Too many things happened this weekend guys. Here is the rundown:

What you can't miss in AI this week: 1/ @OpenAI drops $38B on @amazon AWS compute, breaking up with @Microsoft 's exclusivity while spending toward a $1.4T infrastructure plan 2/ Extropic's thermodynamic chips claim 10,000x energy savings, could this be the solution to AI's power

If you want to know all that is happening due / in / because of AI. This is a famtastic read. 👌

Covfefe

1 Trillion IPO? How much will openAI valuation be worth after 24h? 👀

Here's what actually matters: Cursor's own team was using Composer for their daily work BEFORE release. They built a model they'd actually reach for. That's the difference between "AI features" and rebuilding everything around intelligence. Cursor 2.0 is live.

Then the next level: They embedded a browser directly in the editor. Why? Because agents need to SEE their work. Now agents can: 1- Test changes live 2- Select page elements 3- Use full dev tools 4- Iterate until correct No more blind coding.

Cursor 2.0 can run multiple models on the same problem simultaneously and pick the best solution. Your agents literally compete for the best code. And on harder tasks? This approach significantly improves results.

But here's where it gets wild... Cursor rebuilt the ENTIRE interface around agents instead of files. You can run multiple AI agents in parallel without them interfering with each other. They use Git worktrees to keep agents separated. Finally.

The speed difference is absurd. Composer is 4x faster than similarly intelligent models at coding tasks. Why? Because they trained it via RL to make EFFICIENT choices, fewer tool calls, better parallelism, smarter workflows. It doesn't just code fast. It thinks fast.

During training, Composer had access to the same tools you use: - Codebase-wide semantic search - File editing and creation - Terminal commands - Linter integration It learned to fix its own linter errors, write unit tests, and maximize parallelism. Nobody else is doing this.

They didn't train it on code datasets like everyone else. They trained it with reinforcement learning on ACTUAL software engineering challenges in massive codebases. The model learned by doing the job. Using real production tools. Solving real problems.

Meet Composer. The first coding model built BY an IDE company FOR their own tool. It completes most tasks in under 30 seconds and beats every AI except GPT-5 and Sonnet 4.5. But here's the insane part...

🔥 @cursor_ai just made every other IDE obsolete. - Meet Cursor 2.0 - They trained their own frontier AI model, Rebuilt the interface around multiple competing agents, Embedded a browser directly in the editor! This is what vertical integration looks like 👇

The long tail of software is going to be absolutely massive. And the "professionals" building it will look nothing like developers. They'll be creators. Personalities. Entrepreneurs with $200/month in tools and good ideas. The YouTube moment for software is here.

YouTube revealed that we were desperately short on video content in 2006. LLMs are about to reveal we were desperately short on software (!!) The only reason we didn't know? Only 20 million people could build it. That constraint just evaporated.

Replit just hit $150M ARR this year. Demand is exploding for coding agents. The signals are everywhere: - Kids don't just want to be YouTubers anymore. They want to build apps. - The entrepreneurial urge is finding a new outlet.

But there's a second pattern emerging too: Hyper-personalized applications for tiny audiences. Software freed from the constraint of needing billions in revenue. Just good ideas. Small communities. Real value for specific people.

The most successful apps will be built by individuals with: 1- Existing distribution 2- Natural community gathering points 3- Creator instincts, not engineering backgrounds Think "unboxing videos, but software."

Remember: virtually ZERO incumbents from traditional TV became YouTube stars. Late-night shows treated YouTube as a dumping ground for "bonus content." They completely missed what the medium was actually for. Traditional software builders are making the same mistake right now.