Site icon The Data Exchange

Why AI Makes Games Worse—and How to Make Them Better

Ben Lorica

Piero Molino on Player Backlash, Small Open-Weight Models, and the Economics of AI in Video Games.

Subscribe: Apple • Spotify • Overcast • Pocket Casts • YouTube •  AntennaPod • Podcast Addict • Amazon •  RSS.

Piero Molino, CEO and co-founder of Studio Atelico, joins Ben Lorica to discuss what happens when AI becomes a core game mechanic rather than a cost-cutting tool. They explore why on-device, specialized models could unlock new kinds of games, how post-training gives developers greater control, why artists should share in the value created from their work, and what coding agents mean for the future of software development.

Subscribe to the Gradient Flow Newsletter

Interview highlights – key sections from the video version:

Jump to transcript

Related content:

Support our work by subscribing to our newsletter📩

Transcript

Below is a polished and edited transcript.

Ben Lorica. All right. So today we have an old friend, Piero Molino, CEO and co-founder of Studio Atelico, which you can find at Atellico.studio. The taglines are “AI games with a human touch” and “AI tech that runs on device.” And with that, Piero, welcome to the podcast.

Piero Molino. It’s so nice to chat with you again, Ben. It’s been a while, and we’re looking forward to chatting more with you about this.

Ben Lorica. Actually, yeah. For our listeners who may not be aware, Piero comes from an enterprise infrastructure background. There’s the open source project Ludwig, which he created when he was at Uber AI, and which became a startup called Predibase. But now he’s in mobile and mobile games, which is fascinating and will be the focus of our conversation. Obviously, I’ll try to broaden it out so that people who are not working in mobile games will also find it interesting.

But I guess, Piero, let’s start here at a high level. What do you think the AI industry gets wrong about games, and vice versa? What does the gaming industry get wrong about AI?

Piero Molino. That’s a great question. When I started this company, I brought together people from both the AI industry and the gaming industry, exactly because those two spaces don’t interact a lot.

In particular, it’s been surprising for me to see that gaming has been an early application of every new technology in the last 30 years. Imagine 3D physics, networking with online games, mobile games—all of these things. Whenever there was a new piece of technology, one of the first things people would implement with it would be a game. That has not been the case with AI.

I think it is because of what you just said: maybe the AI people get something wrong about games, and the games people get something wrong about AI. To a certain extent, that’s what I’m trying to reconcile with the company.

I’ve been doing this for about a year and a half now, so despite being relatively green in the gaming industry, I have a deep passion for games. I’ve been playing and buying games throughout the years, and some of my games have now become retro games. But when I bought them, they were not retro games.

I’ve been observing the space, and I’ve been very close to it. During my college years, I was a video game journalist for a little bit, doing reviews of video games and things like that. So I’ve been following the space for a long time.

For me, working on these two things together was an opportunity to bring some of my technical knowledge from the AI industry into the gaming industry, and also finally do something that has been my passion for a long time.

What I’m seeing is that a lot of the gaming industry has, on one hand, been late to adopt AI technologies at the development level. I think maybe starting this year, people are really beginning to use coding agents to build games. Now you can see that whenever there’s a new model that comes out, some of the demos are always, “Oh, how well can you build this game?” or “Can I one-shot this game?” and things like that.

On the other hand, there has been asset production—creating images, audio, voices, and things like that.

Ben Lorica. And NPCs too, right?

Piero Molino. Yeah, for NPCs. And on those things, there has been quite a bit of pushback because it seems like players feel it’s cheapening their experience instead of adding the real thing, like a human being drawing something or a human being performing some lines.

It seems like AI is basically a mechanism for cheapening the development cost and cheapening their experience. I’m sympathetic to that kind of argument because games are an incredible combination of tech and art. If you make it so the art is not as good as it used to be, but it’s cheaper, then it feels like the worst version.

Ben Lorica. And I guess what you’re pointing out there is that the focus is on cost reduction—

Piero Molino. Exactly.

Ben Lorica. —rather than actually improving the game.

Piero Molino. Exactly. And that’s actually what I’m trying to do—the opposite of it. I’m trying to make it possible to use AI within the game in order to give game designers and game developers one more tool to make new games that would not have been possible before.

I think a good analogy here is physics engines. If you imagine a game before physics engines were a thing, you would push a box, and your character would push it by a specified distance in the gaming world. That gave developers and designers perfect control over where the boxes were, but it didn’t allow for emergent things.

Once you have a physics engine, you can use these boxes for doing puzzles, like in Portal, for instance, or other games like that. Or you can do things like building. There’s a famous game called World of Goo where you build bridges with gooey little creatures. You stick them together, and basically you create physics-based soft bodies to see if they can support the passage of some cars and things like that.

Fundamentally, adding these new capabilities, like physics, enabled a set of games that were not possible before. That’s what I want to do with AI and games: give people AI tools so that they can create new games that were not possible before.

Ben Lorica. So it seems like the distinction here—maybe this is an oversimplification—is between AI as a gimmick and AI actually adding something substantial to the game and the game mechanics.

Piero Molino. Yeah. I want to make it a foundational piece of the game mechanics of all the games that we’re going to make, because that enables something new.

I can give you a cool example. Our first game that we are developing, Bobium Brawlers, embodies these concepts and these values. It’s a game where you create your own creatures just by describing them. These creatures are represented both with images and portraits, but also by a deck of cards that you can use to battle against your friends. Both of those depend on what the player describes.

If I describe a dragon that spits fire, then I’m going to do fire attacks, and the picture of the creature will be a dragon. But if I create an ice wolf, then my attacks will be more ice-oriented, and the picture will depict a wolf.

Because this is done in real time, and there’s an infinite number of possibilities that players can create, it would not have been feasible to do it in any other way than having AI generate it in real time. So it becomes core to the experience.

This is just an example. We’re also developing some other prototypes. This just shows what it means to have AI at the core of the gameplay.

Ben Lorica. Yeah. In fact, as I recall—I may have read this somewhere—you built a bunch of prototypes, and your initial conclusion was that AI, in many instances, makes the game worse. Right?

Piero Molino. Yeah. What we discovered right at the beginning was that first we created a basic technological layer, which I can tell you a little bit more about later because I think the technical aspects are also interesting for your audience.

Once we built the base of that layer, we were trying to figure out what fun game mechanics that technological AI layer could enable. We created many prototypes along a spectrum, from ones where there was little to no AI—just a little sprinkle on top of a classical game mechanic—all the way to game mechanics that were, I would say, 99% AI.

To give you some examples, you can imagine something that looks more or less like a board game, but where some of the pieces of the board are generated by AI. That would be an example of very little AI.

Or, to think about a lot of AI, you can imagine playing Dungeons & Dragons where the AI plays the role of the game master. You tell it what happens, and it tells you how the world changes. These are the two extremes.

What we found is that if you go one route—the very little AI route—it doesn’t add a lot, so it’s not particularly useful. Why would you want to have it at all?

If you go to the other extreme, what happens is that it feels very gimmicky, and it doesn’t feel like a game, because games need challenge. They also need rules that the player can learn and master so they can feel like they’re getting better at it.

But if everything is governed by the AI, you cannot create a theory of mind about the rules of the game, and you cannot get better at it.

In our experience, the prototypes that were more in between these two extremes were the most fun. You have some aspects of the game where AI is fundamental and the game would not have been possible before, but some other aspects where classical game mechanics and game design become the predominant part.

If you blend them together, that’s what we found to be the most fun. This is how we got to Bobium Brawlers, where the creature-creation side is all AI-governed, but when you fight battles with your friends, that’s more classical game mechanics—a card-and-dice game that is more familiar to players, where they know how they can get better, how they can progress, and what challenge they have to overcome. That creates the fun.

Ben Lorica. So I guess I interpret what you just said as basically: the sweet spot is when, as a player, I’m not thinking about AI at all. I’m just lost in the game. So it’s an implementation—

Piero Molino. Detail.

Ben Lorica. Yeah. If you broaden that out to non-games, maybe good AI products are the ones where they stop feeling like AI products.

Piero Molino. I 100% agree with you.

Ben Lorica. And I guess part of that makes sense, right? If you’re a knowledge worker or a game player, you want to be in a state of flow where you’re not thinking about what’s in the tool.

Piero Molino. I think you shouldn’t be thinking about which model you’re using, how many tokens per second, how many tokens you’re consuming, and all of these implementation details. They get you away from the game you’re playing or the task you’re performing as a knowledge worker.

I think the more AI becomes invisible, the better the application.

Ben Lorica. So the points that you’ve covered so far, how widely known or accepted are they within the gaming community, not within AI?

Piero Molino. I would say not as much, honestly, for at least a couple of reasons.

On one hand, there aren’t many games that have AI in their core mechanics. There have been a few examples, notably AI Dungeon, Retail Mage, Suck Up! There are a few examples.

Ben Lorica. How would you describe these games in terms of popularity?

Piero Molino. Suck Up! had a moment when it became pretty popular for a short amount of time. It was everywhere on YouTube, and all the streamers were playing it for a little while.

Ben Lorica. And were people talking about it as an AI game?

Piero Molino. Yes. They were talking about it as an AI game because that was the main mechanic and main gimmick of the game.

At the same time, I would argue that it was a relatively simple and relatively shallow mechanic, although it was one of the first attempts. It was a really courageous attempt that I respect deeply.

But because the depth wasn’t there, I think the interest in it didn’t last very long. That said, I think they gave some good examples showing that you could do this and build mechanics with AI in the game.

The reason I think it’s not as popular or as explored yet is that right now there are at least two big blockers. One is technical and the other is economic.

From an economic perspective, there are fundamentally two mechanisms for how games make money. One is: I sell you the game, you buy it, and you can play as long as you want.

The other mechanism is: I give you the game for free, and while you’re playing, there are some things in the game—

Ben Lorica. In-game purchases.

Piero Molino. Exactly. Sometimes they’re cosmetic. Sometimes they’re more like boosters for experience, things like that, which you pay for.

In both of these business models, the more players play, the more the company makes money. In the first case, if you pay upfront, I’m going to give you new pieces of the game, like downloadable content. In the second case, the more you play, the more in-app purchases you make.

If you introduce a variable cost where, fundamentally, the more players play, the more I as a developer have to pay the cloud provider, it kind of breaks the business model.

For instance, in our case, in Bobium Brawlers, if we were to use a cloud provider, it would cost us about 20 cents per creature creation. If we had, I don’t know, a million players and they created five creatures a day, we would have to pay a million dollars a day to a cloud provider.

That makes it very difficult because then you would have to make at least a million dollars a day just to pay your bill. That’s not how game development and game business models actually work right now. It’s very, very expensive, even with all the improvements in cost per token over the last few years.

Ben Lorica. All right. We’ll drill down on this topic of edge versus cloud in a second. But I do have one quick question for you.

For those of us on the outside, what we read is that, at least for asset creation—say I’m a game designer, I’m supposed to build this game quickly, but now I have to create all these trees and houses in the game—is that a place where AI is heavily used now?

Piero Molino. I would say these days, in the last year or so, that’s where many gaming studios are focusing.

Ben Lorica. Yeah.

Piero Molino. With not a great response from at least one side of the public, in particular in the U.S. and other Western markets.

Ben Lorica. So the reason, Piero, is that they can tell it was created by AI?

Piero Molino. Yeah. On one hand, they can tell, but I also think it’s a more philosophical argument: these models have very often been trained on data from the internet without the consent of the original creators and things like that.

So it’s a combination of perceived lower quality plus the perceived—

Ben Lorica. Ethical issues. So is the perceived lower quality justified?

Piero Molino. I would say right now, yes, still—although that may not be the case soon enough.

I also think most companies will start to do something relatively smart. If there’s a character in the game and this is the main character of the game, you probably always want to have it drawn by hand by a human. But if it’s one of the hundreds of tree assets that you have in the background—

Ben Lorica. Yeah.

Piero Molino. —maybe that’s less important. I think developers will get smart about how they use these tools too.

Ben Lorica. All right. So edge versus cloud. That’s an important part of what you do, and you’ve already touched on cost as a main motivation.

First of all, before we go any further, in your case, the models you need are multimodal, right? They’re not just text. They need to be able to create images, sound, video. Is that correct?

Piero Molino. Yeah. Basically, that technological layer that we created, which we call the AI engine—similar to how people call them physics engines—can do many different things.

It can generate text. It can generate structured text. It can generate images and audio. It can take user input in text, images, and audio.

It’s basically a multimodal mechanism that also has additional capabilities like embeddings, search, semantic search, and classification.

Ben Lorica. What about CAD design, where you have these 3D assets? Do you need models that are able to do that?

Piero Molino. Right now, we don’t support the generation of 3D assets on the fly yet because that’s a little bit more involved and—

Ben Lorica. It seems like that’s a direction some of these world models are going—

Piero Molino. Yeah, the world models are going in a direction—

Ben Lorica. The ones that are targeting games.

Piero Molino. Yeah. There are many people doing many different things.

Some of them are pure frame-to-frame, image-to-video or video-to-video generation. Others are more like: I render something in a bare-bones 3D world, like a gray-box 3D world, and then I slap some style-transfer mechanism on top of it that changes the style of the game in real time.

There are many different people doing many different things there.

Ben Lorica. But the bottom line is that when you say “model,” you’re actually describing something beyond text generation.

Piero Molino. Yeah. It’s text, image, audio, all together.

Ben Lorica. And our listeners who interact with some of the frontier models like Gemini or ChatGPT know they can generate multimodal assets, but they are in the cloud.

So one of the things you’re trying to do is lower your costs by not using these expensive APIs.

Piero Molino. Yeah, exactly.

Ben Lorica. And secondly, to run on device, where “device,” broadly speaking, could be a laptop, mobile phone, and so on and so forth.

Which means, what is your base model? Is it one of the open-weights Chinese models?

Piero Molino. Basically, for each of the tasks, we have separate models. They’re kind of like building blocks.

Ben Lorica. Is it something Studio Atelico created from scratch, or did you take one of the open-weights models and start from there?

Piero Molino. It depends on the specific component. In some cases, we are building our own models. In other cases, they are open models that we are adapting, compressing—

Ben Lorica. Building your own model means data—

Piero Molino. Right.

Ben Lorica. Getting the data and then the compute. So what you’re saying is that this is doable for you?

Piero Molino. Yeah. For large models—say, 100 billion parameters plus—that requires a lot of data and a lot of compute.

Ben Lorica. That’s a lab.

Piero Molino. The nice thing for us is that the models we run are in the millions of parameters or very small billions—like one billion or three billion at most.

Ben Lorica. And the reason, Piero, is that ultimately you want to be able to run it on device.

Piero Molino. Exactly. The fact that we use these smaller models makes it so that training them from scratch is not as expensive as training 100-billion-parameter models.

Ben Lorica. And the data for pretraining is just public data?

Piero Molino. Exactly. For instance, if you look at models from both Hugging Face and NVIDIA—and now Hugging Face plus NVIDIA—for text, they’re all using data where the recipes are shared, if you want.

Ben Lorica. You mean the data recipe? Is this the Nemotron family?

Piero Molino. Yeah, exactly, the Nemotron family. They have the recipe for training and the data they’ve been using all available.

Ben Lorica. But your models are multimodal.

Piero Molino. Yeah. We actually have different models for each modality, and then we put them together in different ways.

For some modalities, like speech, for instance, you can get away with smaller models. We want to be on the Pareto frontier of how big the model is—meaning how small a footprint it can have in RAM on a device—versus how fast it can generate things, versus how high quality it is.

Ben Lorica. Throughput and latency.

Piero Molino. Throughput, latency, and quality. Because you can get something very, very fast, but if it’s not good enough, it’s not usable. But if you have something super high quality that is super slow or requires more RAM than the device has, that’s also not useful.

So you want to be at the intersection of all of these.

Ben Lorica. And there are models that you use to develop your own games, and then there are models that have to run on device as people are playing the games.

Piero Molino. Exactly. For building our own games, we use coding assistants and things like that, so it’s not like we’re doing anything substantially different from anybody else.

But for what we run on device, it’s a combination of some open source models that we fine-tune for the specific game—because those models are small enough that they’re also cheap to fine-tune—and some of our own models.

If you don’t fine-tune them, they’re usually not high quality enough for running the game.

For instance, for speech, we have our own text-to-speech model that is on the order of 100 million parameters. So it’s not that large, but it’s high quality enough to sound good to the players and can also run on device. It’s a good intersection.

Ben Lorica. And so when you run the models on device, on a phone or a laptop, how does that impact someone’s battery life?

Piero Molino. Right.

Ben Lorica. And is the inference always being done on GPUs or CPUs?

Piero Molino. It depends on the different devices.

For instance, on iOS devices like the iPhone—our first game, Bobium Brawlers, will release first on iOS—Apple has, first of all, very good hardware, but also a layer that is not exactly like CUDA but is similar in spirit. It’s called Metal.

With Metal, you can write kernels against it, and it decides where to run your computation—on the GPU, NPU, or CPU.

Ben Lorica. And you can use coding agents to write your kernels.

Piero Molino. Right. Yeah, we also use them for that, for sure. But writing kernels is a little bit of an art, right?

Ben Lorica. Oh yeah. You need people who have actually done it.

Piero Molino. Exactly. I’m fortunate enough to have spent some time in the Hazy Research Lab at Stanford, where FlashAttention comes from and things like that.

I was exposed to these kinds of things very early on, and that keeps helping me these days to write very efficient kernels that otherwise would not have been as easy to write, even with the use of coding agents.

We have a little bit of secret sauce there. I won’t get into the details of it, but you can imagine it as a compiler—a kernel compiler—that makes it so we have very, very efficient kernels.

Ben Lorica. So then you have a pipeline for customizing models, or post-training them—whatever that means: fine-tuning, reinforcement learning, whatever it takes to make them better.

Piero Molino. You already mentioned it, actually. We are using more or less Ludwig for doing that.

Ben Lorica. Yeah. You’ve developed a pipeline for doing this, but to some extent the pipeline depends on how good the base models are. And you’ve described the base models as a mixture of things you’ve pretrained from scratch and a bunch of open-weights models that you can then customize or post-train.

So generally speaking, Piero, if you were describing this to people who haven’t gone that far down the road of post-training or customizing models—especially the ones they want to run locally—

Piero Molino. Right.

Ben Lorica. I’ve heard Pete Warden describe these local models as actually not being that far behind the frontier models, depending on the task. You know Pete Warden?

Piero Molino. Yeah.

Ben Lorica. So we’re not talking years; we’re talking months. Do you agree with that assessment?

Piero Molino. Yeah. I have a couple of pointers that could be interesting for the listeners.

On one hand, if you go to Apple AI, they do a bunch of analysis and comparisons. They have some results that are actually from last year, but they apply this year too. Fundamentally, they were showing that, depending on the task, open models that you can run on devices lag behind by maybe six months on some tasks and nine to 10 months on others compared with—

Ben Lorica. Without customization and fine-tuning.

Piero Molino. That’s without customization, exactly. So already, the lag isn’t that large. And you can fill that gap by doing customization, post-training, fine-tuning of sorts.

There’s another example that’s a little older. I think it’s from the beginning of 2024. When I was at Predibase, we released this paper called LoriLand, where we trained 1,200 fine-tuned models on 30 different tasks, each of them using different base models.

Some of the base models were one-billion-parameter models, some were eight-billion-parameter models, and there was everything in between.

After fine-tuning on the specific task, they were reaching the same level of quality as what were the best models at the time—GPT-4, basically—at a fraction of the cost.

The vast majority of the value was coming from the data for the task. The base model was more or less an independent variable. The actual effect was coming from the fine-tuning and the increased performance that you get after post-training.

Ben Lorica. So in many ways, these small specialized models are what make your architecture possible.

Piero Molino. Yes.

Ben Lorica. The question is, if I’m a game developer or game designer, how easy is it for me to replicate what you’ve done?

In other words, does every game studio need to do what you do, unless of course they work with you? How easy is it for a game studio to develop the assets that you have?

Piero Molino. There are two aspects to it.

One is the inference side—not the training or post-training part, but inference. On inference, I would say we’ve solved, or are solving, those kinds of problems for other people too.

At some point—we haven’t announced it yet—we want to release the engine in a way that other people can license it and use it.

Ben Lorica. So you’ll be like Unity.

Piero Molino. Not like Unity, because it’s not a game engine. It’s just an AI engine. So it’s just the AI side.

I would say the most similar thing is Havok, which makes a physics engine. It integrates with Unity, Unreal, Godot, but it’s not the entire game engine. It’s just that specific component.

On the other hand, to make the game work really well—in particular with the LLM side of things, but also with the other aspects like image generation and audio generation—and also to govern it and get it to do what you want, there is this process of post-training.

We are trying to make that process as easy as possible for ourselves. Again, my experience with Ludwig is what I’m bringing from my previous company and my previous open source contributions.

In that case, it’s just a configuration file. You don’t have to know all the details of how to implement the post-training, how to collect the data, and all of that.

There’s also a little data collector that you can put in front and after. You can wrap a classic OpenAI call with that thing, and it collects inputs and outputs. Then you can use those for doing the post-training.

You can run the game with large models, collect inputs and outputs, and then fine-tune a small model on those inputs and outputs to mimic the larger model.

It’s still more involved, but it’s not extremely complicated. With the help of someone like us, I think for gaming companies it’s going to become easier and easier to do this kind of adaptation.

Another reason for doing it is not just quality, but also control.

For instance, in our game, Bobium Brawlers, we have image generation, and we’re doing post-training of the image-generation models not because the base models aren’t good enough, but because we want the output to follow the style of our own artists.

The best way to do that is to have your artist draw what you want and then train on that. At runtime, when a user or player asks for a specific creature, they get an image of that creature that represents what they asked for, but the thickness of the lines, the style of the colors—everything about the art style in general—reflects the style of our own internal artists.

So that’s very important. It’s not just a matter of quality; it’s also a matter of control.

Ben Lorica. Because you rely on the device for inference, in many ways the device becomes your cloud—your deployment boundary.

In the cloud, I’m only as good as the GPUs in the cloud provider. In your case, if my iPhone isn’t good enough, then it won’t work. So the device becomes your deployment boundary in some ways.

That’s why I think you have a minimum version of the iPhone that you support, right?

Piero Molino. Yeah. For this game, we’re basically targeting an iPhone 12 Pro and an iPhone 14.

In the engine, though, for each modality we support many different models of different sizes and, obviously, different quality.

You can imagine, from both a developer perspective and potentially a user perspective in the future, that there could be a slider—like the resolution slider in a video game for visual quality—except there could be a slider for intelligence, if you want.

If I’m on weaker hardware, I can only run so large a model. If I’m on stronger hardware, I can run a larger model. So I can interpolate between those things in the future.

Additionally, I would say we have tailwinds on the hardware side because a pretty consistent observation throughout the years of hardware innovation is that whatever is today the best workstation hardware, costing $5,000 to $10,000, will in a couple of years be your average laptop, more or less.

You can already see it with hardware like the NVIDIA DGX Spark, the new laptops, AMD Halo, and the Mac Studio.

The direction everybody is going in is unified memory combined with strong inference hardware. The DGX, for instance, has 128 gigabytes of unified memory. If you have that much unified memory, it becomes relatively easier to run a model that is large enough to be pretty capable.

Today, the DGX Spark costs whatever—$5,000. Two years from now, that’s going to be what everybody has in their pockets.

Ben Lorica. Yeah. You raise an important point. A lot of articles are written about GPUs and GPU shortages, but actually the bottlenecks are electricity and memory.

Have you been following high-bandwidth flash?

Piero Molino. Oh, this one I’m not familiar with.

Ben Lorica. It’s interesting because there might be some things you can offload to high-bandwidth flash in the future. Obviously, right now it’s a bit early.

Piero Molino. Flash meaning nonvolatile flash?

Ben Lorica. Yeah, similar to what SSDs did—

Piero Molino. Got it.

Ben Lorica. —but for memory. It’s early, but I think people are excited about it.

Your focus is on mobile phones, right? What about games for laptops? Obviously, the same technology would apply there. Is anyone doing something similar to what you’re doing for games on laptops?

Piero Molino. Not that I’m aware of, honestly.

For us, doing a mobile game first was a little bit like fighting the final boss. Because if you can make it run on an iPhone 13, obviously it runs much better on laptops and desktops.

To give you an example, our image model—the one that we’re using on the phone—is an adaptation of an open-weights model, but squeezed, improved, quantized, and fine-tuned.

We made it run on an iPhone 14 in four to six seconds for a 512-by-112 image. On my laptop, which is now four years old, it takes one second to generate an image. If you run it on a 5080, it takes 20 milliseconds.

So it scales up pretty nicely to stronger hardware.

Ben Lorica. Interesting. By the way, all this talk about on-device inference in particular got me thinking—and again, this was based on a recent conversation with Pete Warden.

You look at a company like Apple that hasn’t really gone all-in on this crazy AI data-center capex spending. But on the other hand, if you can run more and more on device, maybe in the future you’ll have models on macOS or iOS that developers can tap into, and a lot of routine tasks will just be supported by whatever model Apple has.

What’s your take on the possibility of Apple shipping something within the OS?

Piero Molino. I think part of the reason they are pushing on the size of RAM and unified RAM is exactly for that. At some point, there could be a part of that RAM reserved for a model that is always in memory or something like that, and you can constantly query it.

That would be, I think, an amazing world.

I believe the name of the game is “good enough.” You probably don’t always need the frontier model.

Maybe a good example is already Opus 4.6, which is a very competent coding model. For the vast majority of tasks—say I have to make a basic website that doesn’t have super-important security or other special features—I think Opus 4.6 is good enough.

If you have a model that runs on device that is good enough for those tasks, then maybe for some specific things, like more research-oriented applications or scientific discovery, you may want to go to the latest and greatest cloud model.

But maybe for 90% of tasks, you may not need it.

Ben Lorica. And by the way, it’s not one-shot either, because every time they ship an update to the OS, they can update the model.

Piero Molino. Exactly. So they can get better too.

What I envision in general for applications is a hybrid state where, over time, you will be able to do more and more locally. Then, for just a few relatively more complex and involved tasks, you ship them to the cloud only when you actually need it.

Ben Lorica. So tell us about the—what is it called?—the artist rights contract.

Piero Molino. Right. Our observation is that what makes games special is the art created by the people involved in the development.

That includes artists who draw images or voice artists who dub the games. It could be 3D artists or any other artistic endeavor that is part of the game.

We believe that if we are training on assets they create in order to expand what is possible, then it’s only fair to treat them with respect and compensate them accordingly.

The main principles of this contract, which is the contract we use with our own artists, are that, on one hand, they are compensated for their work. If they draw images, we pay for those images. Very simple.

The other important aspect is that they receive royalties from the games as a mechanism for compensating them for licensing their style.

Ben Lorica. Because if they draw a tree for you, the generative AI could draw a thousand trees based on their style.

Piero Molino. Exactly. If you just pay them for the one tree they created, but then based on that style you create a million trees, it’s unfair not to pay them for the million.

It doesn’t have to be a one-to-one mapping, because in our case—

Ben Lorica. Tell me that you’re paying in cash and not in crypto.

Piero Molino. We want to pay royalties based on the game.

We don’t know exactly how valuable one more creature in Bobium Brawlers will be, because people aren’t paying for the creatures. But we will know how valuable the whole game is because we will know the revenue for the game.

If the game is made possible by the style of our artists, I think it’s only fair to compensate them in a way that is somehow proportional to that.

Ben Lorica. In your previous incarnation, which is where I knew you, you had Ludwig—open source. So what is open source now with Atelico?

Piero Molino. Right now, we haven’t released anything—actually, no, we released one thing open source, now that you’re making me think about it.

We’re using Godot as the engine for the game, into which we plug our AI engine. I created a little linter and formatter for GDScript, which is the language that Godot uses internally, and I made it open source.

If people are curious, they can go there, download it, try it, and use it. It’s very cool because it’s very fast and very customizable.

But the core technology of Studio Atelico, like the AI engine, is not open source for now. We’ll see in the future what we’re going to do with that. It is possible that we may—

Ben Lorica. The impression I have, Piero, is that in the game industry, open source is not a big deal anyway. As long as your tools lead to good games—popular games—do they care?

Piero Molino. My perception, as a former outsider who is getting into the industry right now, is that, in general, game developers don’t particularly care.

But I think recently Godot has gained steam and a following because of the changes Unity tried to implement at one point.

Ben Lorica. So Godot is like the open source Unity.

Piero Molino. I would say, yeah, more or less. Not exactly, but it’s the open source alternative to it.

Ben Lorica. So what are we talking about in terms of parity? Is it like Solaris versus Linux in the year 2000, where Solaris is still clearly better than Linux?

Piero Molino. Yeah, I would say probably right now that’s where we are.

And I wouldn’t say clearly better in absolute terms of quality or the quality of the games. I’m talking more in terms of adoption.

Ben Lorica. But if you look three or four years ahead, does the arc of history favor Godot because people will gravitate toward open source? Is that a fair prediction?

Piero Molino. I would say I hope so, honestly. I hope so because it feels like a better open alternative.

It’s already really good, but an even better open alternative—

Ben Lorica. What is the funding mechanism for Godot? Is there a foundation?

Piero Molino. There’s a foundation, and I think there are a bunch of companies that have spawned around it, created mostly by the maintainers to do services.

Ben Lorica. So it seems there might be a way to sustain the development.

Piero Molino. Yeah. Now it seems pretty self-sustaining, and I think in the past they received some grants from Microsoft and others.

Ben Lorica. Because Godot is open source and gaining popularity, you’re saying that there might be some pressure for different components of the game-development stack to become open source as well.

Piero Molino. That is definitely possible, yes.

Or, on the other hand, the closed source ones need to become substantially better so that their value is much more understood and much more worth it for developers when they’re making a decision.

Ben Lorica. Or what could happen is you folks could be the AI engine for Godot. Then, as it becomes popular, game developers will go, “Well, wait a minute. Why don’t we create an open source version of Atelico, since that’s an important part of the stack?”

But I don’t think, in the beginning at least, people will care, because they just want to create games.

Piero Molino. Exactly. I think it’s a matter of capabilities, really, more than anything else.

Also, making a game usually takes a long time. The cycles of adoption of any gaming technology are usually longer because, in particular, very large games take five or six years to make.

The decisions you make now, you’re going to stick with for six years.

Because of that slowness, let’s assume I’ve been building a game in Unity for the last four years and now a better engine comes out. I’m going to finish my game in Unity. Maybe for the next game I’ll consider an alternative.

The cost of switching would be very, very high.

Ben Lorica. What if Tencent releases an open source AI engine like Atelico? Is that plausible? I don’t want to make you nervous, but—

Piero Molino. No. That’s always the case, right? With every single thing you do as a company, there’s always, “What if Google does it? What if the incumbent does it?” and whatever else.

Ben Lorica. Well, in this case, Tencent is also heavily invested in games.

Piero Molino. Exactly. But Tencent also owns—I don’t remember exactly what share—part of Epic Games, which makes Unreal Engine.

There are many things happening in parallel.

For any startup, I think the advantage is, first of all, how fast you can move and how focused you are.

Tencent is a behemoth of—I don’t even know how many people, probably hundreds of thousands—versus us. We are seven people, and we are focused only on this. They do a million different things.

I think for every startup, at any stage, the focus and smallness of the team become assets that let you move fast, try different things, react very quickly, and be trailblazers.

It’s almost always the case that small companies are trailblazers, and then the other ones follow later. If the small company is successful, that’s always true for every startup.

Ben Lorica. All right. Let’s get you out of here with two quick questions.

The first one is coding and coding agents. How much do you rely on them, and what’s your take on the future of coding?

Piero Molino. Personally, and as a company, we rely on them. They’re a big part of our tooling.

At the same time, what I’ve discovered over the last year or so, since we’ve been using them more heavily, is that the exact same tool in the hands of different developers ends up producing quite different output.

I think the quality and experience of the individuals using the tools influence the results a lot.

For us, it’s not going to be a mechanism to cut costs of any sort. It’s more a mechanism to be faster. With the same number of people, you can be faster, as opposed to cutting people.

Ben Lorica. That statement you just made—in the hands of people who are good—but in the future, Piero, how do we know who’s good if everyone is using the same coding agents?

I just talked about this with someone. In math, which is where I came from, they now have a situation where generating a proof is easy. Maybe even writing the paper is easy because you can ask the AI to write the paper.

Now, whether or not the paper can be understood by mathematicians is a different thing.

Mathematicians are proposing that not only is generating the proof cheap, and maybe writing the paper is also cheap, but in order for a paper to be accepted, one, it has to be understood by humans who will accept it, and secondly, they’re discussing an additional requirement.

They might say, “Hey, Piero, you submitted this paper. We want you to give a talk about this paper in front of mathematicians.”

In terms of coding, I guess the equivalent would be some sort of code review. That’s how you discover, “Hey, this guy is the next Jeff Dean.”

Piero Molino. Yeah, that’s a great point. And I would say the role and decisions of—I think it’s like a—

Ben Lorica. By the way, for coding and math, you have generation, but the verification step is also doable because you have compilers and formal verifiers. You can even have the AI do the tests.

In other words, generate tests, verify, write the paper—or in the case of coding, generate the documentation. Everyone is on the same footing.

So what distinguishes the star from the non-star?

Piero Molino. Yeah. I think first of all it’s the “what.”

Ben Lorica. Okay.

Piero Molino. You have all the intelligence, quote-unquote, in what to work—

Ben Lorica. On. What to work on.

Piero Molino. What to work on. It’s a very important choice. It becomes the most important skill.

Ben Lorica. That’s still what remains in math, unless the AI solves all of the famous problems. But then after that, what?

Piero Molino. Yeah. Again, I think we’ll figure out that there are more problems.

At the limit, there’s always an infinite number of problems one could be solving.

Maybe it’s different between math and code, but—

Ben Lorica. In math, it’s a non-renewable resource. There are only so many interesting problems.

Piero Molino. At the same time, once you learn something—I’m sure this is always true in science—then you think about the consequences of what you’ve learned.

Ben Lorica. This is the case when the human does it, because the human will try to solve something, get stuck, go this way, explore, do a bunch of things, and then maybe say, “Oh, there are interesting problems here too.”

It’s not clear that the AI will do that.

Piero Molino. Right. You can have an outer loop that is kind of self-improving, but—

Ben Lorica. The human, in the course of exploring, has taste.

Piero Molino. Yeah. And whenever you learn something, that changes your mental model about the world and about—

Ben Lorica. Whereas the AI is just an optimization engine that’s trying to crunch through so much compute to give you the answer.

Piero Molino. Yeah. But you’re directing it somehow, right?

Even if you’re not directing it, you’re still creating the scaffolding that makes it so that it discovers a direction.

Ben Lorica. So there’s still the “what.” What do you think about this whole notion that you should be able to explain to me what—

Piero Molino. For scientific discoveries, definitely that is valuable because I think discovery without understanding—we couldn’t falsify it.

Ben Lorica. It would be like Einstein saying, “I have the theory of relativity, but I cannot give a talk about it.”

Piero Molino. Exactly. It’s not true knowledge if it’s not disseminated, and if we’re not able to disseminate it.

I think it’s similar in a sense for coding. You, as the developer, are the first critic of this thing. The critique is super valuable.

To give an example, let’s assume I’m creating a piece of code that needs to run on a device. Once it has been built, I try to run it on another device, and it breaks.

Maybe I didn’t make it clear from the very beginning which devices I wanted it to run on. But after the fact, I can critique it and say, “Well, this works here, but it doesn’t work there. That’s the new task you have to do. Now you have to make it work there.”

A human can also look at the generalization.

My personal experience has been that these models are extremely good at writing code, and sometimes even redundant code, and sometimes making patches on top of patches on top of patches.

In particular, because it’s so easy for them to write yet another piece of code, generalizing across multiple instances, multiple devices, or multiple different implementations of the same thing using design patterns is not always the first way they implement things or think about them.

There’s an argument to be made that it doesn’t matter—they can create yet another bespoke function. But their memory and context are also limited.

So if you create, let’s say, 100 different implementations of the same function for 100 different devices, they themselves, when something breaks, will not know where to look.

I think good software design principles are something a human can provide and influence the model toward.

Ben Lorica. Which, unfortunately, is somewhat gained through experience. But if you’re not gaining experience—

Piero Molino. Exactly. And that’s a problem.

If you are only using models from the very beginning, you don’t have the opportunity to gain that experience. That is a problem.

I think maybe that’s something to be said for schools, actually. Like, why would—

Ben Lorica. They need to rethink what they do and teach, right?

All right. Last question to get you out of here. Obviously, because of what you do, you need to pay attention to frontier models. We talked about world models, multimodal models.

Is there anything that excites you that you feel is somewhat below the radar that, in six months, more of us will be talking about?

Piero Molino. Interesting.

Obviously, I’m very excited about smaller models and things that run on the device and that whole space. I think some people are starting to pay more attention to it.

Ben Lorica. A lot more are beginning to, right?

Piero Molino. Exactly. I think there will be more of that. I’m preaching to the choir here, I’m sure, but it’s also the kind of thing I’m working on, so I’m excited about it.

Of the things I’m not directly working on, I think there’s an interesting space opening up around world models.

Specifically, I’m not a strong believer in, let’s say, the original formulation of world models as just: past frames plus controls equals next frame.

Fundamentally, I’m not convinced by that because the inductive bias has so little structure that you would require an incredible amount of data to make it do anything reasonable.

And then when there’s a bug, you don’t know what to do. If it does something different, what are you going to do? You’re going to create much, much more data to try to compensate for errors.

I think it’s a recipe for not-great outcomes.

But what I’m seeing is the beginning of people starting to inject some structure. So you get a blending of structures that come from game engines together with some of the techniques of world models and style transfer, et cetera.

The blending of these two things, I think, is interesting and promising.

Maybe not exactly for video games as we understand them today, but for potentially creating new kinds of interactive experiences that we haven’t thought about up until now.

That could be interesting—more interactive-movie-style things rather than, strictly speaking, games. I’m interested in that. I think it could be fun.

Ben Lorica. And with that, thank you, Piero. Always great to catch up with you.

Piero Molino. Likewise, Ben. Thanks for having me. It was a pleasure chatting with you again.

Exit mobile version