> join them

Become an LLM? Probably better to try and differentiate ourselves from LLMs than try to mimic them.

This is the part the AI advocates don't seem to get. There's nothing to learn with AI: each new model is better then the last. Requires less input to achieve a workable result.

The advocacy has always felt like cope to me and you see it in the advertising and LinkedIn: "get ready for AI", "adapt your AI workflows" - it's all centered on saying "you need the skills for the new thing so you don't get left behind".

But I don't need the skills for the new thing, because it does things for me. And each new successive generation will do more. Any time I would've spent bolting together some AI workflow a couple of years ago was wiped out when Claude came along. People are talking about there very clever multi-agent workflows or whatever, but it's all just prompts into the same datacenters and then...wiped out when the next model can just do it.

The advocacy is well...an excuse. The product looks and feels like AI. It's not impressive when it's generated by AI. The user isn't going to improve or build a better one, because they don't work on training new AI models. And a new AI model of sufficient power will just wipe out whatever skills you obtained, and the thing which might be useful - understanding the AI output - you'll never learn because you aren't doing it.

▲

zozbot234 4 hours ago | parent | next [-]

> There's nothing to learn with AI

You need to steer the AI effectively and assess its results, otherwise you just get nonsense. That takes real-world knowledge. In fact availability of AI makes knowledge skills more valuable, not less.

	▲	debazel 3 hours ago \| parent [-]
		The amount of steering necessary is rapidly decreasing. You're looking at a way too small timeline if you think this will be sustainable, or you're hoping that LLMs will hit their peak very soon.

▲

virgildotcodes 3 hours ago | parent | prev [-]

Eh, there’s definitely some value in understanding for yourself via experience which models are actually good for which use cases. The benchmarks are unreliable imo, and as I’ve interviewed developers who don’t really use AI, they say things like how they don’t think the (free versions of) copilot or ChatGPT (requests routed to their cheapest models) don’t seem very good. Totally out of touch with the capabilities of the leading models and harnesses.

I think the real argument is just staying employable. Companies are expecting faster and faster turnaround, and it’s simply becoming impossible to meet these deadlines with fully handwritten code. Even before outright mandates on AI usage. If you refuse to use AI, they’ll bring on someone who will, whether or not the quality drops, high quality code is not the primary goal of the business.

Dogshit, hideous vibe coded messes are launching daily and reaching 6-7+ figure ARRs while leaking customer data. Nobody cares in this environment.

If you’re a freelancer it’s even worse, the expectations are that producing a fully functional moderately complex app shouldn’t take a single person more than a couple months, and ideally one.

Expectation for a contractor coming into an enterprise codebase that’s been running for 11 years with a dozen+ internal devs and a mishmash of legacy and new tech -> they want you to implement a totally new feature which touches half a dozen systems in the app ready to demo in 6 weeks and launch to the public in 8.