It can translate one language into another but we already had transpilers for that.
It can find bugs quickly but we already had static analysis. It also generates many plausible looking false positives.
LLMs can write them for me, and I can happily look through them to make sure what the LLM made was neither brittle, nor testing unimportant things while missing the important stuff.
They can also help with code review, but only to a limited degree; a second pair of eyes to avoid lazy people going "LGTM", or the other way around time-wasting feedback from bike-shedders, but as with unit tests, right now you still need to be the kind of person who is happy to double-check its work — if you're the kind who was previously a little too quick to accept pull requests with a "LGTM", you may well be too willing to listen to irrelevant bike-shedding from the LLM or not realise it is missing the point of the code or the ticket.
But if you use it to replace the stuff you're already good at, and not help with the stuff you're weak at, it'll probably slow you down while also making the code worse.
Millions of programmers out there, and you're saying there's not even a 1% chance that a single one of them is moving 10x faster?
Would that mean that every other claim to be moving 10x faster is an exaggeration? It evidently is, as these claims of massively increased productivity are purely anecdotal.
If the claims were ‘5% productive across the board’, it would be huge, but 1000%? That is a ludicrous tall story that requires extraordinary proof.
It's pretty obvious that's not what's happening. To pretend otherwise would suggest that every software developer everywhere is completely unaware of what tech debt is AND that they don't care at all.
Again: it's obviously wrong if we think about it for longer than 5 seconds. This is just pure emotional cope.
> It doesn't design architecture.
It can do that. Won't say it's amazing, but it can do it.
> It doesn't think. It doesn't reason.
Can you define these in ways that aren't tautologically limited to humans? I have yet to encounter anyone who has managed this while making such an objection to any form of AI.
> All it does is search for and copy-paste code from elsewhere on the internet.
They provably don't work like that. Both because local models exist, can be run offline, still spit out code to solve problems; and because some world records have been set by them, for some CS-related maths problems the best known method was invented by an AI which used an LLM as a component.
https://the-decoder.com/openais-ai-beats-every-human-at-atco...
https://deepmind.google/blog/alphaevolve-a-gemini-powered-co...
No, because that’s the point. It’s not intelligent because intelligence is a human characteristic.
> Both because local models exist, can be run offline
It’s not that they search interactively, it’s that the whole internet has been sucked into the training set.
A rose by any other name, etc.
Defining the meaning of the word you use to suit your preferences and then gluing to that other people's uses of the same word is not a well-reasoned argument, it's the human equivalent of a failure mode we saw in early AI, where a label saying "taxi" can fool a classifier into thinking a clock is a taxi: https://bliss-e-v.github.io/SCAM-project-page/
Will you choosing a tautologically-limiting definition that equals human, prevent AI from performing well enough to replace you at work? No, it will not: when the AI is good enough to do the work, it will do the work regardless of what you call it.
This has already happened in various other fields that have been automated, even though (a) the consensus is those AI are "not intelligent" and (b) workers before that automation insisted their work required intelligence and could not be automated.
(Plus the same example in the opposite direction: while creationists are busy denying that evolution can work and claiming there has to be an intelligent designer, simulated evolution is a standard approach in machine learning).
> It’s not that they search interactively, it’s that the whole internet has been sucked into the training set.
Training.
As in, it learned. First by reading the internet, then by feedback as it tried to make stuff and was rated by the quality of output.
It would be coherent to argue instead that intelligence is the number of examples one needs before one has learned a thing; by this standard AI is as thick as pond scum (rate: evolution), and even the fact that silicon is faster than synapses by the ratio to which joggers are faster than continental drift only makes up for this where the examples exist.
Unfortunately, this topic is software, and the examples exist.
For new code, I treat it like a very knowledgeable mid level engineer who has a lot of patience to search the code base. We make changes a step at a time and I review them carefully and have it make changes I want. It's still a lot faster than doing it myself.