@IngoHeinscher

The difference is that compilers follow deterministic processes that can be understood by humans if the need arises (and the need _does_ arise when you deal with ancient legacy code).

LLM systems, by their very nature, are _stochastic_ processes. They might give the right answer and the right code, but you cannot rule out that they give the wrong answer and buggy code, and there is nothing you can do to prevent that. And then a human has to figure out what went wrong, and if they do not understand how the code was generated, they are already off to a bad start.

@juergen_hubert That is simply not what is happening. The LLMs are faster at writing AND debugging than any human could ever be. I suggest you try it with something like Opencode and an LLM of your choice, or with Claude Code for a simple start.

@IngoHeinscher

So how _did_ they solve the stochastic problem of LLMs?

@juergen_hubert They don't bother, they just do the same tests you'd do with any human-written code. And if there's an issue, ask the LLM to fix it. Might take a few attempts, of course. Just like with any human programmer.

@IngoHeinscher

A human _software developer_ (not a "programmer") can think about the process, and analyze what went wrong.

An LLM, by definition, cannot "think". There won't be any lessons learned, and no institutional knowledge.

@juergen_hubert The bubble might not burst, though, because it's not a bubble. Coding with coding agents based on LLMs is just so much faster that there will be no going back.

Ingo Heinscher That is not what the developers I’ve worked with are saying. Lines of code can be produced faster than ever, but software engineering isn’t about the number of lines of code produced, but about what those lines do.

@juergen_hubert The developer isn't really replaced by the LLMs yet. The programmer absolutely is.

@IngoHeinscher

And how many IT companies still employ "programmers" as opposed to software developers?

@juergen_hubert I am not sure I understand what you are saying. Do you believe coding agents are a productivity leap, or do you not?

@IngoHeinscher

They might be a productivity leap for small, fairly standardized, self-contained projects, but their usefulness decreased geometrically with project complexity.

@juergen_hubert Now you are grasping for straws. This technology is phenomenal and very useful. Now, not every use is maybe a wise allocation of resources. The talk of a "bubble" might be about some of the companies, but the AI technology itself is clearly extremely useful even in its present, rather early stage.

@juergen_hubert Now you are grasping for straws. This technology is phenomenal and very useful. Now, not every use is maybe a wise allocation of resources. The talk of a "bubble" might be about some of the companies, but the AI technology itself is clearly extremely useful even in its present, rather early stage.

@IngoHeinscher

Are there any scientific papers that quantify how useful they are in actual software development practice?

@juergen_hubert Claude Code is just one year old, I doubt anyone decided to write papers about it. But you hear from everywhere how they are no longer coding themselves, but using LLM's for coding.

So do I.

@IngoHeinscher

So no one has yet any experience whether it is viable for large, complex, and long-term software projects, either.

@juergen_hubert I suggest, again, you try it yourself to see what people mean when they say coding by humans doesn't happen any more. Anthropic's Opus 4.6, which just came out this week, will humble anyone.

@IngoHeinscher

I am not a software developer working on complex software products which have been developed over the course of decades. When _they_ tell me that this is genuinely useful after at least one year of usage, _then_ I will listen.

"Extraordinary claims require extraordinary evidence", and all that.

@juergen_hubert You choose to stay blind.

@IngoHeinscher

Or I choose not to get blinded by the hype. We shall see.