[Toutouxc]

> If LLMs reach a point of sufficiently high competence to the point that they're capable of producing usable software from vague, and sometimes contradictory instructions - the same software engineers have to regularly deal with, then at that point software will simply be an expectation of ability in any job.

I don’t understand how people can say this and then continue talking about software. So we’re saying machines can now casually do complex and cognitively demanding jobs like software development (or 90% of all white-collar jobs out there) and we’re NOT worried about the lynching mob going door to door and hanging IT people on lampposts? And I’m being serious, the impact this would have on societies would be unprecedented.

[exitb]

The only truly unprecedented thing is the white-collar part. Otherwise it happened many times over that entire regions where suddenly out of job and job prospects. Automation and offshoring have equivalent impacts on the affected job market.

[sdevonoes]

If LLMs can handle 99% of software jobs, then they can “crack” robotics software and then we would have LLMs on robots to replace the rest of the workers

[hard_times]

If you automate the creation of software, you essentially can automate almost any white-collar job.

[somenameforme]

I don't think this is necessarily true. Software that was mostly tools will be automated because it generally has a "simple" and specific purpose, but consider software that is the final product itself - like a game. LLMs are nowhere even remotely near the level of being able to independently automate the creation of a non-trivial game, though of course they can be a tremendous tool in the creation of such.

And I think that example largely carries over to many other big picture and/or creative tasks. I think it would likely require another revolutionary leap for LLMs to get to that point, and that leap is probably not coming any time soon, if ever, at least not with LLMs.

Another practical issue is that LLMs, as difficult as it can be to imagine at times, are still just glorified autocomplete machines. And so genuinely novel advances will not be coming from LLMs. For instance take LLMs back to the early 20th century and none of them will discover relativity, even with infinite power/time. It's just so unlike their training material, and in direct contradiction to much of it, that it's simply outside their reach, let alone grasp.