But I disagree with the underlying assumption that you can anthropomorphize LLMs. Gradient descent and backpropagation don't take place in the brain. LLMs "learn" in the same way that Excel sheets "learn".
Humans are living beings with needs and rights. A person being able to legally squat in a home doesn't mean that a drone occupying property for some amount of time also has squatter's rights, even though you could easily and affordably automate and scale the deployment of drones to live and hide away on properties long enough to attain rights regarding properties all over the country.
I await the HN ban with fear..
[1] I'm not even doing referencing - so I am surely an LLM.
but, more importantly, OpenAI can also be sued for tortious interference? (basically the civil equivalent of accessory)
That's function of the legal system, not of the technology. If tomorrow someone made a perfect dolphin-Esperanto translator and proved Dolphins were as smart as humans, you still can't sue a dolphin until the legal system says so.
Not exactly, no, but the 'neurons that fire together wire together' way of learning has a pretty similar effect.
> LLMs "learn" in the same way that Excel sheets "learn".
I've never seen an excel sheet do anything like backpropagation.
Not strictly in the sense you mentioned (assuming that you mean "by themselves") but people may find [1] and [2] interesting.
[1] https://pub.towardsai.net/building-a-neural-network-with-bac...
[2] https://towardsdatascience.com/demystifying-feed-forward-and...
Backprop doesn't happen in us, but I think our neurones still do gradient descent – synapses that fire together, wire together.
And ultimately, at the deepest level we can analyse, our brains' atoms are doing quantum field diffusion equations, which you can also do in an Excel spreadsheet, so that kind of reductionism doesn't help either.
> Humans are living beings with needs and rights. A person being able to legally squat in a home doesn't mean that a drone occupying property for some amount of time also has squatter's rights, even though you could easily and affordably automate and scale the deployment of drones to live and hide away on properties long enough to attain rights regarding properties all over the country.
Yes, but we can also do tissue cultures and crude bioprinting, so it's a very foreseeable future where exactly the same argument will also be true for living organisms rather than digital minds.
We need to figure out what the deeper rules are that lead to the status quo, not merely mimic the superficial result. The latter is how cargo cults function.
Sure, that's an interesting path of inquiry, and one should be free to understand themselves as being no different than a machine if they desire.
But the objective of laws is the benefit of (at least some) humans, not machines covered in lab grown tissue. The process of being human is a big part of what makes us human.
I think you're misapprehending — I mean an entity fully 3D printed out of tissue, no machinery (unless you're counting all biology as machinery, but I think you're not doing that).
I recon bio-printing is now where home computing was in the Apple 1 era, so this is a way off, but it's foreseeable.
> The process of being human is a big part of what makes us human.
Mmm. How much has that process that changed since the ancient world?
How do you recon that, Apple 1 was Turing complete. We haven't printed life, that would be a tremendous accomplishment.
I think we're closer to Edison inventing a lightbulb as a step to computers being possible. Printing a conscious thing, at all, would be like the transistor. An Apple 1 analogue wouldn't be likely because of the terrible ethics of a "shitty" printed human.
Sure we have, and in multiple different senses.
The ones which matters here are cell culture, which is nowhere near the fanciest bar that's been surpassed in this field, and tissue culture which is somewhat harder but the reason why I recon it's at the Apple 1 level is that a small number of experimentalists are messing around with it using expensive equipment that you can technically buy at home but you need to be well trained to actually use, for example:
https://youtu.be/Z_ZGq8Tah0k?si=u6bBatjuSWcyNYJ3
And, more broadly, there's bioprinting as a research etc. field:
https://en.wikipedia.org/wiki/3D_bioprinting
And here's a TED talk from ten years ago where they demoed an early research 3D printed kidney on stage:
No. That isn't printing life, that is taking already living cells, priming and transforming them into something useful. Regardless, I'd count it if we could make an entire living organism this way, but we cant. Creating a working organ is no doubt amazing, and proof that this technology is worth pursuing, but it isn't "printing life" any more than producing life saving drugs is.
In your example you are talking about being able to bioprint a person(they have to be a person to have that right) to squat a property. Bio printing an organ isn't an example of that, it's not even close. Saying that we are anywhere near being able to print a human to squat a property is pretty ridiculous.
Which is absolutely sufficient for the usage I described upthread. In fact, I'd go so far as to say it's mandatory for the point I was making, as — fun though bio-printed werewolves, dragons, and fae would be — my point only works if you get humans out of the process rather than some other species. A bioprinted horse is probably slightly harder than a bioprinted human, but the latter isn't getting any squatting rights.
I could've linked to work on synthetic genomes and nucleotides to give evidence for lower-level creation of live, but they don't matter for the same reason:
My point is that there's a pathway heading off into the distance, and somewhere in the distance but before the horizon can be found bio-printed humans with all the same moral issues we're now just beginning to taking seriously thanks to AI being conversational, and if we had something completely customised, that's cool and all, but it doesn't make anyone go "oh, they're people" the way a humanoid body with human DNA getting off a table saying "hello, nice to meet you" does.
> In your example you are talking about being able to bioprint a person(they have to be a person to have that right) to squat a property. Bio printing an organ isn't an example of that, it's not even close. Saying that we are anywhere near being able to print a human to squat a property is pretty ridiculous.
I wrote "an entity fully 3D printed out of tissue […] is a way off, but it's foreseeable" and compared bio-printing today to a nearly 50 year old computer, and one of my references was a link to a youtube channel where someone is attempting to do a small-scale prototype thing along these lines with a handful of organs made from mouse cells grown in his own lab (and mouse cells rather than human because of the disease risk not because something magic happens with human cells). You're mixing up what I think is foreseeable with what I say already exists, and using the nonexistence of what I think can be foreseen to argue against what does exist.
No! Hebbian learning is categorically NOT gradient based learning. Hebbian update rules are local and not the gradient of any function.
Cortical learning is so vastly different from how artificial neural networks “learn” they cannot even begin to be meaningfully compared mathematically. Hebbian learning is not optimization and backprop is not local learning.
Part of the problem of these discussions is a bunch of clueless people talking with authority.
I have to keep reminding myself that outside of my own speciality, ChatGPT knows more than me despite its weaknesses, so I bet ChatGPT knows more about Hebbian learning than I do.
I'll look into that more.
most of the world disagrees with this view, and that means they will create the AI that wins.
You misunderstood me. I was talking about something more fundamental.
Understanding is data compression. They are the same thing. Learning patterns, building mental models, creating abstractions, generalizing, gaining intuition/a feel for something - all the things humans engage in as part of learning and understanding the world - are all acts of lossy data compression.
Humans have rights, software tools don’t.
If you grant an LLM the full set of human rights, then it can consume information, regurgitate copyrighted works, and use it to generate money for itself. However, considering blatantly obvious theft as “homage” goes hand in hand with free will, agency, being in control of yourself, not being enslaved and abused, etc. Pondering various scenarios along those lines really gets to the heart of why an LLM is so very much not a human, and how subjecting it to the same treatment as humans is a ridiculous notion.
If you don’t grant LLM human rights, then ClosedAI’s stance is basically that pirating works is OK because they pass them through a black box of if conditions and it leads to results that they can monetize. That’s such a solid argument, it’ll surely play well in the court of law.
Training data is not an “LLM does it”; first because “it” here is not “learning” or understanding in human sense (otherwise you would have to presume that an LLM is a human), and second because a software tool doesn’t have agency and it’s really just Microsoft using a tool based on copyrighted works to generate profit.
What I expect to happen is whoever has the most influence and power will get what they want and we'll end up raising a generation with the implicit understanding of "that's just how things are," natural order, truth, reality, and all that jazz.
The only thing that ever changes outcomes is if the contradiction status quo is incapable of being managed.
Our collective human limitations(physical, mental and temporal) are sort of invisible implicit rules that we all follow in one way or the other. If an entity is not bound by those rules then I don't see why that entity should be treated the same as a human.
Companies already make this differentiation.
For example take captcha and bot detection. Some of the heuristics are based on inherent human limitations like response time, click time, mouse acceleration etc.
I doubt youtube or any other streaming service will be happy if you want to stream all their videos to train a hypothetical human like AI(which views and prepares notes like a human) at a hugely accelerated speed compared to a regular human. You can guess how quickly they will cite fair usage policies.
What I want to say is there are fundamental differences between a human and an AI. So, we should not be quick to dismiss any concerns just because AI can "mimic" humans in certain areas.
Anyone got more details on this?
Superficially it sounds like total BS; a highly compressed zip file does not exhibit any characteristics of learning.
Algorithmically derived highly compressed video streams do not exhibit characteristics of learning.
?
I’ve vaguely heard the learning can be considered to exhibit the characteristics of compression in that understanding of content (eg. segmentation of video content resulting in more highly compressed videos) can lead to better compression schemes.
…but saying you can “do a with b” and “a and b are fundamentally the same thing” seems like a leap…?
It seems self evident you can have compression without comprehension.
An LLM has limited parameters. If an LLM had infinite parameters it could just memorize the results of every single addition question in existence and could not claim to have understood anything. Because it has finite parameters, if an LLM wants to get a lower loss on all addition questions, it needs to come up with a general algorithm to perform addition. Indeed, Neel Nanda trained a transformer to do addition mod 113 on relatively few examples, and it eventually learned some cursed Fourier transform mumbo jumbo to get 0 loss https://twitter.com/robertskmiles/status/1663534255249453056.
And the fact it has developed this "understanding" as an ability to learn a general pattern in the training data enables it to compress. I claim that the number of bits required to encode the general algorithm is fewer than the number of bits required to memorize every single example. If it weren't then the transformer would simply memorize every single example. But if it doesn't have space then it is forced to try to compress by developing a general model.
And the ability to compress enables you to construct a language model. Essentially, the more things compress, the higher the likelihood you assign them. Given a sequence of tokens say "the cat sat on the", we should expect "the cat sat on the mat" to compress into fewer bits than "the cat sat on the door". This is because the latter is far more common and intuitively more common sequences should compress more. You can then look at the number of bits used for every single choice of token following "the cat sat on the" and thus develop a probability distribution for the next token. The exact details of this I'm unclear on. https://www.hendrik-erz.de/post/why-gzip-just-beat-a-large-l... this gives a good summary.
I fundamentally disagree. That's not some established fact, just a narrative used by those who wish to plagiarize using "AI".