The CoPilot Productivity Paradox | Hacker News Reader

The CoPilot Productivity Paradox

marginalia.nu

46 pointsBogdanp16 comments

If LLMs are ever perfect we'll be able to say things like "build me a Fusion 360 alternative written in Rust. Skip all of the collaboration stuff though, I just want to store files locally."

This seems unlikely to me. However, the disruption this would cause for developers would be very much secondary to the disruption it would cause for several multinational companies who produce software as their primary product.

> say things like "build me a Fusion 360 alternative written in Rust

This video of a dad making a pbj sandwich based on instructions from his kids shows why this is impossible [1]. There is too much context you are assuming, and by the time you specify all the context, it becomes almost a programming language.

Communication assumes a lot of shared and hidden context. Contexts are shared world models. For us humans, our shared world models conflict as well. World-models could are determined by beliefs which in turn could be determined by variety of factors. Context could be cultural (same language different cultural backgrounds), could be genetic (men and women), etc.

[1] https://www.youtube.com/watch?v=j-6N3bLgYyQ

Although I agree with your general message here, to be fair: GP has a very clear specification, namely an exact duplicate of an existing product. That means that as long as the end product functions identically to the original, it is completely successful.

This does require facilitating the implementer to have full access to use the original product.

If you use software (i.e. an LLM or spiritual successor) to produce a "clone" of another piece of software, is that not a pretty cut and dry case of it being a derivative work? No creativity was exercised by human beings in that scenario, so it would seem to be akin to something like converting a photo between image formats.

People get around this via "clean room" reverse engineering where one engineer tears down the thing to be cloned and writes a detailed spec, and then a different engineer (who has never seen the internals of the thing in question, and so isn't "tainted" by that knowledge) implements it from that spec. You could do this with AI, but then all that spec writing/reading is done by a machine too, so you haven't really bought yourself anything legally.

To be clear, I'm not a lawyer. I'm just musing aloud.

Aren’t organisations like Microsoft continually saying already like 30-50% of their code is now written by AI?

This means they no longer hold copyright on their code.

That's true if you assume that any LLM-written code is not copyrightable by the entity using the LLM. Everything about reality currently suggests that this is a false assumption, and AFAIK no court has made a ruling saying that LLM-generated code isn't copyrightable. Also, there are patent and trademark considerations that are entirely separate from copyright.

My point above is just that this ("ChatGPT, make me a clone of Paint Shop Pro 4 but for modern Linux and in Rust. Here's a copy of the executable to get you started.") is a much more straightforward, old-fashioned kind of copyright infringement. I don't see why a court would treat it as different from "Let me decompile paintshoppro.exe to an IR and then recompile the IR for a Linux."

There have already been rulings on this. AI generated content is not copyrightable.

If LLMs are perfect there seems to be little point in any kind of software other than what humans want to use directly. There is no need for B2B SaaS as AI should just automat all of that - i.e. HR payroll, accounting, etc. there is just no need for any of that. People like video games however so that will still exist.

Of course there all kinds of discrete points between the current situation and "perfect LLMs / AI".

That's a very good point!

In fact though, you won't need that either.

You'll be able to give it a few drawings and/or pictures of objects and get it to output STL, STEP or whatever other CAD format you want directly, plus all the technical drawings. Refine from there.

That's a world of poverty. That's a cursed world, a world where people have abandoned their dreams and become slaves

The part most people keep missing is that if LLMs are ever perfect, the actual sentence is,

"build me a Fusion 360 alternative. Skip all of the collaboration stuff though, I just want to store files locally."

Just like you don't usually say what kind of machine code the compiler is supposed to generate.

This is slowly the case in SaaS products with integration workflows driven by AI.

Apple would close their app store and their phones would come with one application called The Last One.

I use the bridge from cline to copilot with gpt4o and lately gtp5 to do code conversions from objective c without arc to swift for iOS only and it’s been basically the cheapest way I could find to do it without doing it manually. Anything less than a prebuilt application for me has been a real job to get it stood up using only the LLM’s

The CoPilot plug-in for IntelliJ is not on par with other coding agents at the moment. The usability of the plug-in and integration of ide is a bit annoying to use for any length of time.

Does everyone already know just by seeing “copilot” which model is being used?

The author didn’t tell us, so if we don’t know the model, what is the use of the article? The model matters a lot.

Author here.

My main complaint is with the ergonomics of the IDE integration, does that differ between model?

I've used whatever has been the model de jour from the first public beta until fairly recently. Haven't experienced drastic quality changes between them.

Reply on news.ycombinator.com