Code is just a precise way of specification (see meme)
- related comic strip
- interesting related articles:
Specification-driven development
Specially-trained LLMs are the new compilers
Cross-compilation = compiling to different programming language targets
Different levels/kinds of specification
- plain English for throwaway applications (Python of specification languages)
- general purpose specifications
- domain-specific specifications
- tests (property, unit, integration)
- types
- post-/pre-req specs
- dependently typing
Different levels of specification give different evaluation guarantees
Add two and two- no real guarantees2 + 2- guarantees that addition will be mathematically accurate- Guarantees can be made for memory, security, resources, etc.
Specifications get us even further from the “how” and closer to the “what”.
There will still be a need for engineers that delve deeper into the actual code, just like there are still times that we need to delve into the generated assembly.
- computer engineers
- security analysts
- penetration testers
- researchers
- verification for actual specifications
Grow in two ways:
- specification languages (higher level)
- stricter typing (more statically-checked specifications) Rust, dependent typing, Dafny
Higher level = we care less about the details of the implementation
LLMs also need to get better, this is a prediction of the future, not the present
Non-deterministic? Yes, but we gave up non-determinism with memory management years ago when we started using garbage collection. All we care about is that the memory is collected, not how it is collected. For the times when we care about how memory is managed, we drop down to a lower-level programming language.
Basically, other languages may not go away, but they will be less frequently used.
This is also an approach for producing, not for learning. Because we are not actively involved in the code, we don’t learn about the edge-cases or the processes that the AI is using. We will need to make a decision about what we want to learn.
Also, maybe experiment by writing a prototype by hand, then get rid of it and write specifications that an AI can then implement.