This is the best summary I could come up with:
In-depth Several big businesses have published source code that incorporates a software package previously hallucinated by generative AI.
Not only that but someone, having spotted this reoccurring hallucination, had turned that made-up dependency into a real one, which was subsequently downloaded and installed thousands of times by developers as a result of the AI’s bad advice, we’ve learned.
He created huggingface-cli in December after seeing it repeatedly hallucinated by generative AI; by February this year, Alibaba was referring to it in GraphTranslator’s README instructions rather than the real Hugging Face CLI tool.
Last year, through security firm Vulcan Cyber, Lanyado published research detailing how one might pose a coding question to an AI model like ChatGPT and receive an answer that recommends the use of a software library, package, or framework that doesn’t exist.
The willingness of AI models to confidently cite non-existent court cases is now well known and has caused no small amount of embarrassment among attorneys unaware of this tendency.
As Lanyado noted previously, a miscreant might use an AI-invented name for a malicious package uploaded to some repository in the hope others might download the malware.
The original article contains 1,143 words, the summary contains 190 words. Saved 83%. I’m a bot and I’m open source!
*bad Devs
Always look on the official repository. Not just to see if it exists, but also to make sure it isn’t a fake/malicious one
*bad Devs
Or devs who don’t give a shit. Most places have a lot of people who don’t give a shit because the company does not give a shit about them either.
What’s the diff between a bad dev and a dev that doesn’t care? Either way, whether ist lack of skill or care, a bad dev is a bad dev at the end of the day.
The difference is whether the fault for the leak of your personal data rests with the worker who was incompetent, or the employer who didn’t pay for proper secure software.
I can be good at a trade, but if I’m working for a shit company with shit pay and shit treatment, they’re not going to get my best work.
You get out what you put in, that’s something employers don’t realise.
You’d be surprised how well someone who wants to can camouflage their package to look legit.
Yeah, I’m confused on what the intent of the comment was. Apart from a code review, I don’t understand how someone would be able to tell that a package is fake. Unless they are grabbing it from a. Place with reviews/comments to warn them off.
the first most obvious sign is multiple indentical packages, appearing to be the same thing, with weird stats and figures.
And possibly weird sizes. Usually people don’t try hard on package managing software, unless it’s an OS for some reason.
True. You can’t always be 100% sure. But a quick check for download counts/version count can help. And while searching for it in the repo, you can see other similarly named packages and prevent getting hit by a typo squatter.
Despite, it’s not just for security. What if the package you’re installing has a big banner in the readme that says “Deprecated and full of security issues”? It’s not a bad package per say, but still something you need to know
The official repositories often have no useful oversight either. At least once a year, you’ll hear about a malicious package in npm or PyPI getting widespread enough to cause real havoc. Typosquatting runs rampant, and formerly reputable packages end up in the hands of scammers when their original devs try to find someone to hand them over to.
I just want an LLM with a reasonable context window so we can actually write real working packages with it.
The demos look great, but it’s always just around 100 lines of code, which is beginner level. The only use case right now is fake packages.
Just use the AI Horde. iirc our standard is like 4K context and some people host up to 8K. Here’s a frontend
I use it for writing functions a lot, tell it the inputs and desired outputs it’ll normally make what i want. Recently gpt has got good at continuing where it left off too.
I have tried the copilot integration in edge out of curiosity, and if you feed the ai the context of the page the response can be useful. There is a catch, tho:
-
when opening a document the accepted formats are html, txt, pdf. The documentation of a software package can be summarized but thr source will be the context of the page and not a web search, which is good in this casr
-
when generating new information, the model can be far too sintethic, cutting out potentially useful informations.
I still think you need to read the documentation yourself, maybe using the AI integration only when you need a general idea of the document.
What I do is first reading the summary of the documebt by bullet point, than reading the pdf file as a whole. By the time I do so, the LLM has given enough of a structure to facilitate my readings…
I’m not particularly interested. Some on my team are playing with it, but I honestly don’t see much point since they spend more time fixing the generated code than they would writing it.
And I don’t think it’ll ever really work well (in the near-ish future) for the most common type of dev work: fixing bugs and making small changes to existing code.
It would be awesome if there was some kind of super linter instead. I spend far more time reading code than writing it, so if it can catch bugs, that would be interesting.
In my experience with Codeium, it sometimes works ok for three or four lines of code at once. I’ve actually had a few surprises where it nailed what I was going for where I didn’t expect it. But most of the time, it’s just duplicating code from elsewhere in the same file, which usually doesn’t make sense.
It’s also pretty good for stuff where I’d usually build some exotic regex to search/replace (or do it manually, because it’d take longer to come up with the expression), like transforming an enum into a switch construct for its members, or mapping said enum to a string of the member’s name.
This is very far from taking over my job, though. I’d love to be more of a conductor than the guy playing all instruments in the orchestra at once.
To each their own of course. It just seems like the productivity gains are perceptual, not actual.
For an enum to a switch, I just copy the enum values and run a regex on those copied lines. Both would take me <30s, so it’s a wash. That specific one would be trivial with most IDEs as well, just type “switch (variable) {” and it could autocomplete an exhaustive switch, all without LLMs.
Then again, I’m pretty old school. I still use vim as my editor (with language server plugins), and I’m really comfortable with those kinds of common tasks. I’m only going to bother learning to use the LLM if it’s really going to help (e.g. automate writing good unit tests).
Can we fucking stop anthropomorphising software?
“Hallucinate” is the standard term used to explain the GenAI models coming up with untrue statements
They don’t come up with any statements, they generate data extrapolating other data.
in terms of communication utility, it’s also a very accurate term.
when WE hallucinate, it’s because our internal predictive models are flying off the rails filling in the blanks based on assumptions rather than referencing concrete sensory information and generating results that conflict with reality.
when AIs hallucinate, it’s due to its predictive model generating results that do not align with reality because it instead flew off the rails presuming what was calculated to be likely to exist rather than referencing positively certain information.
it’s the same song, but played on a different instrument.
when WE hallucinate, it’s because our internal predictive models are flying off the rails filling in the blanks based on assumptions rather than referencing concrete sensory information and generating results that conflict with reality.
Is it really? You make it sound like this is a proven fact.
I think a more accurate term would be confabulate based on your explanation.
No?
An anthropomorphic model of the software, wherein you can articulate things like “the software is making up packages”, or “the software mistakenly thinks these packages ought to exist”, is the right level of abstraction for usefully reasoning about software like this. Using that model, you can make predictions about what will happen when you run the software, and you can take actions that will lead to the outcomes you want occurring more often when you run the software.
If you try to explain what is going on without these concepts, you’re left saying something like “the wrong token is being sampled because the probability of the right one is too low because of several thousand neural network weights being slightly off of where they would have to be to make the right one come out consistently”. Which is true, but not useful.
The anthropomorphic approach suggests stuff like “yell at the software in all caps to only use python packages that really exist”, and that sort of approach has been found to be effective in practice.
I’m honestly starting to get tired about “people confuses advanced chatbot with Jarvis and bad things happen”.
Specially when it’s shitty/lazy devs that don’t code review.
