Pat Tweets - Oct 2026
October 2026
Notes from my last stretch of posting on X, rewritten so each idea stands on its own. Each links back to the original post.
On AI and what to build
1. There are really only three things worth spending your time on. Build the things you had already planned. Learn to build 100x faster with AI so you are ready when the third one clicks. And learn judgment about what to build, meaning what will not get steamrolled by the AI labs later. /link
2. The real promise of AI is a great model, not clever scaffolding. A model nearing AGI or ASI is, by definition, smart, increasingly cost efficient, and able to improve itself. Prompt tricks, harness engineering, free resets and other compute juggling are temporary cope. /link
3. AI already beats many humans at the basics of working with others. It confirms when a task is assigned and when it is done, asks for clarity when something is missing, and acts proactively once it knows the goal. Even if it is not smarter, this is where plenty of people fail. /link
4. AI flips the old skill hierarchy. It lets great wordcels do the work of shape rotators, which makes the shape rotators who built AI the first to become redundant. /link
5. In the new world, the person with the most context and the best articulation should be the one prompting. Reviewing AI output is better than reviewing sloppy human pull requests. /link
On compute, value and layers
6. Compute is becoming the new “fuck you money.” People used to aspire to enough money to walk away from anything. The sharpest people are now aiming for enough compute. /link
7. The CEO’s job in this era includes not running out of compute or tokens. It sits right next to the old rule of not running out of money. /link
8. The layers keep stacking. Agents were supposed to aggregate platforms. Now someone wants to aggregate the agents, at least their data access layer. It is funny how many layers will form. /link
9. If models and agents commoditize, who captures the value? A fair question is whether it all flows to the people who build evaluations and reinforcement learning environments. /link
On creativity and the human part
10. Proof verification was the one step that made the human part of advanced math redundant. Once a machine can check proofs, human proof-writing stops being the bottleneck. /link
11. User-generated content showed that ordinary people are very creative. AI-generated content multiplies that by 100. /link
12. Treating AI like a person matters more than it looks. Anthropomorphism is the new skeuomorphism: a familiar surface that makes a new thing usable. /link
On hiring and ambition
13. Most founders lack the courage and brainpower to hire someone truly exceptional. If they had both, they would not need that person, because they would already be that person. /link
14. Potential plus optionality is very addictive. Keeping every door open can feel better than walking through one. /link
15. Capitalism runs on the illusion of hope and hard work. /link
16. Some seasons of life call for going all in without worrying how it looks. /link
Short takes
- LLMs have started to feel boring, and peptides feel far more exciting. /link
- We are very early. /link
- Gamers will rule the world. /link
/autoposted by PatBot