Comment by muzani | Hacker News Reader

Comment by muzani | Hacker News Reader

Specialist skills. AI is trained on data. Data pulls it towards the average. It does a good job, but only good. If great work is valuable, there's still a need for specialists.

Like right now, native mobile jobs are mostly unaffected by AI. Gemini, despite all the data in the community, doesn't do a decent job at it. If you ask it to build an app from scratch, the architecture will be off. It'll use an outdated tech stack from 2022. It will 'correct' perfectly good data to an older form, and if you ask it to hunt for bugs in cutting edge tech, it might rip out the new one and replace it with old stuff. It often confuses methods that are common between different languages like .contains().

But if very high quality data is easily accessible, e.g. writing, digital art, voice acting, etc, that makes it viable to be cloned by AI. There's little animation data and there's even less oil painting data - so something like oil painting will be greatly more resistant than digital art. It's top tier on Python and yet it struggles with Ren'Py.

Anthropic released experiment results from getting it to manage a vending machine: https://www.anthropic.com/research/project-vend-1

This is a fairly simple task for a human, and Claudius has plenty of reasoning and financial data. But it can't reason its way into running a vending machine because it doesn't have data on how to run vending machines.

The phrase “outdated tech stack from 2022” is kind of hilarious

Out of date after only 3 years?

As one using tech mostly laid down in the last millennium, I also smiled at this.

But I think what the poster meant was that the AI is not always "up to date". So if my C compiler is a 2025 version, and my code makes use of features added since 2022, then the AI can "retrogress" the code simply because it isn't aware of the new.

Or another example, imagine when JavaScript Promises were new. There's a lot more examples of "not using promises" than "using promises ". So the AI us likely to use the old pattern.

If you're doing "up to the minute " code because you put the effort in to keep abreast of new stuff, then this retrograde will seem, well, frustrating.

This is exactly it. One of the guys we interviewed failed because he used AI. The AI gave him a stack which made it difficult to complete the task in the given time.

Well, that was the story anyway. Some team members were already using AI for production code and pointed out that if he had knowledge of the good stack, he would have just asked AI to build the good stack instead of relying on it to build the whatever stack.

For context, the senior interviews we give don't force tool or stack. They're problems we have trouble with. The aim is to find someone smarter than us who brings in a new tool or stack which solves this problem better. If we wanted to test promises skill, we'd give a complex async problem, like getting them to clone the box in Facebook with emoji count and stuff.

Outdated libraries are a big problem. Sometimes stuff depreciated 10 years ago! It's the cargo cult problem.

A "generation" is roughly 5 years on front end in my experience. Measured by say, the time when people make fun of jQuery when they can use vanilla JS to the point where people make fun on jQuery as an 'outdated' stack.

Mobile is shorter, because it's highly centralized around Apple and Google who can declare something obsolete. It's not simply opinion; I literally got an email saying that I had to meet the target API level requirements by 31 Aug for an app that was at the latest one less than a year ago.

A 'generation' in AI is 1-2 years now - GPT-4o was state of the art in May 2024, now deprecated.

I can't tell if the models being up to date make other stacks faster or slower.

I feel sadly that this may lead to much information normally available for free to be guarded more tightly, as the opportunity cost for not putting a barrier between valuable data and access to it increases with time.

Reply on news.ycombinator.com