J
EP. 009
AI for Solo Founders
WEDNESDAY, SEPTEMBER 16, 2026
The people building AI spent the week asking everyone to slow down.
Last Wednesday a researcher quit Anthropic (the company behind Claude) and said the industry is gambling with our lives. Two current Anthropic leads agreed with him in public.
On Saturday Dario Amodei (the CEO of Anthropic) published an essay asking every lab to pace itself. Sam Altman and Elon Musk backed it within hours. Those three do not agree on anything.
Let's jump right in.
THIS WEEK IN AI
1
Why it matters: Jacob Coxon spent three years on pretraining at OpenAI, then Anthropic. He left last Wednesday and said both companies are racing toward self improving AI and gambling with our lives.
The part that made it news: two current Anthropic leads backed him publicly. Evan Hubinger, one of their safety leads, put the odds of AI killing everyone within the decade at more than 10%.
CBC · 3 MIN READ
2
Why it matters: The essay is called We Must Pace the Frontier. His concern: models are now improving models faster than the engineers can understand them.
Three steps. Outside evaluators sitting inside every frontier lab with employee level access. Shared limits on how fast capabilities are allowed to grow. Then some kind of agreement with China. Anthropic committed to the first one on the spot. The other two need everyone else.
Altman and Elon both backed it publicly. Those two agreeing on anything is the headline.
DARIO AMODEI · 12 MIN READ
THE OTHER SIDE
The most interesting take I saw on it
Roberto Nickson (@rpn) laid out both arguments in two and a half minutes.
I'm sharing his take because I genuinely think he summed it up (and more) perfectly.
The short version. Either something happened we do not know about, this is regulatory capture, the labs hit a wall, or it is PR. And even if the risk is real, you cannot just slow down, because China said it will not, and you need powerful AI to defend against powerful AI.
Watch it, then tell me what you think...
The warning.
The boys all working together (sort of?)
TOOL OF THE WEEK
ChatGPT Work
Free on the desktop app. More room on Plus, $20/mo
ChatGPT Work is game changing for how you use AI. Stop using "chat". I literally have not opened regular chat in weeks.
Work reads a folder on your machine before every job. Make a project, pick Local, drag in a folder, and everything it builds lives there.
Inside that folder sits a file called AGENTS.md. Say "Remember this rule" and whatever comes next lands in the file. New session, new window, it already knows.
Then plugins. Connect Gmail, Notion, your calendar, whatever you actually run on, and it does things instead of describing them.
I put the whole thing into a six module crash course, about 40 minutes end to end. It is in the Dojo.
The plugins panel. Every app you actually run on, wired in once.
PROMPT OF THE WEEK
The two words that make ChatGPT remember
You never open the rules file. You talk, it writes. Anything that starts with these two words gets saved to the project. Anything without them dies with the chat window.
Say these three today, one at a time, then open AGENTS.md and read what it wrote.
Remember this rule: never send anything as me without asking first.
Remember this rule: flag what you are unsure about instead of guessing.
Remember this: I book meetings in 45 minutes, never 60.
COPY · PASTE · REPLACE THE BRACKETS
FROM A DOJO CALL
Context is GOOD. Too much context is BAD.
A Dojo member gave Claude everything she had. It handed back generic.
She had done the work. A research library, competitor breakdowns, her own master's thesis, and 447 of her own posts, all loaded in. Every script still sounded like AI.
Here's the thing. That is too much. Not too little.
Her first post and her latest post do not sound like the same person. Feed it all 447 and it averages you into nobody.
What I told her: pick 15 to 30 pieces that sound exactly how you want to sound right now. Yours, or someone else's you admire. That is the whole training set. Build the framework off that.
Same rule for you. The brain works when it is curated, not when it is complete. This is how you properly train your AI.
道
The ChatGPT Work crash course is inside
Six modules, 40 minutes, one folder per project. Watch it once and you are running Work properly by the end of the day.
Plus weekly live calls and a room full of solo founders shipping instead of researching.
First 7 days are free.
Join the Dojo
100+ FOUNDERS INSIDE
One last thing.
Wherever you land on it, the tools got better this week and they will keep getting better whether you use them or not.
Go set up one folder. Teach it one rule. That is the whole assignment.
Keep pushing forward,
Justyn
Justyn
by Justyn. Shipped with 
by AI.