karmickoala

Thank you for bringing it up. I added BERT intentionally to draw attention on this fact that I'm unable to convey properly yet: I was okay with the use of previous generation of models, so it's not that I'm against LLMs since BERT-like architectures were indeed used.

What worries me about the current generation of models is privacy. In another comment, I tried to draw the line where models start to worry me regarding it. Discussing it with whimsicalism, we noticed that it's not the generative aspect either; model size is probably the main factor. For instance, a larger encoder could make its embedding transform almost lossless and thus more prone to data leakage again, although a decoder-like makes it more susceptible.

Non-associativeness of floating point arithmetic plays a major role as well.

I know, I mentioned it because there's a category of models that I'd be okay with, but I can't exactly name it. I thought the difference lied in the encoder/decoder distinction. As a sibling comment mentioned, generative models do offer more privacy risks.

But it's not only that. Model size, for instance, also plays a role. Will modern BERT keep their lossy compression nature? Or could there be a privacy risk as well?

Thus, what are the characteristics of privacy-aware models? So far: discriminative vs. generative, model size, and lossy transforms (related to size). Perhaps looking at pre-2020 models could help to narrow other necessary attributes.

Somewhat related: did you know that you cannot disable GMail's GenAI features without disabling the traditional ML powered features, such as smart inbox, which predates LLMs? In this case, Google/Alphabet is conflating GenAI/LLMs with traditional machine learning, so you can't opt out of genai---which necessarily requires access to your data by the models---without deactivating the previous non-LLM powered features.

Which I get it, they indeed can use LLMs to improve the previous traditional ML pipeline. However I believe bundling these options together shouldn't be legal and I never expected Google to employ dark patterns like this. When I subscribed, these features were provided by traditional ML, perhaps some BERT-like encoder architecture at most, however now I cannot keep that previous functionality if I don't wish to grant LLMs access to my data.

First they came for the artists

And I did not speak out

Because I was not an artist

Which one is more at risk of having the wipe-out and cost levers: your FOSS language/compiler or your paid third party LLM?

Sure, there are similarities, but while we're figuring out a FOSS solution that will run on your machine and we'll have access to more than the weights, the abstraction ladder isn't as equivalent as we think they are.

> I'll read over more of your thoughts and update if anything else sticks out to me.

Thank you for your replies, I'll muse over this discussion as well.

> From a legal-representation-POV, obviously you want a bigBAD asshole representing you (it's just how the game works).

This one I cannot think of a counterexample right now. It does seem necessary.

> Once you get past the outer asshole wrapper, these are the people that are driving our societies (for better or worse) "forwards" – do not forget this – they are assholes because they care so much about their work/goals [∑]

I forgot to address that you mentioned explicitly that assholes are pushing our society forward and they are assholes because they care about their work. Be careful to not conflate expertise and assholeness. As I mentioned in my comment, assholeness is a disguise; it's not a necessary condition to be excellent in your field. It may help one to achieve their own goals sometimes --- which in general revolves around getting more money to the highest bidder and themselves. They care so much about their work that sometimes they have a huge blindspot that prevents them from seeing the tragedy of the commons right in front of them (per Upton Sinclair's law). No, it's not expertise, it's lack of character. They are simply unaligned with what's good for society vs. what would bring them more money. Steve Jobs was an expert in some fields, but he needed Steve Wozniak. Wozniak is a great counterpoint to Jobs expert-assholeness: he was an expert (a programmer's programmer), cared about his work, and for all accounts, was kind and ethical. As an example of just an asshole, there's Martin Shrkeli and several other politicians that fit the bill --- just assholes, not experts, pushing their own goals.

The argument here is that "know-it-all assholes" arise spontaneously if you excel in a field and care so much about it, which I kind of agree with. The corollary here seems to be that several people associate being an asshole as meaning you're good at something. However, that's just an excuse for being an asshole, as it's not a necessary condition to excel at something. It's only through our lack of discipline to tame our own minds that leads us to becoming assholes in general and using expertise as an excuse.

Those should not be ideals we strive for; you can excel without sacrificing social niceties and we should strive for intellectual humility. As one gets better in their respective fields (specially in ours that is always changing so fast), one should try to keep its ego in check. There is a saying in my country: "No one is so big that they cannot learn, nor so small that they cannot teach."¹

The problem with "know-it-all assholes" is that most think they are always talking to someone below them and their expertise carry on to all other fields². Their pride may prevent them from admitting from learning from even the most ignorant of beings.

As an exercise, ask your brothers the times that they were wrong (specially if there was anytime that came from someone that surprised them). If they fail to recollect many stories, I fear you're only hearing one part of the story and you'll never know it yourself. Steve Jobs himself used to do that: people would bring ideas to him that he would dismiss. A week later, he would come back with the idea presented as his.

We should strive to be gentle and humble to anyone, despite success in a field: we'll never know who we're talking to and what they can teach us. Perhaps is our misconception that we think that we're average (or above average) that prevents us from learning from other people.

-----

¹There is a variation, "No one is so ignorant that they have nothing to teach, no one is so wise that they have nothing to learn.", but I kinda like the former because it fits perfectly with the Lord of the Rings.

²I'm not raising the Dunning-Kruger card here.

They may have been made scarce by many methods, perhaps a low print run; court rulings; burning in a revolution to suppress dissenting views; scanning to avoid competitors to get the content and prevent other LLMs to reference, search, or train on it. Perhaps it's a particular edition that is rare (e.g., the first edition had a different description and was thoroughly changed in the second edition of the book, which may be important to have all the facts from someone's biography).

I think it's your (implied) second suggestion. It's interesting that someone here even predicted that AI companies would walk back such statements, I think it was in the Eric Schmidt booing thread.

O.J. Simpson comes to mind.

"I'm doing 1000 calculations per second and they're all wrong"

There's a big asymmetry of power here and "information wants to be free" was about empowering the people. Currently, corporations are bad faith actors that corrupted the idea, making it free only for themselves. Can't you see the asymmetry here? For instance, they should release the weights of the models they trained on everyone else work, but we're not seeing that except for Meta and some other groups.