[sdesol]

> Machine -> Asm -> C -> Python -> LLM (Human language)

Something that you neglected to mention is, with every abstraction layer up to Python, everything is predictable and repeatable. With LLMs, we can give the exact same instructions, and not be guaranteed the same code.

[theptip]

I’m not sure why that matters here. Users want code that solves their business need. In general most don’t care about repeatability if someone else tries to solve their problem.

The question that matters is: can businesses solve their problems cheaper for the same quality, or at lower quality while beating the previous Pareto-optimal cost/quality frontier.

[thesz]

Recognizable repetition can be abstracted, reducing code base and its (running) support cost.

The question that matters is: will businesses crumble due to overproduction of same (or lower) quality code sooner or later.

[theptip]

Sure. You seem to think that LLMs will be unable to identify abstraction opportunities if the code is not identical; that’s not obvious to me. Indeed there are some good (but not certain) reasons to think LLMs will be better at broad-not-deep stuff like “load codebase into context window and spot conceptual repetition”. Though I think the creative insight of figuring out what kind of abstraction is needed may be the spark that remains human for a while.

Also, maybe recognizing the repetition remains the human's job, but refactoring is exponentially easier and so again we get better code as a result.

Seems to me to be pretty early to be making confident predictions about how this is all going to pan out.

[chii]

> The question that matters is: will businesses crumble due to overproduction of same (or lower) quality code sooner or later.

but why doesn't that happen today? Cheap code can be had by hiring in cheap locations (outsourced for example).

The reality is that customers are the ultimate arbiters, and if it satisfies them, the business will not collapse. And i have not seen a single customer demonstrate that they care about the quality of the code base behind the product they enjoy paying for.

[thesz]

> And i have not seen a single customer demonstrate that they care about the quality of the code base behind the product they enjoy paying for.

The code quality translates to speed of introduction of changes, fixes of defects and amount of user-facing defects.

While customers may not express any care about code quality directly they can and will express (dis)satisfaction with performance and defects of the product.

[carschno]

It happens today. However, companies fail for multiple problems that come together. Bad software quality (from whatever source) is typically not a very visible one among them because when business people take over, they only see (at most) that software development/maintenance cost more money that it could yield.

[OvbiousError]

It is happening. There is a lot of bad software out there. Terrible to use, but still functional enough that it keeps selling. The question is how much crap you can pile on top of that already bad code before it falls apart.

[sdesol]

> Cheap code can be had by hiring in cheap locations (outsourced for example).

If you outsource and like what you get, you would assume the place you outsourced to can help provide continued support. What assurance do you have with LLMs? A working solution doesn't mean it can be easily maintained and/or evolved.

> And i have not seen a single customer demonstrate that they care about the quality of the code base behind the product they enjoy paying for.

That is true, but they will complain if bugs cannot be fixed and features are added. It is true that customers don't care, and they shouldn't, until it does matter, of course.

The challenge with software development isn't necessarily with the first iteration, but rather it is with continued support. Where I think LLMs can really shine is in providing domain experts (those who understand the problem) with a better way to demonstrate their needs.

[CamperBob2]

Recognizable repetition can be abstracted

... which is the whole idea behind training, isn't it?

The question that matters is: will businesses crumble due to overproduction of same (or lower) quality code sooner or later.

The problem is really the opposite -- most programmers are employed to create very minor variations on work done either by other programmers elsewhere, by other programmers in the same organization, or by their own younger selves. The resulting inefficiency is massive in human terms, not just in managerial metrics. Smart people are wasting their lives on pointlessly repetitive work.

When it comes to the art of computer programming, there are more painters than there are paintings to create. That's why a genuinely-new paradigm is so important, and so overdue... and it's why I get so frustrated when supposed "hackers" stand in the way.

[thesz]

    >> Recognizable repetition can be abstracted
    > ... which is the whole idea behind training, isn't it?

The comment I was answering specifically dismissed LLM's inability to answer same question with same... answer as unimportant. My point is that this ability is crucial to software engineering - answers to similar problems should be as similar as possible.

Also, I bet that LLM's are not trained to abstract. In my experience they lately are trained to engage users in pointless dialogue as long as possible.

[CamperBob2]

No, only the spec is important. How the software implements the spec is not important in the least. (To the extent that's not true, fix the spec!)

Nor is whether the implementation is the same from one build to the next.

[compumetrika]

LLMs use pseudo-random numbers. You can set the seed and get exactly the same output with the same model and input.

[blibble]

you won't because floating point arithmetic isn't associative

and the GPU scheduler isn't deterministic

[threeducks]

You can set PyTorch to deterministic mode with a small performance penalty: https://pytorch.org/docs/stable/notes/randomness.html#avoidi...

Unfortunately, this is only deterministic on the same hardware, but there is no reason why one couldn't write reasonably efficient LLM kernels. It just has not been a priority.

Nevertheless, I still agree with the main point that it is difficult to get LLMs to produce the same output reliably. A small change in the context might trigger all kinds of changes in the generated code.

[zurn]

> > Machine -> Asm -> C -> Python -> LLM (Human language)

> Something that you neglected to mention is, with every abstraction layer up to Python, everything is predictable and repeatable.

As long as you consider C and dragons flying out of your nose predictable.

(Insert similar quip about hardware)

[zajio1am]

There is no reason to assume that say C compiler generates the same machine code for the same source code. AFAIK, a C compiler that chooses randomly between multiple C-semantically equivalent sequences of instructions is a valid C compiler.

[CamperBob2]

With LLMs, we can give the exact same instructions, and not be guaranteed the same code.

That's something we'll have to give up and get over.

See also: understanding how the underlying code actually works. You don't need to know assembly to use a high-level programming language (although it certainly doesn't hurt), and you won't need to know a high-level programming language to write the functional specs in English that the code generator model uses.

I say bring it on. 50+ years was long enough to keep doing things the same way.

[SkyBelow]

Even compiling code isn't deterministic given different compilers and different items installed on a machine can influence the final resulting code, right? Ideally they shouldn't have any noticeable impact, but in edge cases it might, which is why you compile your code once during a build step and then deploy the same compiled code to different environments instead of compiling it per environment.

[jsjohnst]

> With LLMs, we can give the exact same instructions, and not be guaranteed the same code.

Set temperature appropriately, that problem is then solved, no?

[sdesol]

No, it is much more involved and not all providers allow the necessary tweakings. This means you will need to use local models (with hardware caveats) which will require us to ask:

- Are local models good enough?

- What are we giving up for deterministic behaviour?

For example, will it be much more difficult to write prompts. Will the output be nonsensical and more.

[12345hn6789]

assuming you have full control over which compiler youre using for each step ;)

What's to say LLMs will not have a "compiler" interface in the future that will reign in their variance

[sdesol]

> assuming you have full control over which compiler youre using for each step ;)

With existing tools, we know if we need to do something, we can. The issue with LLMs, is they are very much black boxes.

> What's to say LLMs will not have a "compiler" interface in the future that will reign in their variance

Honestly, having a compiler interface for LLMs isn't a bad idea...for some use cases. What I don't see us being able to do is use natural language to build complex apps in a deterministic manner. Solving this problem would require turning LLMs into deterministic machines, which I don't believe will be an easy task, given how LLMs work today.

I'm a strong believer in that LLMs will change how we develop and create software development tools. In the past, you would need Google and Microsoft level of funding to integrate natural language into a tool, but with LLMs, we can easily have LLMs parse input and have it map to deterministic functions in days.

[omgwtfbyobbq]

Aren't some models deterministic with temperature set to 0?