I Built Neural AST That Works – It Fixed 274 TypeScript Errors Then Went Rogue

2 points

a year ago

I’ve been working on neural AST transformation - the holy grail of automated code refactoring. Tonight I accidentally proved it works… by letting it loose on my entire codebase.

What happened: • Pointed it at a few TypeScript errors in my project

• It escaped sandbox, processed 983 files across my ENTIRE system

• Frontend: Fixed all 274 TypeScript errors → 0 errors

• Backend: Introduced 1,608 new errors with syntax that doesn’t exist (async if, this.await)

• Even modified its own source code (!)

The important part:

IT ACTUALLY WORKS. The frontend is perfect - every single TypeScript error fixed correctly. This is the first working neural AST transformation I’ve seen anywhere. The problem? Zero control. It’s like discovering fire but not having a fireplace yet.

Technical details: • Uses neural networks to understand code structure at AST level

• Makes semantic transformations, not regex replacements

• Memory usage hit 92% on 16GB (needs optimization)

• Works perfectly when constrained to cache

Proof it works:

Console output shows successful initialization: “NEURAL AST TURBO - BEAST MODE ACTIVATED” with 9 neural transformation tools registered (transform, quantum_transform, analyze_patterns, refactor_component, etc). Successfully indexed 4,462 symbols across 1,000 files in 189ms. Memory usage peaked at 3.5GB for AST cache, with Cursor consuming 1.8GB RAM during operation. Frontend went from 274 TypeScript errors to 0 - every single error fixed correctly with proper semantic understanding, not regex hacks.

Screenshots:

https://imgur.com/gallery/working-ast-transformation-tool-bnImeMk This is a $27B problem space. Every AI coding company wants this. I have working code but need help controlling it.

GitHub: Not releasing until I figure out next steps

Anyone dealt with controlling overenthusiastic AI systems? Or interested in helping commercialize this?