I Am Often Wrong
In January I made a prediction about gene editing. Five months later it broke. I got the right street but the wrong address.
Blog posts by Colin Steele
In January I made a prediction about gene editing. Five months later it broke. I got the right street but the wrong address.
The market thinks Meta is in the AGI race. Meta is in the cat video race. There is a difference, and PE of 22 is low.
What it actually looks like to cross from symbol-world into atoms-world, brick a board in the process, and spend the day at a Latin Mass written in 1975.
I have a stupidity detector. It fires whether I want it to or not. Two market shapes set it off (larval and fossilized) and the tax shows up whether or not anyone credits the work.
I built a product that solves a real problem. The internet doesn't care. Here are my four options, all of them bad.
You don't generalize from cups. You specialize from physics. Cognition is navigation of a single manifold where intent is the tangent vector, risk is the curvature, and the metric is the scar tissue.
Dassault sells you the right to read your own design files. NAR sold you the right to see the housing market. Same grift, different industry. One got sued. The other won't.
The enterprise UI was a 30-year workaround for not being able to talk to databases. The workaround is over. A builder's guide to the $5-50M vertical opportunities the giants will never touch.
The industry forgot how to build software that runs on the user's machine. Some of us didn't. Here's what happens next.
Why CRAWL! combat works. One action, two trays, and nowhere to hide.
99.99% of agentic frameworks are evolutionary dead ends. The Permian extinction is coming.
There's a layer in your brain between reality and symbols, that both language and logic rest on. Transformers lack this layer.
Lonely fun is lost. What remains is pale in comparison.
A collection of deliberate bear poking.
You can't file a flight plan for a million delivery drones.
Die, usage die! The usage die can kiss my d20.
A Stanford survey cataloging ~300 papers on LLM reasoning failures just confirmed every architectural diagnosis I made.
Two months ago I proposed layered cognitive-kill drone defense built around cheap autonomous interceptors. Yesterday, Raytheon demoed exactly that. They also revealed the gap I'm most worried about.
I wrote that whole departments would vanish. Then I proved it twice in a month. AS9100-grade QMS and a complete Team Operating System, built part-time with LLMs. Nobody got fired. The jobs just never existed.
Technology has always created more jobs than it destroys. That has been correct for two hundred years. It is about to stop being correct.
The heist shouldn't feel like a spelling bee
The failure to reason is a symptom. The failure to relate is the disease. Frontier models do not relate.
Are you paying attention? Granna has some riddles for you.
LLMs are trained on the compression of human experience, not experience itself. The territory they'd need to become intelligent is fundamentally alien to ours.
The same optimization that turned feeds into dopamine casinos is now operating on AI assistants.
'Massively polygenic' was our security blanket against designer humans. AlphaGenome just pulled it off. The only levee left is ethics. Don't laugh.
Meetings replaced memos the same way the feed replaced thinking. Same thermodynamic disease, different host.
$100 helmet-mounted edge AI that buzzes 'INCOMING DRONE' before the human ear even registers the sound.
Sequoia just declared AGI. Their definition is 'the ability to figure things out.' My Roomba is AGI now. Someone call the LPs.
RLMs and MoR are real progress. They're also still not reasoning. A critique of the cutting edge.
Train your satellite AI for Tuesday in LEO with bad data. A 'messy reality' simulation environment for autonomous space traffic management.
The accelerationists expect the Rapture. The doomers expect Ragnarok. Both are fairy tales. What actually arrives is worse: narrow superintelligences that solve everything except the thing that matters.
Algorithmic feeds are bad for your brain. This startup is vitamins for your brain.
A manifesto for human/machine polity charting civispace. Neither as apex mind. A forest.
Your ASI needs a hug.
I proposed that goal-space geometry could replace executive function in AI architectures. Then I discovered the math already exists.
Your attention-algorithm feed isn't corrupting you. It's completing a transaction your biology was always ready to make. And a population optimized for minimum cognitive expenditure is maximally controllable.
A TTRPG about self-aware puppets who know they're being controlled, and the only freedom left: laughing at it loudly enough that it hurts.
A robot caring for your grandma needs to fluidly inhabit radically different forms of cognition: creativity, logic, and crisis. Nothing in current AI architectures explains how. Worms might.
Transformers don't reason. They perform reasoning kabuki theater. What's missing isn't capability, it's coordination.
As automation commoditizes execution and cognition, the most durable form of human labor is not leadership or knowledge work, but illegible, risk-bearing, political coordination. An essay on why corporate politics survives every technological wave.
If $500 drones can destroy the power grid (or an aircraft carrier), million-dollar missiles aren't a defense strategy. They're bankruptcy. A look at the 'Vermin' problem and why we need an immune system for critical infrastructure.
Why I'm intentionally building Taleway wrong (no accounts, no UI, no engagement metrics) and why that might be exactly right for 2025
Prognistication and contemplation of the nature of sustainable competition, circa 2025
I am the captain of my soul
5E: the mall food court of roleplaying games
Player skill? Character skill? Chocolate AND Peanut Butter?
Building an ambient audio app revealed Apple's blueprint for the future, and it's not about innovation. It's about control. Here's what they're building next.
Your website doesn't matter anymore. Your API does. AI agents are killing web traffic the same way Amazon killed Sears catalogs. Here's what happens next.
Thoughts on static site generators