I am ethically and morally opposed to gen AI as a concept. As far as the breaking point, let me explain. I’ve been a machinist for ~18 years. It’s been really bad for the last 9 years and I’ve more or less had to live off of savings from the good years (those savings are basically gone).
I’ve just recently gotten legitimate job offers for ~$125 an hour remote because I own a CNC. The catch is that I have to record all of the CAM programming and machine time in order to train an AI to replace potentially hundreds of thousands of jobs.
This fucking sucks. I’d love to sit on my ass forever at 43, but I absolutely fucking hate the potential that I could bring that would crush the working class and elevate the owning class.
I wish it wasn’t this hard, because I really do believe in socialism and have labored hard for a long time to make it happen - but there’s a chance I can put everything I have into true anarchism cause and make a legitimate difference.
Please help.
I’ve worked on the part of the pipeline that annotates the type of videos you would record. Things like cleaning the data or marking down what the machines are doing.
Contrary to what others have said, it would be difficult to poison the dataset. Training data works in aggregates. Anything you do out of the ordinary will be flagged by the human annotators or the automated systems looking for deviations from the norm.
Many people are already doing what you’ve been offered. Their data is already being processed and fed into LLMs. They want as much data as they can get. That’s why they’re paying so much to every professional in the world right now.
I honestly don’t know how one might sabotage the system but it’s not as easy as you all think. The militantly anti-AI crowd likes to think of LLMs as more incapable than it actually is. The systems will spot obvious things like randomly tabbing and spinning the cursor. AI is better than that now. Human annotators will flag those easily as junk data too.
I’ve faced as similar dilemma as you. I need money so I did that job even though I hated it. I’ve seen obvious acts of attempted sabotage in the data. We get penalized for not spotting these things, even as bottom rung data cleaning peons.
I wouldn’t waste my time trying basic trickery. This is like day 1 stuff they tell us to look out for. Also you will immediately be noticed and fired for giving them junk data.
Logically the next thought might be to insert some high level errors that only the top experts could spot. I honestly don’t know if this even works, but I doubt it. They have experts in the pipeline too. So you might end up giving them data about errors that only experts could spot. That would only make their dataset even better.
Thank you for an insight from someone that actually seems to understand what’s going on. My stomach twisted and turned as I signed a contract for General Dynamics because I had to eat. They won’t let you poison it because they need a measurable good part or assembly at the end of it. Small changes can easily affect small tolerances. “Oh just do it sloppily!” absolutely does not hold up to scrutiny when they have CMMs measuring what you’ve done vs. what they see. If I have to hold a good tolerance of 0.002" they will measure and with that along with finishes etc. they will know it is not up to snuff.
They have multiple people annotating the same video. They look for discrepancies in how data is annotated between people. Probably most often it’s annotator error. It also allows them to spot actual oddities that maybe only one person noticed. There’s a lot of redundancies. I think some might have a conception of AI-bros recklessly feed raw data into LLMs and sell people the garbage-out machine. It’s not that at all. The source content and the written records are validated. It goes through refinement again and again to check. Who knows how many times. I’ve gotten by own annotations back at much later stages for further verification. That comes after even more people have looked at it and done whatever their assignment was with the data.
This inherently has nothing to do with AI. It’s mass data harvesting and refinement. It’s bigger than LLMs. Whatever technologies supersede LLMs will be fed these datasets. If they do it right once, they’ll never need you or any professional ever again regardless of whether LLMs ultimately flop. They won’t need you for the next iteration of AI technology. In theory of course.
The business here isn’t AI. It’s what big tech has always been after. Data.
Even if you get fired you wasted their time and made some money. Just have another job ready.
Introducing high level errors costs lots of money. The later in the pipeline stuff gets noticed, the worse it gets.
Don’t listen to someone telling you to not listen to other people. I’m a level 69 LLM wizard and clearly outrank the guy annotating stuff. I never faced his dilemma because I’m so far into the ass of AI companies that they ask me to polish their plumbobs.
Jokes aside.
Poison the data. Get money.
You won’t significantly poison the model, but you’ll make it more expensive to train.
If it needs to be flagged it doesn’t enter the dataset. LLMs can’t be fed with errors, they’re regurgitation machines, not thinking machines.
The things that could be made better by you are adversarial networks, which can punish the LLMs in reinforcement learning but there’s an inevitable glass ceiling of model collapse in any case. These slop machines will always make errors, the broader you want them to get – that’s their nature. Nobody is doing AI right nowadays, in 95% of cases they’re just bumbling idiots in a gold rush.
Some errors will always slip through. The more insidious your errors are, the harder to spot they are, the higher the chance they won’t notice.
Do your worst.
The thing is though, when I say they need physical products as a result, they will be measuring them with a CMM (Coordinate Measuring Machine) that not only takes very accurate measurements, but also surface finish. I can’t do both incorrectly - ever. Also, see my update: https://lemmy.sdf.org/post/61854279
Can you make it take ages without the surface finish? Can you make the tools break after 2 or 3 of these? Or have very high wear and tear on the machine? Obviously not killing your own machine…
No, you can’t feasibly do these things because if the tool is cutting it’s making a finish no matter what axis it is cutting in. If they have to measure it and judge it by it being a finished part, it will be rejected.