- cross-posted to:
- cybersecuritymemes@lemmy.world
- cross-posted to:
- cybersecuritymemes@lemmy.world
You must log in or # to comment.
The very fact that Anthropic is now injecting a kind of watermark into every output, is solid proof that such a Ken Thompson hack is a inevetable risk
Ugh. Fuck “AI” and fuck Anthropic. But please read how the “watermarks” work. TL;TR its like a seeded run in a video game. With the seed and pseudo rng, you get the outcome i.e. the extruded text. In the watermark its the reverse, outcome + prng = seed. The result will be the same “quality” extruded garbage as before.
Good luck.
Lemmy is damn near the when it comes to wanting to hold an opinion on a topic without having first understood that topic.
That’s just a reflection of the real life and not unique to Lemmy, dw
A bit of a late response, but why are they mutually exclusive? I get that the “seed” can be extracted if you check for it, but who says that seed can’t be somrthing nefarious? Like, someone vibecodes a website. You visit the website, extract the seed and it shows you the complete userbase of that website.
I’m just sptiballing here though. It might not be the best example. I get where you’re coming from, but why we cant we both be right.
I don’t think you got how it works. Not being mean, might have explained it badly. Maybe read their blog https://www.anthropic.com/news/claude-text-watermark
Thanks, I’ll read it tommorrow (it’s almost midnight here)
Devs should be “dev managers and executives”. Real developers know LLMs are basically just a tool for finding examples and helping with syntax. Sure they’re useful, but I’d never let them write code, much less compile it. Who knows what they’d inject into a build.
What does that even mean. Whoever said that just uttered some empty but smart sounding catch phrase. Such is all the talk about the wonders of Ai
Linus Torvalds
Not exactly, but something along those lines.
We are talking about a guy still using mailing lists and patch files to conduct development on one of the largest codebases in the world, not exactly someone who jumps on any new shiny thing just to sound smart.
https://thenewstack.io/torvalds-ai-programming-productivity/
just to clarify, he’d not so much called LLMs “the new compilers” as he compared both to each other in a sense that an LLM is just another layer of analysis tooling between the developer and the final machine code, which, IMO, sounds much more reasonable than calling LLMs “the new compiler”.
I think it’s supposed to be that how AI turns high level instructions into code is compared to how compilers turn code into assembly. Implying that using AI is just a natural extension of the handing off work to the computers that we’ve already been doing.
I wonder if you gave different AI models some c++ or something and told them to write assembly based on it how well they would do compared to an actual compiler
I highly doubt they would manage to make AI spout out good assembly, they fuck up with high level code, imagine assembly
It’s a tool, use it where it works and don’t where it doesn’t.
It’s a tool, use it where it works and don’t where it doesn’t.
But that doesn’t work with the people creating AI as they are totally dependent on the believe that AI can do absolutely everything (and an artificial general intelligence is just moments away…) to justify they insane investments. So they will make up a million stupid narratives why some people are “actually” not using AI as it obviously can’t be because of AI shortcomings…
If the tool don’t work then don’t use it. I wouldn’t try to unscrew something with chopsticks
You are probably also one of those insane ideologues that refuse to hammer in a screw for some reason, although you know how well that hammer worked on nails… 😂
People probably say that about asbestos too.
And they’d be right
Of course, it still has uses. Just not in home insulation.
If a tool costs more to be made than the equivalent of buying every single person on earth a loaf of bread every single day for 300 days, only for it to have niche uses, that only helps what is possible with already existing tools, is it worth it?
That all depends on how it was made. Not all models cost that much. Chinese models were mostly produced for far less mostly.
And then your boss says: “Use it for everything or you’re fired. Why are you using less tokens than anyone else? You must be more productive or else…”
Na, fuck that shit. Just don’t use it.
Why? Do you have that feeling towards other algorithms?
Yes I do 😊
Like which ones and why?
Usually the ones that accelerate the destruction of our planet, perpetuate sexist and racist biases, pray opon the weaknesses of how our minds work, exploit the underprivileged, devalue humanity and artistry, are annoying and overhyped, I think you get the picture.
So which algorithm is that? If you’re hinting at manipulating the populace because in fact they are easily programmed via social network manipulation. Yeah anything done maliciously is of course bad but I think that comes down to the use of the tools and not the intrinsic nature of the tools themselves. Algorithms are just math, it’s the will of the individuals wielding them that is corrupt and that seems to be your actual chief complaint.
I did not say that. You, like everyone else that says “its just a tool and technology is neutral”, should read more books about ethics, politics and sociology. Everything we do is political. Nothing is “neutral”.
Pmuch this.
How do they feel about machine language translations? It’s been “AI” (ML model) since the late 90’s.
Traffic prediction?
Medical scan data processing?
And so so many more specific uses of ML models that have been a part of daily life for literal decades.
Call out LLMs specifically.
Did we burn down that planet for that stuff? Exploit workers to do weather forcasting?
Maybe? If you poison the prompt then there’s evidence and it can be undone. Poison fragments of the source training data, however, and that’s some KT shit right there. Enterprise foundation models cost bonkers money to train and pretty much slurp up all the data on the internet for mostly automated annotation. Stick something in an obscure part of the internet which becomes part of the training and produces the malicious response and it’s going to be both hard and expensive to detect or correct.
except for poison to take in, it should be a pretty significant part of the dataset. Also, ngl, i’m not much informed on the topic, but aren’t all the datasets, if we’re talking about generic diffusion models and LLMs, already been formed? From what i gather, the innovation in AI mainly comes from utilizing new architectures, rather than training a model on something unique.
The datasets are constantly expanding as new content is generated online. There’s a degradation issue currently where the models are training on incorrect data generated by previous iteration of their own or other models and effectively poisoning itself to more confidently give the same incorrect information in future.
i’ve heard of the dataset poisoning and degradation caused by llm-generated content present in the dataset myself, but i’m not sure whether it was a practical observation, or a mere experiment. And I still fail to see how new datasets are really useful for developing a new llms, or how it’s a problem for the devs to switch back to the older datasets.
And the cornerstone stays the same: to have any significant effect on the final LLM quality, shouldn’t the poisoned (either by llm-produced content, or by intentional poisoning) data portion be… well, statistically significant?
source?
https://www.nature.com/articles/s41586-024-07566-y
LLMs already tend to be the average predicted output for a given input; training on LLM data makes this worse and causes them to become less varied, less dynamic, more towards the mean generated by previous models, and more likely to spit out hallucinations
Should be relatively easy, with the amount of once trusted packages that become attack vectors
At work, I use AI for some things. Right now I’m rewriting some legacy spaghetti code that’s had a bunch of things hacked into it over the years. I spoke to the person most familiar with the expected behaviour and used AI to combine his info plus the existing code and unit/integration tests into a list of requirements.
I wrote the new code and tests based on the requirements rather than based on the old code. After each commit, I used AI to check for parity between the old and new code, and it keeps a Google Sheet up to date with the progress (which features were fully implemented, and which ones were missing or had gaps). I had AI write some tests cases too - given the list of requirements, write integration tests for them based on the style of a few tests I wrote by hand.
It has some quirks (eg for tests it loves over-mocking even though our skills tell it to mock as little as possible) but it definitely speeds things up.
I use AI for small side projects at work too. Tweaking and adding features I want to shared libraries, internal tools to help our team debug stuff and automate triaging of bug reports (they’re all still reviewed by a human), etc.
The entire reason I can trust its code is because I can read it and tweak it myself. I sometimes need to go through a few iterations to get AI code into an acceptable state. AI writing machine code directly, like what’s been talked about recently and what this post is referencing, is such a dumb idea.
There’s other people at work that use AI for absolutely everything. Writing code, reading code, writing posts in our internal groups, etc. That’s something I don’t understand. Some people that are all-in on AI produce so much low-quality AI slop.
I think that’s the distinction between an expert using a tool diligently and responsibly, and a lazy person using it haphazardly as a crutch.
If that tool ever gets ripped out from under you, you’ll possibly suffer a loss in performance, but you’ll still be able to perform and do your job.
If their crutch is kicked out, they’ll crash.
Yeah, I’m in a similar camp. AI is very good at dummy data and unit tests, fairly good at log traces and making fiddly changes across an entire codebase, somewhat good at boilerplate and making derivative code based on something similar, and somewhere between “passable” and “awful” at everything else. I use it for what it’s good at with heavy manual review, and just about the only time I run with the first thing it spits out without looking it over is if it’s something that is never intended to leave my machine.
I can also usually sift the usable from the not-usable in the automatic PRs that our company has turned on. In the past couple of months, the ratio of good comments to bad ones has gotten slightly better; from about 20% good to about 55% good.
But I’d never let an AI write a team message or acceptance criteria or anything. That’s a human job.
Time to get back to Assembly, I guess.
But do you trust the CPU?
Well, fuck
The CPU runs Minix, are you going to tell me you don’t trust Minix now?
Who??
notice
javascript required to view this site
why
measured improvement in server performance
awesome incremental search
Boo! Just give me the text!
Edit: It’s long, but here’s the opening section, at least:
In 1984 KenThompson was presented with the ACM TuringAward. Ken’s acceptance speech Reflections On Trusting Trust (http://cm.bell-labs.com/who/ken/trust.html) describes a hack (in every sense), the most subversive ever perpetrated, nothing less than the root password of all evil.
Ken describes how he injected a virus into a compiler. Not only did his compiler know it was compiling the login function and inject a backdoor, but it also knew when it was compiling itself and injected the backdoor generator into the compiler it was creating. The source code for the compiler thereafter contains no evidence of either virus.
Ken wrote, In demonstrating the possibility of this kind of attack, I picked on the C compiler. I could have picked on any program-handling program such as an assembler, a loader, or even hardware microcode. As the level of program gets lower, these bugs will be harder and harder to detect. A well installed microcode bug will be almost impossible to detect.
Ken does not mean bug in the sense of error, but in the sense of listening device. And it is “almost” impossible to detect because TheKenThompsonHack easily propagates into the binaries of all the inspectors, debuggers, disassemblers, and dumpers a programmer would use to try to detect it. And defeats them. Unless you’re coding in binary, or you’re using tools compiled before the KTH was installed, you simply have no access to an uncompromised tool.
In fact, given the amenability of microcode to the KTH, not even then.
All manner of controls and monitors could be secreted this way in the OSes of all the devices we all use day to day. It isn’t very far fetched to suggest that the hack, in software, can create an updatable backdoor. This way every piece of software on the planet can be KTH bugged without any possibility of detection by any mortal engineer anywhere.
Well, maybe with the diligent use of an electron microscope.
Given last week’s horrifying revelations concerning the US government’s TotalInformationAwareness of every US domestic phone call, it is difficult to imagine that the ThreeLetterAgency’s KTH-hacked binaries are not omnipresent. I mean, can you really imagine AdmiralPoindexter would pass up an ability like this?
not really related to the post, but NGL, it kinda fascinates me, how with the advent of LLMs, they became the ultimate punching bag for whenever something works bad, as if people didn’t write even more horrible things without them (my regards to javascript and python).
Will share the down votes with you, because I think the same.
Claude opus is better than a significant amount of people I’ve worked with, and it’s still significantly cheaper when now. Mind, we are not professional developers, but we need to develop a lot a DS solutions.
Even if AI was not hot garbage the other [insert number] % of the time, would that equate to better material outcomes for actual humans? Not under capitalism, that’s for sure.
What even is a “professional developer”? I mean, given that you aren’t an egg-headed RnD developing new technology for someone like Nvidia, the job of a developer is 95% mind-numbing routine and typical tasks solved by applying ready-made patterns.
deleted by creator
I think about this every day.
AI or not to AI is the same stories I heard from old time machinists when cnc productivity came for that industry.
deleted by creator
