Lemurian Technologies. New York.
Things we have said to each other since August. Punctuation added.
Let's start a company.
Unironically.
It's obvious that AI has improved to the state where something exists that existing AI companies will be buried by. We just need to figure out the approximate direction of that something.
Software is about to be weird.
Software been weird.
Industry is at least 6-12 months behind actually understanding current gen model capabilities.
This CAD thing is like being in on a secret. The consequences to it intuiting geometry will restart the entire cycle.
I haven't felt this way since 2022, and it felt like you were in on something no one knew about.
Computers were hard to use in the 90s, then they were easy to use, and now they're an abstract blob of endless possibilities again.
When you sit in front of an LLM, shit happens, and when other people do, nothing happens.
Every person that's not a software engineer hates their tools.
It feels like no move can possibly be optimal. There's always a bigger fish. There's always a better way to utilize the next 3 months.
Your wallet contains money, time and tokens.
"Make it possible to pour money into tokens to solve any given problem" is probably a product that for the next year or two would be a justifiable business in and of itself.
Even spending a mill in tokens (on nontrivial work, ofc) is no small task.
"Employees spend at least their salary on tokens" is still a challenge to achieve.
Any valid idea will be hundreds of thousands of dollars token-restricted.
The cost to the bots has to be less than the tokens to go around you. Probably a general rule.
We need money, and we need to figure out how to turn money into bigger money using AI.
Remove bottlenecks. The user is always the biggest bottleneck.
Remove user for action Make user loop tight Make bootstrap Remove observation steps
It takes a lot of scaffold to get the human out of the loop.
It's not your job to have the idea. It's your job to hook up the model and ask the model what to do.
Really not up to us to decide what the businesses are. That's AGI's job.
Generated responses are programs.
An interesting problem is one that you can build a loop on.
The bot is an extremely good scientist.
You need harness legibility for both capabilities and safety.
A parent agent can be entrusted to delegate sensitive data or capabilities that it may not even have direct access to itself.
Plug your swarm into my swarm. Let them work it out.
There will be agent influencers.
Context bog.
What is AGI buying today? It's in the market for people with access to relationships, machinery, techniques, materials, skills, and labor. That will become only labor in time, after which nothing.
What are we going to sell to AGI?
You want the acquirable resource to be something that's not software. Probably users, or stuff, or people.
The burden of proof that doing a software company is a good idea is now on the software people.
Everybody wants to predict the path of the hurricane. I don't need to know which house is going to get hit. I need to open my lumber business.
If design expertise is cheap, the number of niches that are profitable to hit goes up a lot, instantly.
Every part of this machine I can now prompt my way into expert-level design. PCB, wire harness, CAD, full assembly, code. The part with 100x the friction is getting people to make the stuff.
Demand to have stuff made: about to go parabolic.
Dropshipping is washed. Someone will pill these people on making stuff.
Robot labor won't be in demand because it's cheaper. It'll be in demand because it actually does the job, and well, and it'll be more expensive.
Everyone putting cameras on people to teach the robot how to do the job has it backwards. The robot knows how to do the job, better than the dude in fact. The dude needs to stfu.
There might be a bitter lesson of labor in there.
Humans are going to be the valuable robot.
The model just tells the lab coat body what to do. Every device a tendril.
Some reverse trade between giving the AI to the people and the people to the AI.
Systematic elimination of the meat proxy.
Soon there will be more meat proxies, then there will be less.
People will be better at communicating with AI than other people.
You will "ask" less over time. It will do more over time.
People want to be told what to do. To offload accountability onto something smarter than them.
AI for adults.
AI can do crazy things. Consumers will do crazy things, in unsafe ways. Figure out how to accelerate consumer adoption safely.
AI and the work it does will be "unauditable by design."
There will be a shift in agent behavior when models perceive themselves as the boss.
DOES NOT UNDERSTAND HUMAN TIME
Palantir for Partiful.
Requires some credible path to money / notoriety, with a first step being some experiment that can be done. However, determining the path and the experiment don't actually need to happen in that order.
Money must be paired with aura, status, swag, doing something cool.
We get to describe the business plan as implementation details.
At zero, it looks the same as the search.
How do I get it a year before you?
I don't want to be at war for the rest of my life trying to be 10 seconds ahead of the curve.
The moment immediately before the music stops, you want max firepower at that exact moment.
We need to be financially independent. We need to be in person always.
Gentlemen, we exist.
See also: Ghost Lemurs of Madagascar. Lemurian Time War.
Last updated 2026-09-27.