I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.
I built Ditherlab to turn photos into dithered images and looping GIFs.
Play with the colors and fades, and export what you make.
It’s free, with no signup. Your images stay on your device.
https://t.co/Xl6pPfgWoq
Give yourself a year to find out what you're capable of.
Pick one thing you keep putting off. Close the door for two hours and leave your phone outside. Show someone your work each week so you can improve it.
Keep going through the part where you're bad at it.
I'd use some of this reset to fix the agent's dev setup.
If you're copying errors or sending screenshots by hand, ask Astra to set up access to logs and a way to test the running app.
Have it document the steps for the next session.
My 15-prompt pack covers code audits, performance work and verification:
https://t.co/YPVWszLAK2
I write more about building with agents in my newsletter:
https://t.co/JNjDJJDDZE
Try the instruction audit on one repo and share what you find. 10/10
GPT-6 Astra follows instructions more closely. That makes your old skills and AGENTS.md worth another read.
OpenAI recommends auditing them.
I'd check the rules that make the agent stop, ask for approval, or keep testing after the work is done. 🧵
I'd pay for max where repeated trials show fewer failed fixes or less human repair. I'd use high where it passes the same checks at lower total cost.
Count failed attempts in the cost per accepted result, and record review time alongside it. 9/10