[joduplessis]

I really wish seemingly intelligent people would stop using the abstraction analogy (like the article does). The key word is: determinism. Every level of abstraction (inc. power tools, C, etc.) added a deterministic layer you can rely on to more effectively do whatever it is that you're doing - same result, every time. LLM's use natural language to describe programming and the result is varied at the very best (hence agents, so we can brute force the result instead). I think the real moat is becoming the person who can actually still program.

[phpnode]

People always say this but it’s misguided imo. Yes LLMs are not deterministic, but that’s totally irrelevant. You aren’t executing the LLMs output directly, you’re using the LLM to produce an artefact once that is then executed deterministically. A spec gets turned into code once. Editing the spec can cause the code to be updated but it’s not recreating the whole program each time, so why does determinism matter?

[michaelrpeskin]

In my experience, I'm using LLMs as my abstraction to "junior engineer". A junior engineer isn't deterministic either. I find that if you treat the LLM output like a person's output, you're good. Or at least in my projects, it's been very successful. I don't have it generate more code than I can review, or if I give it a snippet to help me fix it, if it ends up re-writing it like an ambitious engineer would do, I tell it to start over and make minimal changes.

I guess I'm not spun up about the determinism because I've been working at the "treat it like a person" level more than the "treat it like a compiler" level.

To me, it's really like an engineer who knows the docs and had a good memory rather than infallable code generator.

I work at a small company, so we don't have tons of processes in place, but I imagine that if you already had huge "standards" docs that engineers need to follow, then giving the LLM those standards would make things even better.

[skydhash]

The thing is you can quickly teach a Junior how to respect a specification contract, so that with very minimal oversight, you get the wanted implementation. And after a few years (or months), the communication overhead get shorter. What would have been multiple rounds of meetings and review sessions are a short email and one or two demos.

[QuercusMax]

What I've been learning as a 20% "harness engineer" is that in order to get the models to "learn" you need to add both documentation and static checks, as well as often custom skills. My main project at work has issues where the AI will often get super confused and step on itself trying to run tests - so the answer is writing better docs (AGENTS.md) and providing deterministic tools to work with the projects.

Large software projects (I'm thinking google3) often have large amounts of both of those things, as they're always getting new developers joining.

[AstroBen]

If it's not deterministic you can never fully trust it. In a deterministic abstraction I don't need to audit the lower levels.

[ex-aws-dude]

Who said you need to trust it? Reviewing code is still way faster than writing code.

[bluefirebrand]

> Reviewing code is still way faster than writing code.

Writing code results in a much better understanding of the code than reviewing it

In fact I would say that in large complex codebases, in order to develop the same understanding of what the code is doing might actually take longer than writing it from scratch would have

[esafak]

But it's written to your spec; there should be no surprises!

[bluefirebrand]

That's the fun part! The surprise is that it's actually not written to your spec at all! It just kinda smells similar to your spec

[HDThoreaun]

You fully trust your coworkers?

[nozzlegear]

If you don't, you may want to find a different company to work for.

[mrbananagrabber]

this is the way LLMs _should_ be used, as an assistant to create reliable, deterministic code. and honestly, they're fantastic when used this way. build the thing you need with the LLM, then put the LLM away.

but in practice, the current obsession with agents means people are creating applications that depend entirely on sending requests to LLMs for their core functionality. which means abandoning the whole idea of deterministic software in favor of just praying that all of the prompts you put around those API requests will lead to the right result.

[udave]

try distributing this spec amongst your team members, ask each of them to drive it to completion. no follow up edits. deploy to individual environments and then run a rigorous test suite against all of the deployments. see if all of them behave the same way.

[phpnode]

They won't. So what? This is not how specs are used, no one is saying that they are a replacement for source code.

[ex-aws-dude]

Exactly, the argument makes sense if its about inference at runtime

But that's not the case here

[knivets]

how do you know the artifact is correct?

[NiloCK]

I grant that there's a definition of abstraction that LLMs don't fall into. But people describing LLMs as another abstraction layer aren't all misunderstanding this. Instead, they are using the term ... more abstractly.

EG: How did Mark Zuckerberg make software five years ago?

He's as capable of opening up an editor as I am, but circumstance had offered him a different interface in terms of human resources. Instead of the editor, he interacts with those humans, who produced the software. This layer between him and the built systems is an abstraction, deterministic or not.

Today, you and I have a broader delegation mandate over many tasks than we did a few years ago.

[indiosmo]

The way I frame this is that LLMs are not replacing the tools, whic are are deterministic. They are replacing the humans, which are themselves non-deterministic, as in your Zuckerberg example.

[qnleigh]

LLM's don't have to achieve perfect reliability to replace lots of work. They just have to reach the balance of reliability and cost suitable for a given task. This will depend on the task.

[ansk]

I see what you're getting at, but determinism isn't the right word either. LLMs are fundamentally deterministic -- they are pure functions which output text as a function of the input text and the network parameters[1]. Depending on your views on free will, it could be effectively argued that humans are deterministic as well.

The concept you're touching on is the idea that LLMs (and humans) are functions which are inscrutable. Their behavior cannot be distilled into a series of logical steps that you can fit in your head, there are no invariants which neatly decompose their complexity into a few interpretable states, and the input and output spaces are unstructured, ambiguous, underspecified, and essentially infinite. This makes them just about impossible to reason about or compose using the same strategies and analysis we apply to traditional programs.

[1] Optionally, they can take in a source of entropy to add nondeterminism, but this is not essential. If LLM providers all fixed their prng seeds to a static value, hardly anyone would notice. I can't imagine there are many workflows which feed an LLM the exact same prompt multiple times and rely on the output having some statistical distribution. In fact, even if you wanted this you may just end up getting a cached response.

[udave]

Let's be real, if you and I both ask claude to generate a feature on the same project, what are the chances that it spits out 100% replicated code? But if we are to build the project using a Dockerfile, we will get the same binary and the same image. Products around LLMs are non deterministic unlike compilers.

[ansk]

I can assure you that a fully deterministic and equally effective claude is possible to build. And yes, that would mean identical prompts would yield 100% identical output 100% of the time. It would still make the occasional logical or factual error, but it would do so deterministically. Would this solve any of the problems with building reliable programs using LLMs?

[pzo]

it's nondeterministic because we chosen it by having higher 'temperature' in settings. I bet if you run open weights model with temperature 0 and on the same device the same prompt and turn off parallelism you will have more deterministic result (excluding some floating point operations).

[skydhash]

> Optionally, they can take in a source of entropy to add nondeterminism, but this is not essential. If LLM providers all fixed their prng seeds to a static value, hardly anyone would notice

Everyone added /dev/random to their offerings, so every LLM tools for coding are non deterministic.

[arecsu]

There's something to be said about the fact that the very people who would use deterministic layers to build stuff are... non-deterministic. We, as humans, have our set of pros and cons, wins and failures. Even the most brilliant coders on earth will make mistakes from time to time. I often fail to see this getting accounted in any conversation when there is a critique towards LLMs, as if we humans are not flawed in our own ways, with a huge degree of variance across individuals. Good and bad code existed prior to LLMs. If you're hiring someone to do code, you're basically using some heuristics to trust this person will do a good job. But nothing is ever guaranteed 100% deterministically ever. Without thinking it that much, LLMs will sometimes produce better code and manage systems that some people who are earning salaries out there. Possibly sub-par developers if we were precise, but professionals in the meaning of the word (that are being paid to do work).

At the end of the day, what matters is how willing the person behind a given task is when it comes to deliver quality work, how transparent and honest they are, to understand requirements, and a pleasure to work with along other humans. AI/LLMs are just extra tools for them. As crazy as it might sound, but not so many people are willing to push boundaries and deliver great work. That is what makes the difference.

[careers_terria]

Are other layers truly deterministic though? Do I know for sure whether that object has been garbage collected or not? How many cycles this instruction will take to run?

[danaw]

every time a person uses the abstraction argument, an angel dies