The Island of Dr. Brain

My grandfather Lou got me excited about computers and math before I was ten, in the late eighties, reading me things out of PC World. The game I remember playing with him is The Island of Dr. Brain, a Sierra puzzle game from 1992 where you work your way through a mad scientist’s security system one logic puzzle at a time. Somewhere in there a computer stopped being a machine and became a puzzle box, a thing that would give up its secrets if you were patient and curious enough, and I have loved them that way ever since.

I am a child of the internet. Not the one that exists now, the one that came after that game: modems, shareware, a browser with View Source (and the occasional website hiding secret text in the source comments, ahhh, the nineties), forums full of strangers who would answer a question from a fourteen-year-old at two in the morning for no reason at all. Most of what I know how to do, somebody I never met taught me for free.

The tools I use all day are now built by companies, and companies have shareholders. I don’t think the people inside them are any different from the strangers who answered me at two in the morning. It’s just that nowadays one of them would have to explain to their boss why taking five minutes to post an answer on Stack Overflow increases shareholder value.

The ideas underneath that old internet shaped how I think about software more than any job I’ve had, and the thing I’m building now comes out of them directly.

Where the manners came from

Most of it was written down before I could read.

Steven Levy published Hackers in 1984 and wrote down the ethic that had been floating around MIT since the fifties: access to computers, and anything that might teach you something about the way the world works, should be unlimited and total. All information should be free. Mistrust authority. Judge hackers by their hacking. It reads a little like scripture now, and the people involved would have hated that, but it was a description of how they behaved around a PDP-1.

A year later Richard Stallman published the GNU Manifesto, which is where the manners got stated as an obligation:

The fundamental act of friendship among programmers is the sharing of programs.

His version of the rule is that if he likes a program he must share it with other people who like it. You can argue with the man’s positions for a week. The instinct underneath is the one I grew up inside.

In August 1991 a student in Helsinki named Linus posted to comp.os.minix that he was doing a free operating system, “just a hobby, won’t be big and professional like gnu”, and asked what people wanted in it. In April 1993 CERN put the World Wide Web software into the public domain. The web could’ve been a product. Somebody (was it Al Gore? I forget) decided it would be a thing you were simply allowed to have.

I didn’t read any of this at the time. I was a kid with a modem, and what I absorbed wasn’t the manifesto but the behaviour it produced: you asked, a stranger answered, and when you knew something you answered back. Nobody explained it as a philosophy. It was just how you acted if you wanted to be part of that place, the same way you learned not to type in all caps. I took the lesson before I knew it was one.

I was ten for most of that, on AOL, so the manifestos reached me second-hand. Games reached me first-hand.

The open game

In January 1993, before the game existed, id Software sent out a press release for Doom that said, among a great many other claims, this:

DOOM will be released as an OPEN GAME. We will provide file formats and technical notes for anyone who wants them. People will be able to easily write and share anything from their own map editors to communications and network drivers.

They had watched what happened with Wolfenstein 3D, where people built map editors and trainers without being asked, and instead of fighting it they designed the next game around it. The first episode was shareware. You copied it from a friend, or a BBS, or the cover disk of a magazine, and if you liked it you sent them money for the rest. The WAD format meant that within months the internet was full of levels made by strangers, most of them terrible, some of them better than the ones that shipped.

At the time I was just playing Doom and relishing it, and had no idea what an Open Game was. My contribution to the medium was more of a hint of things to come: I worked with my buddy Zack Snow on a COMPLETELY UNRELATED game called Hell Fighter, where you go to Hell and fight Hitler, and it was pencilled into the margins of my ninth-grade social studies book. I digress.

In December 1997 John Carmack released the Doom source, “at long last”, with a note that they couldn’t release the DOS sound code because of a copyrighted library (“wow, was that a mistake – I write my own sound code now”). Two years later, on December 21, 1999, Quake went out under the GPL as “a Christmas present for the coders”. The readme explains why he chose the GPL over something looser: there had been unfortunate losses to the community from mod teams keeping their sources closed and sometimes losing them.

The guy who wrote Doom and Quake was worried about mod teams losing their code.

Valve shipped the Half-Life SDK on terms that read, even now, like a dare: you may distribute your modified Valve game in source and object code form, but only for free. In 1999 a website called itself “Counter-Strike: A Half-Life modification” and announced that Beta 4 was out and nearly every file in the mod had changed. A few years later some people made a custom map in the Warcraft III editor called Defense of the Ancients. Two of the biggest competitive games on the planet are, at the root, things strangers made for free with tools a company decided to hand out.

That was normal. That was what a game was: the thing, and the way the thing was made, in the same box. I grew up assuming you could open the box and look inside at all the pieces.

View source

The web had the same property built into the browser. Every page was its own tutorial. Every page told you how it was made whether the author meant it to or not, and there was no other way to learn it, because the tool came free with the browser and nothing else did. Find a page that did the thing, look at how, copy the part you needed, break it, fix it. I remember CSS breaking my brain the first time I tried to understand how one file could change every page on the site at once. It seemed like cheating.

The protocols underneath assumed the same thing about people. Postel’s rule in RFC 761, 1980: “be conservative in what you do, be liberal in what you accept from others.” Read it as a security posture and it makes no sense. It only works if you assume the stranger on the other end of the wire is trying to cooperate and might just be slightly wrong. Dave Clark’s line from 1992 about how the IETF ran, quoted in RFC 7282: “We reject: kings, presidents and voting. We believe in: rough consensus and running code.”

Wikipedia launched in January 2001. Jimmy Wales, in 2004: “Imagine a world in which every single person on the planet is given free access to the sum of all human knowledge.” In April 2008 GitHub went live with a three-line post thanking the beta testers. Six days later Joel Spolsky announced Stack Overflow:

Free to ask questions, free to answer questions, free to read, free to index, built with plain old HTML, no fake rot13 text on the home page, no scammy google-cloaking tactics, no salespeople…

Atwood’s post the same day says the intent was “collectively increasing the sum total of good programming knowledge in the world.” Every answer went out under a Creative Commons licence, which is a detail that will matter in a minute.

It was never a utopia

Look, I’m not going to pretend. In 2009 I wrote on my blog that the internet was one of the strongest strongholds of racism in the universe and that I’d heard more slurs in five minutes of Call of Duty than in my entire life walking around in meatspace. Both of those things were true at the same time as everything above.

And the commons was always being carried by somebody. Nadia Eghbal’s word for the shape of modern open source is a stadium: a couple of people at the helm and a very broad base of users with little context for what’s actually happening, and even the maintainers who are paid can burn out just from the scale of the public interaction. xkcd 2347 is the whole argument in one drawing: all modern digital infrastructure balanced on a project some random person in Nebraska has been thanklessly maintaining since 2003.

Daniel Stenberg, who maintains curl, has been documenting the newest version of that load. In 2024 he wrote that with AI-generated bug reports, the better the crap, the longer time and the more energy we have to spend on the report until we close it. By mid-2025 about a fifth of curl’s security reports were slop and the valid rate was around five percent.

So no. Not a utopia. A gift economy with a couple of exhausted people at the bottom of it. But the direction was clear, and the default was to share.

Everything eventually gets a tollbooth

Here is what I think went wrong, and it wasn’t one villain. It was one move, made over and over, by people who each had a perfectly good reason.

Take a commons. Realise you need, or want, to make some money from it. Put a layer on top. Charge for the layer. Keep what’s under it.

The web’s first layer was advertising, and I think it was the original sin. Not my phrase: Ethan Zuckerman, who wrote the code for the first pop-up ad at Tripod in 1997, used it to apologise in 2014. “I have come to believe that advertising is the original sin of the web.” Once the page is paid for by whoever is watching you read it, the page’s job stops being to help you and starts being to keep you, and that is the straight line from a GeoCities guestbook to the attention-grabbing slop that fills every feed now. It is the reason the web feels the way it feels. But it is a different essay, because it is about what happened to readers. The version of the move that happened to software, to the people making the thing, starts in March 2008, when Apple announced the App Store and the terms in one sentence: developers set the price and retain 70 percent. Be fair to this one: it worked. It made paying for software normal for people who had never done it, and it killed the CD-key and the shareware nag screen in one go. It also put a landlord between every developer and every user, and the landlord decided what you were allowed to ship.

And it put monetisation at the front of the room. Before the store, you made a thing and then, maybe, figured out how to get paid for it. After the store, every tiny app had a lottery ticket attached: one of them was going to be Flappy Bird, and the guy who made Flappy Bird pulled it because it was making him miserable, not because it stopped making money. You could still build a good tool. There was just no guarantee anybody would ever see it, and a very good chance that the thing next to it, faster and more addictive and built in a weekend, would. The store didn’t reward the way a thing was made. It rewarded the thing that got tapped.

Two years later Wired ran “The Web Is Dead. Long Live the Internet”, and the dek said it plainly: sleeker services, “think apps”, were “less about the searching and more about the getting.” There is no view source in an app.

In May 2013 Adobe stopped making Creative Suite. There would be no new version to buy. You would rent Photoshop from now on, and if you stopped paying, the thing you had learned on for a decade stopped opening. The industry watched this succeed and took the note. Almost nothing you use today is something you own.

Games took the move too, and it hurt more because of where they had started. In June 2019 EA’s vice president of legal sat in front of a UK parliamentary committee and explained that they didn’t call them loot boxes. They are “surprise mechanics,” and they are “quite ethical and quite fun.” The industry that had once handed out the level editor and the source was now charging for the surprise.

That was the loud version. The quiet version is volume. Steam shipped more than nineteen thousand games in 2025, a record, and roughly half of them never collected ten reviews. Anything that works gets cloned within a month, the clones get cloned, the price of the clone goes to zero, and the only thing left to compete on is how well you can keep somebody spending after they bought it. So every game needs a battle pass, and a season, and ideally a subscription, because the people paying for the studio need recurring revenue and a game you buy once isn’t that. I spent years shipping live-service games. I’m not throwing this from outside the house. The box that used to come with the editor now comes with a store.

Cory Doctorow eventually gave the pattern a name, enshittification, and his definition is worth quoting in full because it is the move again, in three beats:

Here is how platforms die: first, they are good to their users; then they abuse their users to make things better for their business customers; finally, they abuse those business customers to claw back all the value for themselves. Then, they die.

Even open source did it to itself. Elastic moved Elasticsearch off Apache 2.0 in 2021 because the cloud providers were selling it as a service without contributing back, which was true. HashiCorp moved Terraform to the Business Source License in 2023 and the community forked it as OpenTofu within six weeks. Redis went source-available in March 2024 and, a little over a year later, added the AGPL back with a line I keep rereading: “This achieved our goal—AWS and Google now maintain their own fork—but the change hurt our relationship with the Redis community.”

Every one of those companies had a real problem. The move was still the move.

No commons, no model

The models learned to code from the commons. Not metaphorically. From it.

OpenAI’s Codex paper, the model that became GitHub Copilot, says its training set was 54 million public software repositories hosted on GitHub, 159 GB of Python once filtered, collected in May 2020. The Pile, the open training corpus a lot of 2021-era models were built on, is 5 percent Stack Exchange and 7.6 percent GitHub, with Wikipedia, Hacker News, and the Ubuntu IRC logs in there too. Meta’s LLaMA paper lists Stack Exchange, GitHub, and Wikipedia as three of its seven sources, and notes that the GitHub slice kept only projects under Apache, BSD, and MIT licences. The Pile’s authors say they could use Stack Exchange because the dumps are public under CC BY-SA. That is the Creative Commons licence from three sections ago. Atwood and Spolsky chose it in 2008 so that the answers couldn’t be enclosed. Fifteen years later that choice is why a model was legally allowed to read all of them.

Take Stack Overflow, GitHub, and Wikipedia out of the training data and you do not get a model that can code. You don’t get Codex. You don’t get Copilot. You don’t get Cursor. Every answer a stranger ever wrote for free at two in the morning, every WAD, every view-sourced page: that is what the thing is made of. The manners built the machine.

And the place the manners lived is emptying out. Stack Overflow received 108,563 questions in November 2022, the month ChatGPT launched. In December 2024 it received 25,566, fewer than any month since May 2009. In January 2024 Stack Overflow’s own blog wrote that LLM developers have an obligation to contribute back to the communities that create the data they leverage; four months later it licensed the corpus to OpenAI. Nobody is being evil here either. You ask the model because the model answers faster and doesn’t close your question as a duplicate. I do it too.

But a Stack Overflow answer was never just the fix. It was the fix with the reasoning under it, three comments arguing about whether it was safe, an edit from 2019 saying this is deprecated now, and a link to the docs the stranger had read to work it out. You got the thing and the way it was made, and you learned something on the way past. When the model answers, you get the thing. Whether you also get the way it was made is now entirely up to whoever built the tool you asked it through.

I can’t do much about whether the commons survives. What I care about is where the manners go now that the answers come from a machine.

The newest layer

Which brings me to the tools I use all day.

Andrej Karpathy coined “vibe coding” in February 2025, and the tweet is more honest than the term became: “I ‘Accept All’ always, I don’t read the diffs anymore… The code grows beyond my usual comprehension.” The tool in the tweet was Cursor.

Cursor isn’t a villain, and it can’t be bad software; SpaceX finished buying it in August in an all-stock deal that valued it at sixty billion dollars. The deal had been in motion since April, after SpaceX had already acquired xAI and put its debt and spectacular appetite for compute on the rocket company’s balance sheet. SpaceX’s own filing says what Cursor adds: the application layer on top of its compute infrastructure and models. It already had the compute bill and Grok. Cursor gave it somewhere developers might actually pay to use both.

It is a very fast, very good editor with a model layer handling four hundred million Tab requests a day, and the economics of that are brutal: frontier tokens cost real money, a subscription resells them, and at least one report has put the company’s gross margins below zero. If you are in that position, every layer you put between the user and the model is a place you can protect a margin.

And so the layers went in. “Auto” picks a model for you, and when a user asked on the forum which models were behind it, the answer from staff was that the pool “is not fixed and can change over time” and there is no public list. Composer 2 shipped in March 2026 as a model of theirs, at a price of theirs, without saying what it was built on; a user found the Kimi base in the API traffic within a day, and a co-founder said it was a miss not to have mentioned it. The 2025 pricing change needed an apology post and refunds. In fairness: Cursor does show a collapsed thinking block, and GitHub Copilot’s auto-router at least tells you which model it picked. But it isn’t one company. Windsurf’s in-house SWE-1.5 was post-trained on “a strong open-source model” it doesn’t name either, and Kiro sells you credits that even your hooks burn.

The thinking itself got enclosed one level down. When OpenAI shipped o1 in September 2024 they decided not to show the raw chains of thought to users, and listed “competitive advantage” among the reasons in their own words. Anthropic’s docs say what you see is never the raw chain of thought, and you pay for the full thinking tokens either way. I understand why. It is still the move: the thing, without the way it was made.

What gets me is how much this looks like another chance to extend the old manners. Thorsten Ball built a working coding agent in under four hundred lines to show there is nothing magic in the loop, and Geoffrey Huntley’s Ralph, the early precursor to every agent loop people run now, is literally one line of bash: while :; do cat PROMPT.md | claude-code ; done. That’s it. That’s the whole secret sauce, out in the open, the same way the WAD format was. An agent leaves a fuller record of its work than I ever left of mine: what it read, what it ran, what it changed, what it cost. That record could be the most legible thing about software work. Or it could be the new thing under the layer.

What a child of the internet builds

Look, of course I would like LodeFlow to make money. Everybody would like the thing they keep coming back to after work to turn into a business, and I’m not some special exception to this, so I keep thinking of ways to monetize it.

Memory could be a paid tier. The useful analysis could run on LodeFlow credits instead of the Claude or Codex account somebody already pays for. The instructions that make it work could stay hidden, because now the hidden part is the thing I have to sell. These aren’t evil ideas. They’re extremely normal ideas, and that is how the tollbooth gets built.

I keep coming back to a simpler bet: make great software, solve a real problem, and leave people enough of a handle to build something with it I couldn’t have imagined. The cost of rebuilding software keeps falling anyway. A hidden prompt isn’t much of a moat.

LodeFlow already hosts the real Claude Code, Codex, and Kimi CLIs in real terminals. I made that argument before as a fidelity argument. It is also this argument. Whatever the CLI exposes gets to stay exposed, without a LodeFlow-shaped version of the agent in the middle.

Now LodeFlow is starting to learn. The first piece is memory: it reads recent sessions, finds the decisions, traps, and conventions worth carrying forward, and proposes them for review with the evidence attached. It uses the model account you already pay for, and what you reject doesn’t come back next week wearing a fake moustache. What you keep lands in ordinary files on your machine.

Earlier this year Anthropic accidentally shipped a source map that exposed roughly half a million lines of Claude Code’s internal source. Not its CLAUDE.md, exactly; people got the prompts, the memory plumbing, the tool orchestration, the whole pile of judgment around the model, and then spent weeks losing their minds over how much was in there.

Seeing it didn’t make Claude Code less impressive. It made the thing more impressive, because now people could see how much thought had gone into turning a model into a tool. And why was that part secret to begin with? Claude Code still burns Claude tokens. Anthropic still makes money. They could make the client incredibly thin and it would still be an incredible product.

Giving people the WAD format didn’t make Doom less of a game. Seeing how Claude Code was put together didn’t make it less of a product. The second people had a handle, they started building with it.

That’s what I want for LodeFlow. I want it to solve the annoying part: gather the work, run the job, keep the evidence straight, make it pleasant to review, and carry the result into the next session. I also want the person using it to be able to open the box, understand why it behaved that way, argue with the instructions, and hand back something better.

If somebody uses LodeFlow to build something amazing that I never saw coming, that isn’t a hole in the moat. That’s the win.

And if you happen to run a rocket company currently moonlighting as Grok’s balance sheet and want to buy this editor for sixty billion dollars in stock, feel free to give me a call.