

The thing you get as a human is weirdly connected neurons that LLMs don't really have — that spark in weird ways, that are non-deterministic, that fundamentally give you a bit of an interesting perspective on how to interpret this piece of data. And LLMs can probably do, again, like 60 to 80% of that.

But I think it is that 20% that is truly unique and creates a very interesting insight into understanding fundamentally what your consumer wants, that is the thing that is uniquely human, in my opinion.

Patrick FitzPatrick: My CTO said something recently that really stuck with me. 

He said that the role of humans in a world where AI does lots of the production is going to be to inject entropy and randomness into the systems that the AI is developing. That it is this imperfection, the rough edges, the surprise — that's the job of the human now.

Agree or disagree?

Youssef Rizk: Yeah, I think it's a different reframing of what I've said earlier, which is that you need human supervision there.

Interestingly, there are companies trying to, you know, record your meeting notes so that you can inject them into your agent, and the agent has that. 

TBD, again, on whether that works.

I do think that the human fundamentally is the owner of the narrative, and the narrative is fundamentally what shapes the logic and the decisions that you make. So injecting entropy and randomness, to me, just means a human is a bit more creative and owns the actual direction that the agent runs in.

Patrick FitzPatrick: Knowing where to point the thing, and making those decisions as they're going forward.

Chris Sacca is a cowboy-shirted ex-Googler, and he wrote early checks for Twitter, Uber, Instagram, Stripe, Twilio — turned USD 8.4 million into billions. Some people would argue the most successful venture capitalist ever.

And he told Tim Ferriss last year that he is only investing in tech if it's tech to enable in-real-life experiences. That everything he's doing is analog. It's the messiness of humans. Quote: "The last bastion of humanity is going to be in the random, unpredictable messiness of humans. The rough fucking edges that make no sense."

The dude's buying warehouse space in rural United States where he can bring people together and have art and sports.

Make me the case about what survives of that human creative endeavour going forward. 

Is it in-real-life experiences? Is Chris Sacca onto something? What would you do today in that respect?

Youssef Rizk: I think we can look a little bit at the past few years, and one thing to look at is how the price of luxury goods has fared relative to the price of luxury experiences.

What you'll find is that, at least at secondary market rates, the price of a Louis Vuitton bag isn't holding its value the way it used to in the past. You can see that reflected a little bit in the stock prices of these companies as well. All this to say that I think there's a bit of either fatigue or saturation within that luxury product market.

If you look at luxury experiences — if I'm remembering this right — their prices have on average tripled in the last seven or so years. And, you know, we can argue about what the dynamics that led to this are, but that seems to at least be the fact.

In a world where that fatigue or saturation of product catalogues was taken to the limit, what remains is what remains scarce — or, we value what remains scarce. And what remains scarce is that which cannot be automated, which is fundamental human-like relationship.

So I think he's onto something, for sure: that these real-life experiences will become drastically more valuable than they are today, in the future.

Patrick FitzPatrick: What is the human part of building L&D software, or making a video, or doing a great broadcast, that you can't automate away — that you can't put into the LLM and Wondercraft?

Youssef Rizk: I mean, ever since we started Wondercraft, and even in the podcasting use case, we never fully set out to replace your big famous podcasters, because the value prop was different.

For anyone who listens to the Enterprise podcast — the daily one — it's actually quite functional. Fundamentally, you're listening to it because you care about efficient information consumption. I just want the news as I drive, rather than having to read it in the morning. So that's functional.

I think the appetite for something that is AI there is a lot higher. No one's lying about it being AI. You know—

Patrick FitzPatrick: You're right at the top of the show.

Youssef Rizk: Fundamentally, the reason you like a particular podcaster is very different from that. You like them because you like their banter on screen. You like their opinions. You like their humanness a lot more than—

Patrick FitzPatrick: It’s incredibly intimate to put those EarPods in your ears and walk around with somebody for an hour.

Youssef Rizk: It is an entertainment value. It is a connection and a relationship that you have.

And it's very hard, you know, if you have a podcast with guests, to bring a lot of the interesting guests that people will want to have that conversation with, and to build that credibility, that you ask the right questions and actually create a pretty intellectually stimulating conversation. 

That you cannot replace with AI — and even if you tried, the AI version of it would become not that interesting, because what you care about fundamentally is the scarcity, which is human relationship.

Patrick FitzPatrick: Absolutely.

So, staying with podcasts for a second: the web is being reshaped by AI right now. You've got audio living out on Spotify. You've got audio living out on YouTube, on Apple.

How do people five years from now discover the stuff that you help people build? How do the future listeners of Diary of a CEO discover Steven, in a world where there's less one place you can go to find all the information in the world?

Youssef Rizk: That's the trillion-dollar question. If you have a good answer for that, you should quit your job.

Patrick FitzPatrick: Quit my job and go start something else.

Youssef Rizk: Fundamentally, I've always maintained this position, which is that in a company there are two roles essentially. There's one, which is I need to build a product that works. And the other one is I need to grow it.

Even before AI, the product has always been easier. The growth teams, in general, they run experiments. The experiments are: well, let me try this thing and see if it actually results in any growth.

And I think that there is no good answer to that. SEO used to be the biggest thing. Programmatic ads used to be the biggest thing. These things arguably still work to an extent. 

Now the new thing is: how do you get recommended by ChatGPT instead of showing up on the first page of Google? These things are there, but I don't think they're an explosive new paradigm shift in how you discover.

Patrick FitzPatrick: Not a one-for-one replacement of Google.

Youssef Rizk: I don't think that's been discovered yet. I don't think we've figured out what the right mechanism for that is. I think it's going to live a lot less in the digital world and more in the physical world.

Patrick FitzPatrick: Long-form journalism, like the type of stuff that we do — do we survive?

Youssef Rizk: I think you're making a comeback.

Patrick FitzPatrick: We're making a comeback.

Youssef Rizk: I think long form is making a comeback, correct.

Patrick FitzPatrick: Tell me more.

Youssef Rizk: Shorts are notoriously getting sloppy. The kind of content that you see there isn't that great.

Podcasts, fundamentally, are long form. Podcasts have been on the rise for a while. I do think that there is an inherent quality that is perceived in long-form content, at least partially, if not in large part, driven by the fact that AI isn't able to produce great long-form content. 

So it becomes a bit more of a distinction between what is this artisanal, handmade content versus—

Patrick FitzPatrick: The New Yorker versus a three-minute summary of a news story.

Youssef Rizk: Though I will argue myself here and counter my point, which is: I think in the age of AI, where the information is so readily available, do people want to watch a one-hour thing, or do they just want to get the summary of it? Because the summary of it can probably be expressed in a couple of paragraphs that you can take a minute to read.

Patrick FitzPatrick: Yeah, I think it really depends on what you're looking for as a consumer. 

We're not in the business of a scoop that says HSBC got a mandate to do an IPO, because that literally has a shelf life of five minutes and then everybody else has got it. What we want to do is tell stories that make you think, that make you sit back, that make you spend more time with us.

Youssef Rizk: So what's your view?

Patrick FitzPatrick: On which aspect of it?

Youssef Rizk: Long-form content.

Patrick FitzPatrick: Long-form content. I think it survives. I think we're going to see a shaking out of people who are able to do it. I think it requires doubling down on human craft.

I don't want a piece of information to trade on. I want something that informs my worldview. And I think people who are looking for that will find that. And that's where long form survives.

Last question for you, sir. If you were not running Wondercraft — if you had to walk away tomorrow — what would you build or invest in next?

Youssef Rizk: It is such an interesting time to be alive…

Patrick FitzPatrick: Mesh keda? [isn’t it?]

Youssef Rizk: … because at least in my lifetime — and I imagine in the lifetimes of people older than me as well — but there has never been a time where the future is this undefined.

It is an interesting time to sit down and form a thesis about where you think the world is heading, and what you think will be valuable. The conversations we've had today are really conjecture about what is going to be valuable in the future.

And I think the reality is that if you look back three years ago — what has really changed day to day for you? I mean, you still walk up to the office, you still publish the newsletter. A lot of what you do day to day hasn't changed that much, despite having a literal omniscient intelligence available on your phone whenever.

Patrick FitzPatrick: Yep. 

Youssef Rizk: The thing that interests me here is that I do think a lot of human dynamics survive. I don't think this is the end of the economy as we know it. I do think it's very interesting to think about what lies in the future.

Is it to go down the path of, let me build more agentic things — that focus on a niche, on a problem that's experienced by certain industries, and really focus on that, whether that's legal, financial, construction, you name it? Do I focus more on in-person events, and focus on tech that enables the human relation? Do I go open a hotel because I think tourism is going to boom?

It's unclear. 

And I think there are so many different things here that you can explore. But with the backdrop of what I think is true, which is AI can pretty much do a lot of the legwork of white-collar jobs — a lot of the easy stuff, not the hard stuff.

Patrick FitzPatrick: Last-mile stuff. It's really challenged that….

Youssef Rizk: Yeah. And so that's where you come in. How do I build something that rides a wave of imminent model improvement, which will be the case for a few years? And how do you make sure you are in tune with a very rapidly shifting change in behaviours and preferences, that seems to happen way faster than what I perceive it to have happened in the past?

So I don't know what I'd build, but it'd be one of these. Something along the lines of all of these things.

Patrick FitzPatrick: If you had to pick — I put a gun to your head, man, and I make you pick one thing. What's it going to be?

Youssef Rizk: Um. I would actually try to focus something on gaming.

Patrick FitzPatrick: Gaming?

Youssef Rizk: I think there's a lot of room for human entertainment in the future world we're heading for. And I generally think it's a lot more fun than selling B2B SaaS.

Patrick FitzPatrick: Youssef, thank you so much for joining us today.

Youssef Rizk: Thank you, Patrick. It was a pleasure.