Real Talk on AI

The strengths and weaknesses of some of our current AI tools.

I got into game development these last few months and with AI to help me, this felt like a great time to do it.

Games are interesting to me because it seems like the most natural use case for AI.

It feels to me like the whole world is focused on improving AI for productivity reasons. I understand why, of course, productivity moves GDPs.

But we are trying to get this imperfect tool, AI, that is non-deterministic, that hallucinates, that makes mistakes to work on all of our critical things, to do our most critical work for us. Everyone wants to use AI in healthcare and every other part of the critical economy. Only a few people seem to care all that much to use it in entertainment, the one use case where the stakes are low and hallucinating can be considered a feature.

Anyway, my “vision” that drove me to game dev now (if you want to call it that) is that the current state of AI models is perfect for games & entertainment. Not just to help build the game but also to actually be in the game, imagine characters powered by LLMs etc.

But that isn’t the only reason I got into games the last few months.

I have always loved games. I ran teams of gamers since I was a 14-year-old and took them all the way to the Chicago Major League Gaming champions. My video game crew was in the top 10 worldwide of game battles for years on end. Video games are where I learned to lead people. It is where I learned what motivates them, what gets them to work as a team, and how to turn random strangers into champions.

Later in life these leadership skills would prove very valuable when I managed engineers. And so I owe a lot to video games.

And my friend

Chao Lam

reminded me recently that games are how primates learn critical skills. Not just primates, lion cubs learn to hunt by playing games too. So games are serious business.

But while it might be “just games” to some, to me it’s probably also a huge reason I got into computer science in the first place. Since I loved games so much, I wanted to learn how to make them from a young age.

But somewhere along the way I realized I could make a lot more money building software for big banks & hot startups and forgot about games. I think this is a common story. And it does not help that software engineers that work at game companies work long hours and make very little money; they get taken advantage of for their passions.

But while I know I would never do game dev for EA or Activision, there is still an itch back there that was waiting to be scratched to do it for myself.

Now that I have kids, I get an insane amount of joy having them to help me build games. In one game I had my daughter draw the characters, then we used AI to enhance them. My other daughter voiced over all the things that happened.

I mean, this stuff will be a real joy to me decades from now, so of course it’s worth pursuing, even if commercially who knows if I could ever make any money with games.

But having my kid understand that the music classes she is taking are not just for no reason, we can use that stuff in interesting ways, like in our game. Or that her drawing ability can drive the art direction of a real product she can play regularly on her iPad, this kind of stuff is priceless to me.

All of this would be significantly harder for me, most likely impossible, to pull off in the timeframe I have pulled it off without AI. I mean, just in a few months I have built 10+ games on top of some other apps

I’ll share below.

Here are two games that I built with heavy AI help, that are live on the web that you can try right now:

ZansCards.com

Comboku.com (working title!)

I’ve built more games, but will keep it brief for now because I have more I want to tell you about this building with AI world.

That’s because games are not all I’ve built these last few months. Here are two more things that would’ve been nearly impossible to do in the timeframe they got done without AI:

I built a full-blown Google Analytics replacement that basically has all of the functionality I could want out of a web analytics SaaS. Since I needed it to monitor traffic of my own sites, and Google analytics 4 is pretty terrible, I decided to make it a SaaS. After all, why not put more buy buttons on the internet?

Even with my experience, I estimate something like this SaaS app would’ve taken me a few months to build; I did it with AI in under a week. Of course the code AI produced is pretty average, I spent most of that week taking it from average to more than that. Most of the time was spent making it production ready, I use it on my own stuff, and it has a few other users now too.

lightweightanalytics.com

I built a voice recording & transcription app for the iPhone, I built it as a native iOS app. It is live on Apple app store.

I have never in my life built an iOS app; I never programmed Swift before. And this is a common theme with the things I’ve been able to make the last few months. I can program in a few programming languages but truthfully, I still don’t know how to program in Swift. But I learned enough just by asking AI stupid questions to build this.

But this is one of the greatest strengths of AI: there are no stupid questions.

If you ask a question about code, it’ll respond to you with “great question!” and proceed to give you a highly customized answer that is perfect for your dumb question. A human might've gotten offended if you asked such dumb questions that you could've googled and piecemeal the answer to yourself.

But probably a classically trained iOS engineer (and I envision my friend Martin, who built the entire JetDotCom shopping app by himself in 2015 in Objective-C when I say classically trained iOS engineer) would look at the Swift code Gemini, Sonnet, and I wrote and think, “Man, Louie, this is pretty average.”

I open-sourced the WalkWrite note taking app so you can see for yourself, and maybe he very well might look at it after this post and think that.

But nonetheless, Apple approved the app and it is live in the App Store. And you can use the open source version as a template and starting point with claude code, or similar agents, if you want to build an iOS app that hosts local models etc.

But regardless of the average code, the app works and solves a very real problem for me. I was paying $50 a year for Otter to transcribe my articles on my walks, and Otter was shipping my voice notes to a server for transcription. But that $50 a year, etc., is not even the real problem, the real problem is that Otter was trying so hard to do so many things when all I wanted was to transcribe my articles on my walks. My app uses a local AI model to do transcriptions on your device, state-of-the-art Whisper, and that's all it does. It's not interested in making millions as a VC backed thing, it was built in a few days with AI for me to use. If others find it useful, great.

If you are interested in the nitty gritty of all this, I recorded myself building with AI and put it up as a youtube video. In the video I built an eCommerce site for a friend, end to end, with AI in an hour. The eCommerce site will integrate with stripe, it has its own cart, built from scratch, just me and the AI. That video might be useful to see the decisions I made along the way in driving the AI agent. It is how I normally drive these agents.

So where is all of this AI stuff going?

I don’t know. I just know that at the very top meta level, if someone like me can build all this stuff in a few months, a lot more people are going to be expected to build a lot of stuff soon.

Putting on my “Charlie Munger’s multi-disciplinary hat” on: I recognize a capital-T truth here, the world is competitive, and not just markets. I mean nature itself is competitive, and as Munger said the more domains something holds true in, the more likely it is to be true nearly everywhere: and competition holds true in a lot of domains, you cannot escape competition.

So because of that, I feel certain most are only a matter of time away from being expected to produce more with AI, in most domains, one might only be able to be competitive if they are using AI well.

But what does being able to use AI well even mean?

Well it means understanding its strengths and weaknesses. To be able to take average outputs and make it better, quickly.

The truth is that AI as it stands today has lots of strengths but lots of weaknesses too.

Being able to use it well, to me, means understanding those trade-offs. Being conscious of the fact that a lot of work can probably be offloaded to AI, but not everything. Making the judgment calls as the AI agent works on my codebase is still critical; it could go off the rails in ways you cannot imagine. It will literally delete your whole codebase if you aren’t careful.

But still, it’s worth using, and it can do a lot well. So something as simple as: don’t let it run intricate and crazy git commands against your codebase, don’t let it run critical queries against your production DB if you have users, not without vetting those queries or changes anyway.

But ultimately, even though I think we are far from ASI (Artificial Superintelligence, I think this is the new marketing term for AGI) and AI replacing everything humans do, I am still very bullish on these tools. I still think they are immensely useful.

But the area they are most useful is in allowing a total newb, like me with game dev, to get somewhere fast, to bring that newb up to the average. For example, as a total newb at making music you can get some decent sound FX with AI; as a total newb at art you can get pretty decent art from AI; as a total newb at programming you can get a decent app out of AI.

But it's not just useful for areas you are a total newb, you can push AI further in ALL of these domains if you aren’t a newb in the domain, you can get AI to give you the average then spend the time to take it a level further. Like me with the Analytics SaaS app.

So this is why it's worth getting good at driving these AIs, in the domains you are a newbie in and in those that you are an expert in too.

But it’s got its limits of how far it can go; it’s up to you to find those limits. It’s also up to you to decide, OK, I’ve taken this far enough, now I need to get expert human help. I think this is the area I am most interested in.

Even in the games I am making, I dream someday I’ll be able to have something hit so I can hire or get other people involved in helping me make it even better. I dream I could hire a real musician to make some of the sound FX real, or a real artist to redo some of the art so it is unique, but I cannot start there. It is not feasible; it costs too much, and AI is too cheap.

And I suspect the same will apply to devs: a vibe-coding app maker that somehow gets to money probably dreams of getting a real engineer to fix up their security, make sure their architecture scales, and so on.

As my friend

Chris Wong

recently summed up: AI tools are a regression to the mean. If you’re below average in a domain, AI will make you average. For my friends who have never written a line of code in their life, becoming an average programmer at a the cost of a monthly subscription is a pretty good deal. But the opposite is also true, if you are above average in a domain then AI can bring down your work to be average, if you don’t intervene. You need to build up your judgment and experience to figure out where AI has reached its limits.

I think we are quickly moving from a world that had heavy constraints on expertise to get started with anything complex, to a world where the constraints now are on having good ideas. It does not mean expertise is worthless, real expertise will probably become more valuable as more things get to money.

But the other constraint in this new world of building with AIs is that it requires a willingness to take action on those ideas. And taking action is still hard.

Just because most people can, does not mean most people will make stuff with AI. As Pareto famously observed 20% of the pea pods made 80% of the peas and the other 80%, although they could make peas, stood by and produced very little.

I am observing the same thing in this new world where more people can make more stuff. Not everyone is taking advantage. There is a huge incentive-caused-blindness, some people don't want this tech to be that useful because it might hurt their job. So they play it down, say it does not work right, point out the times it deleted a database. Those tweets go ultra viral. Some people hate this stuff because they feel it's stealing from artists. Some other people find it makes slop.

It is my contrarian opinion here, from the trenches of working closely with AI, that all those takes are wrong. They’re not wrong about what they are saying, the tool might have very well dropped the DB. But they are wrong because they are blaming a tool, and putting all of that on what is basically just a useful tool.

But either way, the tools themselves won't care if some chooses to ignore them. And because so many are still ignoring them, that can be your competitive advantage.

There has never been a better time than now to be in the twenty percent of the pea pods making stuff.

— —

P.S. If you are a beginner at making apps with AI, I am teaching two free workshops this week on how to make an app in an hour with claude code.

Tuesday, 7pm ET: https://lu.ma/jztk5hdh

Thursday, 1pm ET: https://lu.ma/w46n7r5m

We will be vibe coding with the Act Two guys, Will & Dan live. This will be a sort of hands-on workshop, during that hour we will be building a very simple app, and deploying it live on the internet.

I’d love to have you if you are free.

---

Thanks for reading this far!

—Louie