https://www.techrepublic.com/article/ai-generated-code-outages/
Businesses using artificial intelligence to generate code are experiencing downtime and security issues, according to Sonar CEO.
cross-posted from: https://sh.itjust.works/post/25239919
"i don't know what happened, the truck was cruising just fine when we put the toddler on the wheel"
LLMs will save us from having to work on features now that we nearly ironed out all the issues introduced by Kubernetes.
Someone will have the "brilliant" idea to fix this by having chatbots review code in 5... 4... 3...
Welcome to my new startup where we train LLMs on compiled binaries. Now you can just prompt and get a complete executable, no coding knowledge needed. We value our company at $5b, product launch date indeterminate
I could swear I’ve seen a shartup with this pitch
will try check tomorrow, rn I’m enjoying the sounds of the first thunderstorm of the season
Thanks now you've sent me down the rabbit hole since I searched for this and clicked on the first ad: coderabbit.ai
One of the code reviews they feature on their homepage involves poor CodeRabbit misspelling a variable name, and then suggesting the exact opposite code of what would be correct for a "null check" (Suggesting if (object.field) return; when it should have suggested if (!object.field) return; or something like that).
You'd think AI companies would have wised up by this point and gone through all their pre-recorded demos with a fine comb so that marks users at least make it past the homepage, but I guess not.
Aside: It's not really accurate to describe if (object.field) as a null check in JS since other things like empty strings will fail the check, but maybe CodeRabbit is just an adorable baby JS reviewer!
Aside: the example was in a .jsx file. Does that stand for JavaScript XML? because oh lord that sounds cursed
Because rabbits are cute and fluffy and good and it is the solemn mission of all terrible tech companies to take the things you love and make you associate them with useless AI products.
JSX is JavaScript, but you can also just put HTML in it (with bonus syntax for embedding more JS expressions inside) and it can get transpiled into function calls, which means it'll result in an object structure representing the HTML you wrote. It's used so that you can write a component as a function that returns HTML with properties already computed in and any special properties, like event listeners, passed as function references contained in the structure.
You’d think AI companies would have wised up by this point and gone through all their pre-recorded demos with a fine comb so that
marksusers at least make it past the homepage, but I guess not.
The target group for their pitch probably isn't people who have a solid grasp of coding, I'd bet quite the opposite.
as I have said here some time ago, these chucklefucks are a goldmine waiting to happen. just not the kind of gold they think.
"When asked about buggy AI [code], a common refrain is ‘it is not my code,’ meaning they feel less accountable because they didn’t write it.”
Strong they cut all my deadlines in half and gave me an OpenAI API key, so fuck it energy.
He stressed that this is not from want of care on the developer’s part but rather a lack of interest in “copy-editing code” on top of quality control processes being unprepared for the speed of AI adoption.
You don't say.
have they tried writing better prompts? my lived experience says that because it works for me, it should work as long as you write good prompts. prompts prompts prompts. I am very smart. /s
I fixed the quote from the article "programmers are not known for being great at writing prompts because many of us find the whole idea offensive and stupid"
I'm reminded of the guy in a previous thread who claimed LLMs helped him as a rubber duck partner. You know - the troubleshooting technique named for its efficacy when working with a bath toy.
Incidentally, not even 5 years ago, you could just go on stack overflow and copy and paste a whole topic into your IDE and compile that bitch, and you'd probably wind up with fewer bugs than the normie executives who just learned about ChatGPT yesterday from their 8 year old nephew who uses it to write catfish lovenotes to his teacher from Gordie Howe or some other retired sports guy.
Reminds me of the story of the old engineer asked to come in and fix some machine in a factory.
The engineer inspects the machine, marks it with some chalk, then strikes the chalk mark with a hammer.
The machine works again.
The company asks for an itemised invoice after seeing the initial invoice for $10k.
To which they received:
- hitting chalk mark with hammer: $1.
- knowing where to place the chalk mark: $9,999
GPT suffers from garbage-in garbage-out just as much as a search engine does.
Knowing how to find search results to fix your specific situation is a skill.
Utilising GPT for such a task is equally a skill. With the added bonus of GPT randomly pulling the perfect API/Library out of its ass
on a slight tangent, I often think about this piece of writing. in general, but I've also started wondering what that picture's going to look like after the tsunami of LLMs suddenly finds it's actually made of air and not water
Yeah I feel like once people realize AI chatbots like ChatGPT are largely just search engines with AutoTldrBot built in, they'll be better at using them. ChatGPT is great for bouncing ideas off of or rubber-ducking through a solution. But just like with StackOverflow answers, you as the developer need to be able to recognize when ChatGPT is just spouting garbage, when it's getting you close to the answer, what adjustments you need to make to make its answers work for your situation, etc. In it's current state, it will never just magically hand you a fully developed, robust, well-integrated, complete solution though, as much as tech CEOs want it to.
Sounds like a great solution people will be prepared to pay OpenAI $100B in the future for, and not at all like an incremental upgrade over StackOverflow with extra ecocide added.
Yeah after rereading my comment it's not super clear, but I'm not trying to endorse ChatGPT/OpenAI. I agree that AI is a pretty terrible solution for the use case of "search engine with a built in AutoTldrBot", because of all the reasons you mention. I was just trying to point out that it's being marketed as a replacement for actual software developers, and that's very very very far from reality at the moment.
you as the developer need to be able to recognize when ChatGPT is just spouting garbag
easy: all the time
GPT and the whole AI bs we have at the moment excels at being convincing. It's even prepared to back up what it says.
The problem is, that all of that is generated. Not necessarily fact.
It will generate API methods, entire libraries, sources, legal cases, and science publications.
And it will be absolutely convincing as it presents and backs up those claims.
For example, GPT gives some API function of some library that magically solves your issue. Maybe you aren't hugely familiar with the library, but you don't trust GPT - so you research this made up API method and find the actual way to do it. Except you have GPT saying this exists and it works the way you want it to. So you research more, dig deeper.
Eventually you end up reading the source code, have a deeper understanding of the API in general and how to actually find useful answers (IE how to search query for it), and end up using the method you found while trying to find the mythical perfect API method.
I mean, I guess that's a win? You learned some documentation, you solved the problem... Who cares?
Maybe I'm just bitter because that was how I first tried any of the new AI things. And I wasted 2-3 hours instead of actually solving the fucking problem by consulting the facts.
For some reason when I read this I am reminded of our "highly efficient rail" which often derails