Stardock's Brad Wardell on Clairvoyance: Most AI Work Is Not About Replacing the Creative Work

Hello and welcome to the latest edition of AI Gamechangers, your free weekly newsletter where we quiz people across the games industry who are doing practical, inventive things with AI.

This week we’re joined by Brad Wardell, the industry veteran behind Stardock’s Windows utilities and games like Galactic Civilizations and Sins of a Solar Empire. His team has built Clairvoyance, a platform that gives you a team of AI agents living on your own computer. He explains why he thinks small teams could soon outpace AAA studios, why the real threat to big publishers is attention rather than AI, and how his Hivemind tech will put idle GPUs to work overnight.

If you’re new here, welcome! Our archive is packed with conversations with the people building with AI, and it’s free to browse and share. As ever, scroll down for our round-up of AI news, links and tools.

Brad Wardell, Stardock

Meet Brad Wardell, the founder and CEO of Stardock. Brad started Stardock at college in 1991 and has spent three decades making PC strategy games, including Galactic Civilizations and Sins of a Solar Empire, alongside desktop utilities such as Object Desktop and Fences.

Today, he’s talking about Clairvoyance, an application that began as an internal tool for Stardock’s developers and is now a free app that runs a team of AI agents on your own machine.

In this interview, Brad explains how Clairvoyance and its upcoming Hivemind feature work, and why he thinks small teams will outpace AAA.

Top takeaways from this conversation:

1. A ton of game-making time goes on grunt work such as build errors, importers and hunting for assets, so AI’s best use is freeing artists and designers to create rather than replacing them.

2. Clairvoyance gives you a team of AI agents that live on your own machine, sitting on top of tools like Claude Code and Codex, working round the clock, and remembering your projects between sessions.

3. Brad argues digital distribution, not AI, is what’s breaking AAA economics. With around 30,000 games expected on Steam this year, attention is the scarce resource and distinctive human creators will stand out.

AI Gamechangers: Let’s start with your background. Our readers will be familiar with Stardock, but tell us about the journey. How did you get to this point?

Brad Wardell: I started Stardock back in college, in 1991, and then incorporated in 1993. I was always really interested in tech. We were doing OS/2 software, and we looked at our job on OS/2 as: “Wow, this thing is so powerful, but my God, this is hard to use!” That’s where Object Desktop came from. It came out for OS/2, and a lot of the things that we enjoy today came from stuff we were doing back then that you don’t even think about.

Why is it that compressed folders are ZIP files? Why ZIP, and not one of the others? When we were making Object Desktop, we only had ZIP to work with, so we made ZIP files work as folders back in 1994, long before Windows did it.

“The thing that finally moved us was when OpenClaw came out. We’d already been doing those things for about a year, but we hadn’t released Clairvoyance to the public. It was like: this is going to happen anyway. We should make Clairvoyance public”
Brad Wardell

Back in the day, OS/2 used to hang a lot. It had a single input queue. At that time, hitting Control-Alt-Delete would just reboot your computer. So we thought: we could make an app where, if you hit Control-Alt-Delete, it brings up something like a task manager, and then you can kill the app. Of course, that eventually showed up in Windows as well.

Our thing has always been that we make games because we just like making games, and we like to make software that allows regular people, consumers as well as enterprise users, to get the most out of the tech they have on their desk.

AI Gamechangers: How did that philosophy lead to Clairvoyance? What problem were you trying to solve internally at Stardock?

Brad Wardell: There’s a lot of misunderstanding about AI because the stuff that gets coverage is things like, “Look, it made a picture.” That’s not the use case for AI. It’s like saying changing your desktop background is what computers are for. No, that’s not it.

For us, it started with: how do we get our own developers to use AI? We would have developers spending hours trying to figure out why a build wasn’t compiling. Our art is human-made, but we make custom engines at Stardock: Ashes of the Singularity, Sins of a Solar Empire, Galactic Civilizations. You get that spaceship made in Blender or 3D Studio. An artist might spend four weeks making the model, but then spend a week trying to get it into the damn game, editing a large amount of text files where all the stuff is. We thought: we have got to fix this!

I was already doing a lot with this on my own, running Claude Code and using our own really crude harness for managing old ChatGPT sessions in little terminals. This was before “agentic” even really had a name. But our developers weren’t able to do this easily. We were asking them to manage a bunch of terminals and know the nomenclature.

I’d still walk by and ask, “What are you doing today?” And the answer would be, “CMake is barfing. I spent the afternoon trying to get it so I can make the build again.” It’s like: that just cost us $500 in time. You’re not making the game.

So it started as a crude tool to let our developers, artists, designers, writers – everyone – use AI. A lot of people were already scared off because they were thinking, “AI is going to come and take all the jobs.” This is not about that. This is to save our jobs, because most of the AI work is not about replacing the creative work.

Think about writing an article. What percentage of your time is spent writing the article versus getting the damn thing up onto the site, going through all the forms and CMS crap? “You didn’t pick the thumbnail. You didn’t pick the thing.” It actually gets you to dread doing this stuff.

“Imagine you have tens of thousands of articles and you need to classify them. I’m not going to do that with ChatGPT. That’s such a waste of money. I can do it with local AI. It might be slow, but I don’t care – do it overnight. Then the question is: what machine do I do it on?”
Brad Wardell

One of the first uses of this was on Ara: History Untold, which we were making with Microsoft. The writers would send Microsoft Word documents, and to get them into the game they had to be CSV files. That’s the sort of thing I’m talking about.

I don’t think most people realise how much time is wasted on this kind of stuff. I’m not talking about editing the document or doing something creative. I’m talking about getting the thing up on the page.

If you made a pie chart of what a game developer, artist, programmer, writer, or designer does, what percentage of the time is actually design or development, versus trying to figure out, “Why am I getting error 742?” In the old days, you’d go on a journey through Google trying to find out why the thing is happening. Now it’s: no, go find it and fix it. I don’t have time for this.

AI Gamechangers: So what does Clairvoyance actually do?

Brad Wardell: Clairvoyance essentially allows you to have a staff of AI agents that live on your computer. They’re on your PC, Mac or Linux machine; they’re not in the cloud. They remember what you’ve done.

I boot up my machine and start Clairvoyance, and there they are, right where I left off. They already know how to work out of my project. I don’t have to explain how to do the thing all over again. They remember, and they get smarter and smarter over time.

Then you get into the additional things we’ve added. One of my biggest pet peeves with third-party apps is: I made a thing (like a presentation or a dashboard or a note) but now… where is it?!

So we put in document management for Markdown files, and we have what we call Exhibits.

Say I need to know what the user interface for a game should look like. Should it be tabbed? Should it be a thing on the right? I don’t know. Can you just show me what it would look like? You can ask ChatGPT to make a mock-up, but then where is it later? I made a mock-up last month. Where is it? I don’t know.

In Clairvoyance, I can actually find all my Exhibits and the stuff that’s been made. There’s task management, basically a Jira-like thing for figuring out where everything is and sharing. Anything it makes, I can right-click and share. It creates a share link. And it doesn’t care where the file is; you don’t have to put it into some special directory.

I can also access my AI “staff” from anywhere. I can be on a plane and still tell them, “Hey, I need you to fix that compiler.” The other night I forgot to turn off the sound on the computer, so all my Teams notifications were coming through. From bed I told it: mute the computer.

AI Gamechangers: You’ve also described cases where agents act without someone explicitly prompting them. How much autonomy do you give them?

Brad Wardell: The more you think of the staff as artificial people – artificial employees – the more natural it becomes to get them to do something.

I can say, “I need you to check the logs on blah blah blah and see if there’s an emergency. If there’s an emergency, send me a notification.”

Or you get a message saying, “Hey, you’re getting a bunch of clicks on your ad from Bangladesh in the last hour. I don’t think those are people.” If a bot farm comes in and starts eating up your ad spend, normally the best you can do is pay attention to it once a day. Being able to get something in real time and jump in and go, “No, stop,” is very handy.

AI Gamechangers: What sits underneath Clairvoyance? If I already use Claude Code or Codex, does it use those?

Brad Wardell: When you install it, it goes through a walkthrough and says, “Hey, you already have Claude Code or Codex installed. I’ll use those.” If you don’t have any installed, it has its own token allowance you can use instead.

If you think of Claude Code and Codex as AI operating systems, think of Clairvoyance as a shell.

“There’s a Pareto-like distribution to attention. Human beings can only keep track of so many things. The AAA industry was premised on: we have the marketing clout to make sure we’re in the top percentage of titles people pay attention to. But what really did in AAA isn’t AI. It’s digital distribution. You’re now competing against every game that’s ever been made”
Brad Wardell

It’s extremely customisable, not just cosmetically but in terms of how you use it. In the Home Desk mock-up I made, part of it is being able to take what we call Artifacts – think of them as AI-agent snippets – and put them on your desktop if you want. The Clairvoyance app can sit in your system tray, and you can interact with those as tiles on your desktop.

Underneath the covers, what Clairvoyance is ultimately doing is acting as a data manager/agent orchestrator.

AI Gamechangers: You’ve argued that a small team equipped with AI will soon outperform a 250-person studio. When did you become convinced of that?

Brad Wardell: It wasn’t really until the local AI stuff. I was not a fan of local AI until embarrassingly recently. I should know better. I felt like the old story about the founder of Digital Equipment Corporation laughing at the idea that someday computers might be small enough to fit on a desk. I suddenly thought, “I’m that guy. I’ve become that guy!”

The local AI stuff was so bad until around July. Now it’s October, and it can do all kinds of crazy stuff. That’s when I realised that all kinds of things are going to be automated, where you’re not going to spend any time at all doing the grunt work anymore.

Take a company like Stardock. We’re small, but we’ve been around a long time, and we have a ton of assets we’ve made over the years. I can’t tell you how much time we’ve spent looking for an asset. “I know we have a 3D wolf model that will fit this somewhere. Where is it?” Now I would just have a local AI agent go find it for me.

And not just on your computer – think throughout your entire ecosystem. There’s a ton of AI work people don’t even think of yet. Imagine you’re a shipping company trying to track millions of packages from a zillion different shippers. Doing that with big cloud models could cost millions of dollars. But now that local AI is good, I could set up Clairvoyance, have it do that in the background, and have it report back.

AI Gamechangers: What does that mean for AAA development? Are we entering an age in which smaller teams have the advantage?

Brad Wardell: Oh yeah. AI doesn’t make good games. I’ve seen a lot of people talk about vibe coding and AI slop, and it’s like: I lived through Unity games, right? And Flash games. I don’t want to hear about AI slop. Some of us had to write software to load images. Now you just drag and drop PNGs into Unity. That’s so much easier.

You’re going to end up with a bunch of garbage, for sure, but that’s not really a new thing. I’ve seen enough Unity platformer games to last me a lifetime.

For the big AAA studios, AI is just going to hasten something that was already happening. What it really boils down to is attention.

You can remember games that came out in the late 1990s: StarCraft, Total Annihilation, Grand Theft Auto. On the other hand, Gamescom was [in August]. Do you think the average gamer can name three games that were announced there?

“I think within a year, 90-95% of AI work will be local. Only 5% will be in the cloud, because the local models will be well into ‘good enough’. There’s going to be a lot of importance put on intelligent routing. I only talk to one [agent] and they in turn talk to the others”
Brad Wardell

There’s a Pareto-like distribution to attention. Human beings can only keep track of so many things. If you have nine things, three of them are going to get most of the attention and the other six will get hardly any.

The AAA industry was premised on: we have the marketing clout to make sure we’re in the top percentage of titles people pay attention to. That’s why you go to Gamescom and pay for ads. But what really did in AAA isn’t AI. It’s digital distribution. You’re now competing against every game that’s ever been made.

If there are 100,000 games that people can potentially pay attention to, now you have to be in a tiny fraction of a percent, and AAA studios can’t guarantee that. That makes it a bad bet.

And it isn’t only other games. For a lot of Zoomers, you’re competing with TikTok. It’s not like the old days when it was your Amiga or go outside!

The AI deluge hasn’t really begun. There’s a really interesting chart from GitHub where the number of things being put on GitHub has increased enormously, and you’re going to see the same thing on Steam. The number of titles on Steam this year will probably be about 30,000 games. Break that down, and you’re talking roughly 80 games a day. No one can even keep track of all those games. There are just too many.

And that AI onslaught really only started to become plausible for games once models like Claude Opus 5 arrived, where you could potentially make something that could go on Steam.

AI Gamechangers: If everyone has access to the same AI tools, what makes the difference between a great studio and everyone else?

Brad Wardell: It’s definitely the people. You’ll have lots of people who load up a Unity or Unreal template and whip out a level. That’s not really a change.

But if you’re using AI to create the content, your game is going to be just like a zillion other games, and people can tell. I can listen to a YouTube video now and as soon as I hear, “It’s not this; it’s really this,” I start to tune out because I know exactly how it’s structured.

“Clairvoyance allows you to have a staff of AI agents that live on your computer. They remember what you’ve done. I boot up my machine, and there they are, right where I left off. They already know how to work out of my project”
Brad Wardell

That’s going to happen to these AI-slop games that come out. They’re going to get review-bombed, and deservedly so.

What will make games distinct is the individual creators. The really top artists, writers, designers and engineers will have their excellence magnified. That world-class artist is not going to spend their day trying to figure out why the Maya importer is giving error negative 33. They’re going to actually focus on the art.

AI Gamechangers: There’s still fierce pushback against AI in parts of the games industry. What do you make of that debate?

Brad Wardell: My view is that AI is just a tool. It will be abused.

I think a lot of people don’t realise that the Band-Aid has already been pulled off. Anyone who’s been in the industry a long time remembers when Flash and Unity came out. That lowered the bar a lot.

I also think there’s a lot of misunderstanding of how machine learning works. Some people think of it as stolen, or whatever. There are also a lot of false accusations, which is really discouraging. I’ve had people claim that games we shipped in 2016 are filled with AI art. I assure you they aren’t. 2016?! One of the problems we run into is that we recruited heavily from ArtStation over the years to make our concept art, and AI was heavily trained on that sort of material. So people look at our concept drawings and go, “That’s AI art.” First of all, no: this game was released in 2016. It’s just that AI art looks like this because this is the art it was trained on.

I look at some of the stuff we released and think, yeah, that looks like something out of Midjourney now. But it wasn’t Midjourney in 2016.

AI Gamechangers: Tell us about your Hivemind product. How does that fit into the picture?

Brad Wardell: Like the other parts of Clairvoyance, it started internally. It probably won’t shock you to find out that we have a ton of [NVIDIA GeForce RTX] 3090s scattered around the company. Those computers are on pretty much 24/7, but a lot of the time they’re idle.

As we get more into data classification and local AI models, the idea is: install Clairvoyance on all the machines, let them make their GPUs available, and then the computing can happen on those machines.

If you’ve tried to set up a local model today, I’m not going to say it’s rocket science, but you have to set up your port and your endpoint. My guys aren’t going to do that. But they will install Clairvoyance.

There’s an app called Incredibuild that lets you distribute compiling across CPUs on your network. Hivemind is kind of like that, except it could be anywhere in the world and we have access to a lot of GPUs.

The idea behind Hivemind is that you can make your GPU available for computing. You put on Clairvoyance and give it permission. It needs to be a high-trust environment, because you don’t want to send your stuff to strangers. You do it with your contacts or your domain users.

This isn’t something where you just put your machine up for the public to use. At least not at this point.

I used to run a BBS back in the ’80s, and I still can’t believe I did that. Why would you put your computer up, pay money for the phone line, and let people call your computer and use it? What did you get out of that? I got nothing out of that. All those times I was doing SETI, I could have been doing Bitcoin!

If someone is using your GPU and then you need your own machine, it just kicks them off. The context is still on the other person’s machine; it’s just using your GPU for the computing.

“An artist might spend four weeks making the model, but then spend a week trying to get it into the game. We thought: we have got to fix this! This is to save our jobs, because most of the AI work is not about replacing the creative work”
Brad Wardell

The motivation is the price of GPUs and the need for local compute. We do a lot of classification of our 3D assets. We have 30 years of assets, icons and sprites. I’m not going to pay $20,000 for an RTX 6000 when our studio is filled with 3090s.

And a lot of this local-model work is something you run overnight. Imagine you have tens of thousands of articles and you need to classify them, put metadata in them and organise them in different ways. I’m not going to do that with ChatGPT. That’s such a waste of money. I can do it with local AI. It might be slow, but I don’t care – do it overnight. Then the question is: what machine do I do it on? Not my machine. I’ll run it overnight on one of the QA boxes.

For a Clairvoyance user, it’s seamless. Step one: install Clairvoyance. Step two: enable the feature. There is no step three. By default, it’s people on your domain and your contacts.

AI Gamechangers: Is there another Clairvoyance feature you think people should know about?

Brad Wardell: One of the nice things Clairvoyance does, even if you’re just using it for Codex or Claude, is its terminal. We have our own direct terminal, and we have a feature that integrates with Jev [TypeSafe AI’s non-text model].

Because it’s built into Clairvoyance, you don’t have to do anything. It can save you about 10-15% of your tokens on finding things. So much of the cost comes from having a large model go through its memories and documents to figure out which one to look at. Clairvoyance can use Jev to find that quickly and save you tokens.

We’re always trying to find ways to save tokens, especially as they’re not being subsidised as much as they used to be. I have all these people here using it, and we need to find every way we can to get the most out of it.

AI Gamechangers: Where do you see Clairvoyance and Hivemind going over the next couple of years?

Brad Wardell: I think within a year, 90-95% of AI work will be local. Only 5% will be in the cloud, because the local models will be well into “good enough”. Clairvoyance uses both; it doesn’t care.

I think there’s going to be a lot of importance put on intelligent routing. In Clairvoyance, at a given time I’ll have 50 or 60 agents going. No normal person wants to talk to all of those. I only talk to one, and they in turn talk to the others. That’s where all this is going.

There’s still nothing quite out there like Clairvoyance. The thing that finally moved us was when an app called OpenClaw came out. I don’t mean this as disrespect, but it was kind of a crude version of what Clairvoyance could do. We’d already been doing those things for about a year, but we hadn’t released Clairvoyance to the public. It was like: this is going to happen anyway. We should make Clairvoyance public.

There was some consternation at Stardock because it was a pretty big advantage for us to have this one-of-a-kind app. But what’s going to happen is that users will want to install a thing and have it just work.

If it isn’t Clairvoyance, it’s going to be something that hasn’t been made yet that is like Clairvoyance. It’s not going to be some custom operating system with agents, or some other cryptic thing to use. You’re going to install it, and it just works. It manages my data. It manages my work. The agents will seem as human-like as possible.

It’s not enough just to have a thing that you talk to. It also has to know where your stuff is, and it has to organise your stuff.

Subscribe! It’s free, and you’ll receive interviews with industry insiders and studio leaders in your inbox. We have many conversations in the pipeline, including Luke O’Leary of BisectHosting, who’s rolling out an agentic server assistant called Biko.

Further down the rabbit hole

Some useful news, views and links to keep you going until next time:

• Warhorse Studios CEO Martin Klíma says replacing game developers with AI is a “fundamentally flawed idea”. The Kingdom Come: Deliverance boss nevertheless believes the technology could enable studios to create bigger games for the same budget.

• Sony has patented an AI system that could make boss fights adapt to individual players in real time, analysing their tactics and reactions to change NPC behaviour dynamically.

• TypeSafe AI, maker of the non-text model Jev, has raised $870 million at a $7.5 billion valuation, led by Andreessen Horowitz with Sequoia and DCVC. Jev launched on 15th September. It isn’t an LLM: it outputs probabilities rather than text, and the company says a third of the Fortune 500 are already using it.

• NVIDIA and Microsoft are declaring a new era of local AI on Windows PCs, with RTX Spark laptops now open for pre-order ahead of their 16th October launch. A DGX Station for Windows was recently previewed too, aimed at running frontier-scale models and agents from the desktop.

• Be nice! Anthropic has updated its usage policy to ban “sustained and needless abusive or cruel behavior” towards its models. It’s aimed at extreme cases, and Claude ending the conversation remains the main enforcement tool.

• Magnific has launched Magnific One, an image-generation tool designed to avoid the recognisable “AI look” while enforcing company brand guidelines. It automatically checks generated artwork against a brand kit, potentially making it useful for studios producing large volumes of marketing creatives.

• Google Cloud has given a first look at Whenere, a new storytelling platform from Neal Stephenson and Half-Life texture artist Karen Laur, with a debut experience that brings Pride and Prejudice to Fortnite in UEFN. A tiny team used Google’s Dreamlands to generate 249 assets for Netherfield Park in a day, and fans will eventually write their own stories and watch them play out with AI voices.

• It’s your last chance to get tickets to PGC Nordics (20-21 October), which includes the AI Gamechangers Summit. Remember, we’re running a competition this week to give away 10 tickets!

• Meanwhile, Google and Unity have teamed up to launch Playground, a free experimental platform that generates playable games from text prompts. It initially launched for US users aged 18 and over. A more sophisticated tool, Unity Spark, promising professional-level mechanics and 3D capabilities, is coming later this year.