Anduril CEO's warning on the AI bub

Palmer Luckey’s skepticism about the "AI bubble" usually boils down to a simple point: the gap between a flashy demo and a reliable production system is wider than most VCs want to admit. As someone spending 10+ hours a day in Cursor and Claude Code, I feel this tension constantly. We are in this weird phase where AI can write a perfect React component in 3 seconds, but can still hallucinate a non-existent library method that breaks the entire build.

The "bubble" isn't about the tech being fake; it's about the overestimation of "zero-shot" capabilities. If you treat AI as a magic wand that replaces thinking, you're contributing to the bubble. If you treat it as a high-speed implementation engine that requires rigorous verification, you actually get a productivity boost.

My current workflow to avoid the "bubble trap" is moving away from generic chat and toward strict context steering. I've stopped asking Cursor to "fix this bug" and started using a .cursorrules file to force the AI to adhere to my project's specific architectural patterns. Without this, the AI tends to drift toward "generic internet code" which creates technical debt faster than it solves problems.

Here is a snippet of the constraints I put in my .cursorrules to keep the AI grounded:

- Always use TypeScript strict mode.
- Avoid using any third-party state management libraries unless explicitly requested; stick to Zustand.
- When modifying API calls, check @/lib/api-client.ts first to ensure consistent error handling.
- Do not suggest 'placeholder' comments like // implement logic here; write the actual logic or mark it as TODO: [Specific Requirement].

The real productivity gain comes when you stop fighting the AI's tendency to over-engineer. I've noticed that Claude 3.5 Sonnet loves to add "future-proof" abstractions that I don't need. To kill this, I use a "minimalist prompt" strategy. Instead of "Build me a robust user authentication system," I use:

Implement a basic JWT auth flow using NextAuth. Keep it to a single file if possible. No abstractions. Just the raw implementation.

One major gotcha I've encountered is the "context drift" during long sessions. After about 20-30 prompts in a single Composer session, the AI starts forgetting the initial constraints and begins hallucinating. I've found that the only way to maintain quality is to aggressively kill the chat session and start a new one once the specific feature is implemented.

The tangible wins I'm seeing:
Boilerplate elimination: I can spin up a fully typed CRUD interface in minutes, which used to take an hour of manual typing.
Rapid Prototyping: Using Claude Code via CLI to run tests and iterate on errors in a loop is significantly faster than jumping between a browser and an IDE.
Refactoring: Feeding a messy function into the AI and asking it to "simplify the cognitive load" often reveals logic flaws I missed.

The bubble pops for people who think the AI is the architect. For those of us using it as a super-powered compiler, the value is very real.