these LLMs are extremely good at fooling people that have to make business decisions, whether or not they are tech people is almost irrelevant. from the military to small start ups people have fallen for the siren of a text generator that can somehow do all these awesome things without actually being designed for it, and it just talks so convincingly on topics they already know about (because these topics have been discussed to death before and it can regurgitate it very well).
as a coder i cannot emphasize enough how bad AI code is for new tasks and how much it isn’t production worthy for everything, even the well known tasks. the agentic stuff is just pure nightmare. at best it can be used as a second partner in pair programming, but falling for just hitting tab and taking its suggestions will fuck you up.
Right, the real problem is that people making decisions are largely disconnected from the work and only have a superficial understanding of what the business actually does. The whole problem was already there long before LLMs where upper management tends to run on pure bullshit, and rarely takes any feedback from people actually doing the work. A common thing you see in software development is where the management gets together and decides on features and timelines without involving any of the technical staff, and then just presents these as a decree.
And I don’t find agentic stuff to be a nightmare myself, but its definitely not a tool that can write code unsupervised. You have to give it clear direction on what to do, set up tests, review the code. In general, I find it’s incredibly useful for figuring out a lot of the boilerplate, and figuring out things like APIs. But I completely agree that if you just let it run wild, it will consistently produce complete garbage. Hence why it doesn’t actually make you work all that much faster. The human is still the bottleneck, and simply crapping out more code faster isn’t useful. This is, once again, the part that management doesn’t understand. They think typing out the code is what the bulk of software development is.
You have to give it clear direction on what to do, set up tests, review the code. In general, I find it’s incredibly useful for figuring out a lot of the boilerplate, and figuring out things like APIs.
This is something I still like to do by hand. Especially when I’m using a new API. That stage of reading and understanding the API docs and finding undocumented issues (there’s always tons) is important. I’ll usually end up wrapping the API I’m interacting with in my own that behaves the way I expect for what I’m doing then use that.
I tried the LLM stuff a while ago and it just really rapidly devolved into my having no idea what was actually happening.
My use case is that I often need to work with services, and before I’d actually have to spin up the service, and try making curl calls to it to see how it actually works. Now, I can just ask the LLM to produce a curl or js call example, and it saves me a ton of time. It’ll show exactly which params I need to pass, and the structure of the request and the response. Obviously, I still have to test all this end to end, but I can do that in the dev environment where all the services are running, instead of having to orchestrate everything on my machine while developing.
When I do get LLM to implement stuff, my approach is to just break things up into focused features, and then tell the agent to make a branch per feature, that makes it pretty easy to review the pull requests, because I know what the features is intended to do, and there isn’t a mountain of code to go through.
That’s fair, having a bunch of individual atomic features that you get around to eventually is better than full slop merging by far.
You have to stay in the loop or you’ll get lost
Exactly, and that’s why the human is the bottleneck in the whole process that can’t be removed. Writing code isn’t what takes the bulk of time when doing software development. It’s figuring out what the code is already doing, what needs to be written, and how that’s going to interact with the existing features. We get paid to think, and the LLM can’t replace that.
I am fully convinced that AI coding is why so many Microsoft and NVIDIA products have had major bugs lately.
Relevant Ed Zitron https://www.wheresyoured.at/the-revenge-of-the-business-idiot/
yeah similar take, there really is a perfect match between people who are largely disconnected from material reality and tools that can create a plausible simulacrum of how these people expect things to work.
I mean, before LLMs, they had consultants. The gippities are just very good at reproducing the contentless, cloudy bubble talk 24/7. And they are even more sycophantic and reckless because they don’t have to worry about insurance.
for sure, it’s not a qualitative change, the problem was always there
