Splitting the Atom of Software: From Towers of Babylon to Small, Safe Local Reactors.
When we split the atom, we were promised free energy—instead we got Hiroshima, Chernobyl, Fukushima, and H-bombs, while waiting forever for cold fusion. Yet we also built thousands of safe reactors and compact naval reactors. Can we build the peaceful, compact reactor equivalent for AI: small, self-contained, safe, abundant local software, logic, and automation—guided by Micro-Waterfall instead of building more Towers of Babylon?
- I. Start Steno + Oppenheimer/Ragamuffins
- II. Splitting the Atom & Small Reactors
- III. Pay-to-Win Before AI: Capital vs Clicks
- IV. 2026: Parachutes & Acceptance Testing
- V. Stolen IP Rented Back to Us
- VI. Sarah Connor: “No Fate”
- VII. Local Meeting App & Single-Speaker
- VIII. 1-Day Build & 45m Realtime Bottleneck
- IX. Raggamuffin & Towers of Babylon
- X. Aristotle’s River & Micro-Waterfall
- XI. 15:00 Stop Steno & Revisit at +15:00
Custom Editorial SVG Illustrations (11 Slides)
Eleven self-contained vector illustrations covering every metaphor, technical pipeline, and architectural model in the talk. Select any card to present.
Bullet-Point Synthesis of the Discussion
1. The Atomic & Reactor Metaphor
- • Oppenheimer vs. Ragamuffins: Tech leaders quote “The nuclear bomb, I have called death, the destroyer of worlds,” while on the ground, “Now we’re all ragamuffins.”
- • Splitting the Atom: We were promised free energy. Instead we got Hiroshima, Nagasaki, Three Mile Island / Long Island, Chernobyl, Fukushima, and hydrogen bombs—and we’re still waiting for cold fusion.
- • Small, Safe Local Reactors: Yet thousands of nuclear reactors operate safely, and the U.S. Navy runs compact, self-contained reactors safely (except for using them to go around the world killing people).
- • The AI Equivalent: Can we build the peaceful compact reactor for AI—small, self-contained, safe, abundant local software, logic, and automation?
2. Pay-to-Win (Before AI) vs. 2026 Paratroopers
- • Pay-to-Win (Before AI): No assistant magic—just brute capital: faster hardware, bigger data lakes, and hiring more employees or outsourced contractor agencies.
- • 2026 Paratrooper Clicks: What once required specialized training (e.g. paratrooper school or years of systems hacking) is now accessible in a click. Everyone parachutes straight into the rooftop pool.
- • Direction Matters More at High Speed: The faster you go, the more important it is to go in the right direction. High-velocity vibe coding without direction is catastrophic.
- • The Acceptance Testing Crisis: Human user testing and acceptance testing remain critical. If you had 5 testers before and now output 10x more apps or changes—who will tell you if any of it is actually good?
- • Sarah & John Connor: “No fate but what we make.” Reclaiming agency by running models locally.
3. The Local Meeting App & 15-Minute Live Demo
- • Undistributed Future: “The future is already here, but not even they have distributed it.”
- • Real-World Friction: Open-source Android app (`Fossify`), OS Voice Isolation noise-gate traps disabled, zero-cloud transfer (`LocalSend`), and local models (`NB-Whisper` + `Pyannote` + `Ollama`).
- • Live 15-Minute Cadence: Run `Steno` live on stage from 00:00 to 15:00. At 15:00, stop recording and trigger background processing. After another 15 minutes (during Q&A), revisit the transcript and summary.
- • Single Speaker vs. Meeting: A meeting with one speaker is a (little stiff) presentation. The app uses different modes for a single speaker—it assumes a lecture format. Transcription is similar, but notes are structured as presentation takeaways rather than multi-person action items.
- • 1–2 Hours Work vs. 45-Minute Realtime Bottleneck: Built in 1 day (1–2 actual coding hours), bottlenecked by having a 45-minute meeting where audio conversion took roughly 45 minutes (roughly realtime). Depends on knowing the right questions to ask the AI.
4. Raggamuffin, Aristotle’s River & Micro-Waterfall
- • Raggamuffin & Dancehall’s Long Tail: Digital riddims produced timeless masterpieces alongside a long tail of very poor dancehall music—the exact power law of 2026 vibe coding.
- • Towers of Babylon: “Babylon stands very tall in reggae culture”—and we are building more Towers of Babylon than ever simply because we can.
- • Dlang + C++ vs. Rust Everything: Before, you wrote a surgical Dlang core linked to short C++. Today, people rewrite entire codebases in Rust with 900 crates and zero architecture.
- • Aristotle / Heraclitus: Can’t Swim in the Same River Twice: In software, the code stream is constantly moving. Without human architecture at the top, every AI re-prompt tosses you into an entirely different river of dependencies and bugs.
- • Micro-Waterfall (The Way Forward): Human steps at the top of the waterfall (architecture, constraints, right questions), AI executes downstream, iterating in 1 day or 1 week.
Complete Set of Notes — Everything Spoken Across All 5 Prompts
Preserving every quote, example, timing detail, and philosophical thread dictated during the session.
“Idea for a talk about the state of AI, local AI, and development in 2026. Core idea is some quotes:
- ‘The nuclear bomb, I have called death, the destroyer of worlds.’
- ‘Now we’re all ragamuffins.’
- Sarah Connor, John Connor.
- ‘No fate but what we make.’
- 2023: pay-to-win, doglike, pass through security control quickly, priority boarding.
- 2026: everyone is parachuting into the pool on the roof of the hotel.
- Immense power.
- ‘They stole all intellectual property and rent it back to us.’
It’s not really democratised, but local AI may change that. These are some ideas. Come up with some illustrations and a rough sketch of a talk in self-contained HTML.”
“Example: The future is already here, but not even they have distributed it. When it gets here, try to: record locally via a local LLM, record a meeting, transcribe it, make meeting notes, and identify speakers.
How hard is it to find a good, say, open-source Android app to actually record in good quality and correct microphone settings? How easy is it to get that recording, without copying it to a cloud service, transcribe it, analyse it, and create meeting notes?
Today we’re also going to launch such an example. I have it running in the background here: a Python app that runs on all desktop platforms can do it. It’s safe. Not everyone could have made this in one day, but I did. We want to help everyone to be able to do it in one day with web coding.
We’re going to let this run for 50 minutes, so we’re going to see what it comes up with as meeting notes for this talk so far (50 minutes). It took me one day, and part of that was because I had a 45-minute meeting. Every time I tried to redo the job, it took 4 to 5 minutes—almost real time to transcribe—and then some additional time with the rest of the processing pipeline. Actually, work hours for me were maybe 1 to 2 hours of work time, 1 to 2 days. Still incredibly fast for such an application, but it depends on you knowing the right questions to ask the AI.”
“Final idea: We’re all ragga muffins now. Reggae and dancehall have incredibly good music, with a long tail of very poor dancehall music. But this idea of the Babel Tower / Babylon, and ‘Babylon stands very tall in reggae culture,’ should also stand tall in our minds as developers.
We are building more towers of Babylon now than ever before because we can. Before, you might want to rewrite something in a safe language, so you’d write a small D language core and then link to your short C++ code. Now you rewrite everything in Rust. We already had Rust and D lang, but nobody was using it, so we got Rust in addition: slower compile times, maybe arguably a more powerful language.
What does it mean? That’s why: all these languages, all these frameworks, all this vibe-coded stuff, no architecture. What is the way forward? Is it perhaps micro waterfall? You do the human steps at the top of the waterfall and let the AI do everything below. Iterate quickly, turn around in 1 day or 1 week, and have a lean waterfall hybrid approach maybe.”
“Nuclear fusion / fission is a metaphor for AI and technology change. When we split the atom, we were promised free energy. We got Hiroshima and Nagasaki. We got disasters on Long Island / Three Mile Island. We got Chernobyl. We got Fukushima, and we got hydrogen bombs. We didn’t get cold fusion, which we’re still waiting for—but we also have thousands of safe nuclear reactors that haven’t had an accident.
We have the U.S. Navy able to operate safely smaller reactors, except for the fact that the U.S. Navy goes around the world killing people. Could we do something similar for AI: local AI development—small, self-contained, safe, abundant energy—or in this case, software, logic, automation?
After this I think maybe you should draw up some bullet points from the discussion so far and also a complete set of notes with everything I’ve spoken.”
“Let’s move all of this from this conversation, so state notes everything. Move it into the www top-level repository that’s your sibling because that’s probably where it should go, under the title ‘We’re All Ragamuffins Now’ or just ‘Ragamuffin.’ It belongs there because it’s going to be a blog post, probably at some point.
Please clean up everything we’ve done in this talk in this Python repository and move everything, including a complete transcript, to the www stuff. When you’ve done that let me know and I can move the session over as well.”