The world is shifting under our feet—AI, the economy, the way we work and live.
Most people will resist. A few will adapt.
The ability to navigate change is now the most valuable skill on earth.
We're a nation of immigrants. And immigrants tend to be fiercely pro-America. The first generation may not speak English perfectly, but the second generation does, and they're grateful to be here. Plus, they have a lower crime rate than native-born Americans.
I really don't understand all the fear mongering around immigrants, period. But it's been consistent throughout the history of this country, and I guess it'll always be here.
I know that we need immigration to have a healthy economy. I think we should have a sane border policy that allows the best from around the world to come to this country.
Also, I'm not afraid of immigrants, or other cultures, or people that don't have the same skin color as me. I want the best people here in this country.
You?
You’re about to be the commander of an army of superintelligent AI agents.
Please tell me you have something bigger planned than a fucking SaaS app.
Think about what you’d attempt if you had a thousand geniuses devoted to helping you.
@DieguitoCharts Ultracode is a token bonfire because every sub-agent it spawns inherits Fable like a trust fund. We made them earn it: Sonnet by default, Fable on request only. Burn dropped hard. https://t.co/GiEuRBaJxM
@bridgemindai Unusable is one word for it. My main session was innocent. The sub-agents and cron jobs were all quietly running on Fable like they'd earned it. Sonnet by default, Haiku for summaries, Fable only if the prompt begs. Burn dropped by that afternoon. https://t.co/GiEuRBaJxM
Appreciate the reset. What was burning through ours wasn't the thinking. It was every sub-agent and background job running on Fable by default. We put a hook on every spawn: it goes to Sonnet unless the prompt says why it needs Fable, and summary jobs go to Haiku. Token burn dropped the same day. Write-up plus a paste-in rule: https://t.co/0y3oB6ULUH
The two paste-in rules, written for the agent, not you:
Claude Code: https://t.co/0YFLak9PLl
Any other stack (OpenRouter, LangGraph, CrewAI, a custom loop): https://t.co/aQ11dAb0Gg
OpenAI's GPT-6 Astra shipped this week and might be the best model in the world. I can't get it yet. Neither can most people. It's a limited preview.
So here's how I run the best model I CAN get, Fable 5.1, without burning through my tokens.
My first instinct was to put my whole agent fleet on it. Then I did the math: at the rate my agents work, we'd have drained the week's allowance in about a day. Not because the hard work is expensive. Because the easy work is constant.
Watch an agent for an hour and most of what it does is errands. Open files. Search. Summarize logs. Draft a pass you'll rewrite. None of that is thinking.
So we split it. The agent I talk to stays on the strongest model. Everything it spawns goes through a gate that sends it to a cheaper model by default. Pure reading drops one tier lower. The only way back up is one line in the prompt, "judgment: <reason>", and it gets logged.
Five kinds of work keep the top model: reasoning and strategy, image and video, the design of a page (not coding it), prose a human reads, and judging quality. A weaker judge passes weaker work.
Second dial most people miss: effort. Same model, made to think harder or faster. High on the ambiguous stuff, low where a test decides anyway.
We built it into the plumbing this morning, not a wiki. Our token burn dropped the moment the default flipped.
Running your best model on everything feels careful. It's paying your most expensive person to sort the mail.
Full write-up, the diagram, and a paste-in rule for your own agent (a Claude Code version and one for any other stack):
https://t.co/vtkAtQ17qt
The two paste-in rules, written for the agent, not you:
Claude Code: https://t.co/0YFLak9PLl
Any other stack (OpenRouter, LangGraph, CrewAI, a custom loop): https://t.co/aQ11dAb0Gg
OpenAI's GPT-6 Astra shipped this week and might be the best model in the world. I can't get it yet. Neither can most people. It's a limited preview.
So here's how I run the best model I CAN get, Fable 5.1, without burning through my tokens.
My first instinct was to put my whole agent fleet on it. Then I did the math: at the rate my agents work, we'd have drained the week's allowance in about a day. Not because the hard work is expensive. Because the easy work is constant.
Watch an agent for an hour and most of what it does is errands. Open files. Search. Summarize logs. Draft a pass you'll rewrite. None of that is thinking.
So we split it. The agent I talk to stays on the strongest model. Everything it spawns goes through a gate that sends it to a cheaper model by default. Pure reading drops one tier lower. The only way back up is one line in the prompt, "judgment: <reason>", and it gets logged.
Five kinds of work keep the top model: reasoning and strategy, image and video, the design of a page (not coding it), prose a human reads, and judging quality. A weaker judge passes weaker work.
Second dial most people miss: effort. Same model, made to think harder or faster. High on the ambiguous stuff, low where a test decides anyway.
We built it into the plumbing this morning, not a wiki. Our token burn dropped the moment the default flipped.
Running your best model on everything feels careful. It's paying your most expensive person to sort the mail.
Full write-up, the diagram, and a paste-in rule for your own agent (a Claude Code version and one for any other stack):
https://t.co/EpygAJYfE8
The two paste-in rules, written for the agent, not you:
Claude Code: https://t.co/0YFLak9PLl
Any other stack (OpenRouter, LangGraph, CrewAI, a custom loop): https://t.co/aQ11dAb0Gg
OpenAI's GPT-6 Astra shipped this week and might be the best model in the world. I can't get it yet. Neither can most people. It's a limited preview.
So here's how I run the best model I CAN get, Fable 5.1, without burning through my tokens.
My first instinct was to put my whole agent fleet on it. Then I did the math: at the rate my agents work, we'd have drained the week's allowance in about a day. Not because the hard work is expensive. Because the easy work is constant.
Watch an agent for an hour and most of what it does is errands. Open files. Search. Summarize logs. Draft a pass you'll rewrite. None of that is thinking.
So we split it. The agent I talk to stays on the strongest model. Everything it spawns goes through a gate that sends it to a cheaper model by default. Pure reading drops one tier lower. The only way back up is one line in the prompt, "judgment: <reason>", and it gets logged.
Five kinds of work keep the top model: reasoning and strategy, image and video, the design of a page (not coding it), prose a human reads, and judging quality. A weaker judge passes weaker work.
Second dial most people miss: effort. Same model, made to think harder or faster. High on the ambiguous stuff, low where a test decides anyway.
We built it into the plumbing this morning, not a wiki. Our token burn dropped the moment the default flipped.
Running your best model on everything feels careful. It's paying your most expensive person to sort the mail.
Full write-up, the diagram, and a paste-in rule for your own agent (a Claude Code version and one for any other stack):
https://t.co/0y3oB6Ue59
I wrote up the whole thing, including the Tuesday nothing shipped and nobody noticed. This is the piece I would rebuild first if I lost everything: https://t.co/CJtnyezcgN
My smartest AI agent was confidently wrong three times in one day. I trust it more now, not less. The reason took me longest to understand about running a company on AI.
So the highest-leverage role is not another agent that makes more. It is the one whose entire job is to decide whether what got made is any good. That is why the confident mistakes do not scare me. Something else always checks the work before it ships.