I must confess to getting a little sick of seeing the endless stream of articles about this (along with the season finale of Succession and the debt ceiling), but what do you folks think? Is this something we should all be worrying about, or is it overblown?
EDIT: have a look at this: https://beehaw.org/post/422907
You must log in or register to comment.
This video by supermarket-brand Thor should give you a much more grounded perspective. No, the chances of any of the LLMs turning into Skynet are astronomically low, probably zero. The AIs are not gonna lead us into a robot apocalypse. So, what are the REAL dangers here?
- As it’s already happening, greedy CEOs are salivating at the idea of replacing human workers with chatbots. For many this might actually stick, and leave them unemployable.
- LLMs by their very nature will “hallucinate” when asked for stuff, because they only have a model of language, not of the world. As a result, as an expert said, “What the large language models are good at is saying what an answer should sound like, which is different from what an answer should be”. And this is a fundamental limitation of all these models, it’s not something that can be patched out of them. So, bots will spew extremely convincing bullshit and cause lots of damage as a result.
- NVidia recently reported the earnings they’ve gotten thanks to all this AI training (as it’s heavily dependent on strong GPUs), a trillion dollars or something like that. This has created a huge gold rush. In USA in particular is anticipated that it will kill any sort of regulation that might slow down the money. The EU might not go that route, and Japan recently went all in on AI declaring that training AIs doesn’t break copyright. So, we’re gonna see an arms race that will move billions of coin.
- Both the art and text AIs will get to the point where they can replace low level people. They’re not any danger towards proper experts and artists, but students and learners will be affected. This will kill entry position jobs. How will the upcoming generations get the experience to actually become trained? “Not my problem” will say the AI companies and their customers. I hope this ends up being the catalyst towards a serious move towards UBI but who knows.
So no, we’re not gonna see Endos crushing skulls, but if measures aren’t taken we’re gonna see inequality go way, WAAAY worse very quickly all around the world.
Hi @[email protected], finally got around to watching the video, thanks for letting me know about it.👍 One thing that really befuddles me about AI is the fact that we don’t know how it gets from point A to point Z as Mr. Dudeguy mentioned in the video. Why on earth would anyone design something that way? And why can’t you just ask it, “ChatGPT, how did you reach that conclusion about X?” (Possibly a very dumb question, but anyway there it is 🤷).
We have designed it that way because it works better than anything else we’ve ever had, the kind of stuff you can achieve with these deep neural networks is astounding, but also stupidly limited. As to why you can’t ask ChatGPT: Because it doesn’t know. It doesn’t know ANYTHING. As mentioned in the comment, all it knows is how to sound right, it’s a language model, all it knows it’s language. It knows nothing about the internal workings of AI, because it knows nothing besides language.
A lot of the fearmongering surrounding AI, especially LLMs like GPT, is very poorly directed.
People with a vested interest in LLM (i.e shareholders) play into the idea that we’re months away from the AI singularity, because it generates hype for their technology.
In my opinion, a much more real and immediate risk of the widespread use of ChatGPT, for example, is that people believe what it says. ChatGPT and other LLMs are bullshitters - they give you an answer that sounds correct, without ever checking whether it is correct.
The thing I’m more concerned about is “move fast and break things” techbros implementing these technologies in stupid ways without considering: A) Whether the tech is mature enough for the uses they’re putting it to B) Biases inherited from training data and methods.
LLMs inherit biases from their data because their data is shitloads of people talking and writing, and often we don’t even know or understand the biases without close examination. Trying to apply LLMs or other ML models to things like medicine, policing, or housing without being very careful about understanding the potential for incredible biases in things that seem like impartial data is reckless and just asking for negative outcomes for minorities. And the ways that I’m seeing most of these ML companies try to mitigate those biases seem very much like bandaids as they attempt to rush these products out the gate to be the first out the door.
I’m not at all concerned about the singularity, or GI, or any of that crap. But I’m quite concerned about ML models being applied in medicine without understanding the deep racial and gender inequities that are inherent in medical datasets. I’m quite concerned with any kind of application in policing or security, or anything making decisions about finance or housing or really any area with a history of systemic biases that will show up in a million ways in the datasets that these models are being trained in.
