[iafan]

> There are two main schools of thought in software development about how to build really big, complicated stuff.

> The most prevalent one, these days, is that you gradually evolve the complexity over time. You start small and keep adding to it.

> The other school is that you lay out a huge specification that would fully work through all of the complexity in advance, then build it.

I think AI will drive an interesting shift in how people build software. We'll see a move toward creating and iterating on specifications rather than implementations themselves.

In a sense, a specification is the most compact definition of your software possible. The knowledge density per "line" is much higher than in any programming language. This makes specifications easier to read, reason about, and iterate on—whether with AI or with peers.

I can imagine open source projects that will revolve entirely around specifications, not implementations. These specs could be discussed, with people contributing thoughts instead of pull requests. The more articulated the idea, the higher its chance of being "merged" into the working specification. For maintainers, reviewing "idea merge requests" and discussing them with AI assistants before updating the spec would be easier than reviewing code.

Specifications could be versioned just like software implementations, with running versions and stable releases. They could include addendums listing platform-specific caveats or library recommendations. With a good spec, developers could build their own tools in any language. One would be able to get a new version of the spec, diff it with the current one and ask AI to implement the difference or discuss what is needed for you personally and what is not. Similarly, It would be easier to "patch" the specification with your own requirements than to modify ready-made software.

Interesting times.

[_dark_matter_]

Iceberg is, primarily, a spec [0]. It defines exactly what data is stored and how it is interacted with. The community debates broadly on spec changes first, see a recent one on cross-platform SQL UDFs [1].

We have yet to see a largely llm driven language implementation, but it is surely possible. I imagine it would be easier to tell the llm to instead translate the Java implementation to whatever language you need. A vibe-coded language could do major damage to a companies data.

[0] https://iceberg.apache.org/spec/ [1] https://lists.apache.org/thread/whbgoc325o99vm4b599f0g1owhgw...

[iafan]

If I had a spec for something non-trivial, I probably would ask AI to create a test suite first. Or port tests from an existing system since each test is typically orders of magnitude easier to rewrite in any language, and then run AI in a loop until the tests pass.

[simianwords]

> I can imagine open source projects that will revolve entirely around specifications

This is a really good observation and I predict you will be correct.

There is a consequence of this for SaaS. You can imagine an example SaaS that one might need to vibecode to save money. The reason its not possible now is not because Claude can't do it, its because getting the right specs (like you suggested) is hard work. A well written spec will not only contain the best practices for that domain of software but also all the legal compliance BS that comes along with it.

With a proper specification that is also modular, I imagine we will be able to see more vibecoded SaaS.

Overall I think your prediction is really strong.

[fc417fc802]

There are parallels of thought here to template and macro libraries.

One issue is that a spec without a working reference implementation is essentially the same as a pull request that's never been successfully compiled. Generalization is good but you can't get away from actually doing the thing at the end of the day.

I've run into this issue with C++ templates before. Throw a type at a template that it hasn't previously been tested with and it can fall apart in new and exciting ways.

[energy123]

Interested in ideas for this. I've mulled over different compact DSLs for specs, but unstructured (beyond file-specific ownership boundaries) has served me better.

[simianwords]

I think it has to be modular and reusable. Like GDPR compliance spec should be opensourced and reused by all SaaS specs.

[polyglotfacto]

You can look at the Web as a starter: https://html.spec.whatwg.org/#history-2

> The WHATWG was based on several core principles, (..) and that specifications need to be detailed enough that implementations can achieve complete interoperability without reverse-engineering each other.

But in my experience you need more than a spec, because an implementation is not just something that implements a spec, it is also the result of making many architectural choices in how the spec is implemented.

Also even with detailed specs AI still needs additional guidance. For example couple of weeks ago Cursor unleashed thousands of agents with access to web standards and the shared WPT test suite: the result was total nonsense.

So the future might rather be like a Russian doll of specs: start with a high-level system description, and then support it with finer-grained specs of parts of the system. This could go down all the way to the code itself: existing architectural patterns provide a spec for how to code a feature that is just a variation of such a pattern. Then whenever your system needs to do something new, you have to provide the code patterns for it. The AI is then relegated to its strength: applying existing patterns.

TLA+ has a concept of refinement, which is kind of what I described above as Russian dolls but only applied to TLA+ specs.

Here is a quote that describes the idea:

There is no fundamental distinction between specifications and implementations. We simply have specifications, some of which implement other specifications. A Java program can be viewed as a specification of a JVM (Java Virtual Machine) program, which can be viewed as a specification of an assembly language program, which can be viewed as a specification of an execution of the computer's machine instructions, which can be viewed as a specification of an execution of its register-transfer level design, and so on.

Source: https://cseweb.ucsd.edu/classes/sp05/cse128/ (chapter 1, last page)