Share

The multiplier nobody's pricing in: the web stack is where AI lives. Every model was trained on more TypeScript than on C++, and on more React than on Unreal Blueprints. Agents don't read the engine. They speak it. Fluent agents and 30ms iteration compound. Legacy stacks can't run that loop.

The economics make it worse. Unity charges per seat. Unreal takes 5% off the top, and 20-second loads bleed players. For a casual puzzle game, that's robbery.

I'm building SpellBurst on a modern stack: WebGPU-ready, Capacitor for native and browser for web, HMR iteration in milliseconds. The Firebase backend is efficient and server-authoritative. No royalty, no seat licenses, no middleware tax dressed up as "infrastructure," no clunky editor written for humans. The multiplier pays in full.

Underneath is warp5, my platform. A custom React reconciler rendering WebGL with flexbox layout. It handles auth, analytics, state sync, remote config, and OTA distribution. SpellBurst is just the first title.

I've run this play before. At Compact, the engine I built powered 15+ titles, sold tens of millions of copies, and tripled the company's value.

My bet: the next Candy Crush or Hollow Knight ships on a modern stack. Legacy teams will be at a structural disadvantage, slower iteration, higher rent, and a stack agents can't read.

In 2026, choosing the legacy stack for 2D isn't a decision. It's a failure of imagination.

I've watched this play out several times.

By the mid-90s, I'd been writing Windows code for nearly a decade. I had seen this coming early, but by then, everyone was a Windows developer. 95% of them had never written a WndProc. Everyone had the resume bullet. Some had the vocabulary. Few had the receipts.

Then the dot-com era turned every business into an internet company. I once overheard a guy at the next table tell his date he was a web entrepreneur. He owned a pet store with a website.

If you've watched Silicon Valley, you've already seen this. Pied Piper was a compression algorithm that became a platform, then a video chat app, then whatever the room wanted it to be. The joke was that the technology never mattered. The pitch did. The world doesn't always reward substance, but it almost always rewards theater.

The larger trend gets sorted the way these things always do. Maybe the next model release tells you who built a company and who built a feature. Until then, watch the man behind the curtain.

Subscribe now

The populations are anchored in early Amplitude sessions: word-length shifts across moves, late-game streak maintenance, and how people push when they're one star short.

The model has two skill axes, word-finding and strategic play, each a mixture of Beta distributions tuned by tier. Vocabulary is weighted by real-world frequency from Google Ngrams, so low-skill players only reach the common words, and veterans find the rarer ones. Board scanning is probabilistic, so sometimes they see everything, and sometimes they miss the obvious. A Markov chain models how players shift mid-game as pressure builds.

To fix the big combo reward, I raised the threshold and made it rarer. The levels improved immediately. Not because I added anything. Because I removed the noise.

Every level gets a report card across a dozen dimensions: reward frequency, chain density, board intensity, star distribution, and endgame engagement. A level can look good on nine of them and still be wrong on the tenth. The simulator flags the anomaly before I'd even notice it. The 5.5 combo was one finding. The simulator has surfaced dozens.

By the time SpellBurst lands in the app store, every level will have been shaped by a population that played it thousands of times first. That kind of calibration used to take a big studio months of live data. Mine runs over lunch.

This is how a solo founder gets big-studio polish without a big-studio team.

Subscribe now

Most of what large teams produce isn’t progress. It’s coordination overhead. The meetings about the meetings. The alignment docs. The quarterly roadmaps.

AI doesn’t fix that. It accelerates it. A large company running AI through six layers is just shipping the wrong thing faster.

The real advantage AI unlocks isn’t throughput. It’s the gap between an idea and a working prototype. In a big organization, that gap is weeks or quarters. Decisions move through layers. Priorities get negotiated. Calendars are full.

For a small team or a solo developer, that gap is an afternoon.

I’ve lived on both sides. Early in my career, I shipped CD-ROM titles on factory deadlines. Miss the ship date, miss Q4. Miss Q4, the company folds. No one talked about process. Short cycles, direct conversations, work that mattered. This was before the consulting class co-opted agile and turned it into status theater. When the company was acquired, the engineer who did due diligence on the codebase called it “a work of art.” That doesn’t come from the process. It comes from ownership.

Ownership scales down, not up. The further a builder sits from the decision, the less skin they have in the outcome. Jira is a monument to this failure. A trust deficit that says: I do not trust you to be a professional. The trend toward narrow lanes: engineer, PM, designer, each in their own silo, didn’t just slow things down. It stripped ownership from the people doing the work. Ownership becomes compliance, and large orgs spend enormous energy recreating the culture they’d have had if they’d trusted people. Nobody moves mountains for a ticket. It’s possible that’s starting to reverse. That these roles are distinctions the next generation of builders won’t recognize.

Now multiply all of that by AI. The tool that lets you chase a hunch before the moment passes.

I saw a workshop demo of an OpenClaw application. Although unrelated to my work, it inspired me. I realized I could use the Claude SDK directly in my application. By midnight, I had a proof of concept. By Monday, a working prototype.

That’s what the “AI favors scale” crowd gets wrong. AI is an amplifier. It doesn’t care what it amplifies. In a large org, it often amplifies compromise. The spec no one owns. The direction no one would have chosen. For a solo builder, it amplifies conviction. The deep familiarity with the problem. The clear thesis about what to build.

The leverage was never scale. It was clarity.

Subscribe now

In tech, your history often doesn’t travel. Every new room, you start at zero in someone’s eyes. You were an engineer, a designer, or a product thinker, one slice of yourself sometimes on a dead road already mapped out.

The system preferred it that way: easier to schedule, easier to replace, easier to stack rank. What it couldn’t schedule was the part that actually mattered — the ownership, the agency, the willingness to care about something beyond your own lane. When you constrain a person to a tidy HR-sanctioned role, everything else they have to offer has nowhere to go.

This past year, for the first time in decades, I was whole again and knew it. We are more than the roles we’re handed. A game almost in the app store is my proof: built by the multimedia pioneer whose work lined store shelves, the animation ringer who built a screensaver in six weeks that helped sell a startup for $38 million, the SRE who single-handedly kept a walking corpse with 2.4M players on its feet, and the product thinker who learned to measure resonance before committing to build.

Subscribe now

We're in one of those moments again. The floor is shifting. I know because I'm standing on it. In 1987, the democratization of the interface made computers accessible to millions. Now it's the democratization of software development itself.

I build with AI every day. It's fast, especially with boilerplate. Build me a website. That's not magic, that's a simple and well-understood problem done at extraordinary speed. But that speed changes something fundamental. The cost of trying a hunch has collapsed.

The hard part is everything else. Any serious work is a thousand-bound problems that need to fit together. The AI can solve any one of them. But it doesn't know which ones matter, how they connect, or where the weight is. Three weeks before a game launched, I was handed the requirements for a matchmaking system capable of handling 100 million players. I projected the criteria into a single dimension and repurposed an existing leaderboard service. That intuition comes from experience and made an impossible task possible.

I spend hours before the tools touch anything serious. Decomposing, sequencing, and finding the load-bearing walls. That's the work. The rest is fast. You can't outsource the thinking without first doing the thinking.

All this has happened before. I hear the melody. It was on vinyl the first time. I'm not waiting around for the chorus. The needle has dropped. Again.

Subscribe now

Because it's not about how many times the ground shifts or the studio closes. It's about how many times you can find yourself back at zero and still find a way forward to one.

Maybe that's what the saying really means. Not the surviving. Getting back up and moving forward. Again. And again.

Subscribe now

What if a computer could do this? Not just stitch together pre-written bars, but use probability to generate music. In 1984, that was a wild idea. MIDI didn't even exist yet. There was no way to build a working prototype.

So I wrote the argument instead. Sixty pages. Markov chains were well established in mathematics — I just thought they could model music. Transitions between musical states. Poisson distributions for timing and event frequency. Heuristics to keep it from sounding like noise. I couldn't prove it worked, but I could prove it should work.

I didn't think of it as AI. The word never appeared in the paper. It was just a research paper by a kid who liked Mozart and had a background in math. It was 1984. I was 21.

The scale of today's generative AI is unrecognizable, but the intuition is the same. Markov processes, probability distributions, and pattern completion.

I wasn't predicting anything. I was following a thread that started with a movie ticket and a college elective. The best ideas don't announce themselves. They just show up — if you're paying attention.

Subscribe now

When the company relocated to San Francisco that fall, most of us went together, landing in the old Hamm's Brewery in Potrero.

It was the kind of team where the line between colleague and friend didn't really exist. You don't get that bond often. Maybe once or twice in a career, if you're lucky.

Tom Cole was the glue. When I struggled adjusting after the move, he made a point of hanging out after work. That was just Tom—he was like that with everyone. We lost him too young.

Mark and Sunil had the vision. I got to build a piece of it alongside people I liked. This choice shaped my journey in ways beyond the work.

When the AAA studio I'd been working for closed last year, I recognized the fork. Same one as thirty years ago. Same instinct. I'd arrived where I started—and knew the place for the first time.

The seed was planted just before the studio shuttered. A professor friend asked a question over lunch that followed me home. I began deconstructing the games I loved—the ones that stay on my phone for years.

I found a creator's story I'd never heard before. Seeing how they'd solved a specific problem made everything click. I pulled the thread and landed on a combination of mechanics I hadn't seen anyone else put together. Not a small idea. The kind you can't walk away from.

Within weeks, a prototype. Within months, something I couldn't put down.

And the tools caught up. What once required a team and a room full of extension cords, I can build from a sunroom. Some people outgrow that instinct. I just found a nicer office.

Subscribe now

Up until then, I had worked at big companies where ownership gets diluted across layers. Bob did the opposite. He made us feel like the work was ours, in the way that matters day to day: responsibility, agency, and pride.

Bob trusted you to change the system if you could see a better way. I pushed us toward SGML markup because I’d spent the previous few years at IBM building SGML tooling and saw the fit. Hugh Matlock implemented it, and we shipped our own markup as plain-text files that are still sitting on the discs. Anyone with a copy can open them and see exactly how it worked.

Hugh also did many other remarkable things. One that stands out was adding IDs to tags so content became addressable in code, something we shipped before we had heard of HTML, and long before the web standardized a DOM. Mike Bradshaw ran with this and used those tools to build experiences people later associated with "dynamic web" interfaces. Hugh's insight made all that possible, and Bob's culture made it normal.

You would mouse over one element, and a plane from the other side of the screen would roar into view. A frosted-glass pop-up would crossfade with stats. It looked like a single seamless screen, but under the hood, there was an array of multimedia choreographed by our markup. That's ordinary today, but this was BEFORE NETSCAPE, and IMO came about because of the environment Bob fostered. It felt like the future because, in a small way, it was.

Compact was acquired by The Learning Company (SoftKey), but the part I keep coming back to is this: Bob trusted people and gave them room to innovate. That trust shaped me. It became my north star. I am at my best when I can make a difference, not just execute.

I owe him more than I ever told him. I wish I had.

Thank you, Bob.

Subscribe now

A lot of my thinking came from a simple intuition. If binary search partitions a one-dimensional space with a small decision tree, then a two-dimensional space could be tamed with something one dimension higher. That led me to build a three-dimensional tree that avoided most unnecessary comparisons.

I left Xerox before the project was finished, but I licensed a similar dataset and completed the work around a year later. My search ran about 10x faster than the brute-force version in my tests, which felt like a real breakthrough.
Then reality hit.

I had made the classic mistake. I never measured the real bottleneck. I had optimized the part of the system that interested me the most, not the part that actually mattered. The search accounted for only about 10% of the total time. So even with a tenfold speedup, the overall gain was just a couple of seconds.

The other twenty-plus seconds came from drawing the vectors to video memory. My clever data structure did nothing to help with that. I shelved the work and felt embarrassed. A few years later, I came across an article describing this spatial indexing structure that looked nearly identical to what I had built. That felt good, and it told me the intuition wasn’t entirely off.

But the real lesson had nothing to do with the structure itself. It was much simpler. The most innovative solution in the world is worthless if you solve the wrong problem. Measure first. Figure out where the real time is going. Then decide what to optimize.

It’s a lesson I learned the hard way and have carried with me ever since.

Subscribe now

After decades in this industry, I’m calling this the Engineering Singularity: the gap between architectural intent and production has collapsed. A massive undertaking no longer requires the coordination or friction of a traditional studio.

AI is a bus that can take you from Point A to Point B faster. But it will never tell you that you’d be better off going to Point C.

That part’s still ours.

Subscribe now