AI gfs incoming!!

Alright, I'm really excited. Super duper excited, even.
But I'm not desperate, alright? I really don't have a burning desire to have a gf. I'm just a little curious, that's all. The idea of an AI girlfriend is awesome in that it requires no effort to search for and maintain the relationship, and carries zero risk almost by design. This sounds like a great demo! Or perhaps something even better than the real thing? Either way, I'd like to give it a shot.

There's just one thing that concerns me, and that is the question of privacy. More precisely, it's the open source question, which I believe to be a necessary prerequisite for privacy. Unfortunately, I don't think it's looking good for AI waifus: The reason is that it has become somewhat of a staple of AI development that it's big tech that releases the state of the art, while the open source community keeps lagging behind due to lack of compute power or data. Ignore the benchmarks, just try any of the available models and see for yourself that they are far behind what OpenAI has to offer. Most of the time, these models are dumb as rocks, and when they aren't, then you can't run them on your midrange setup anyway.

I doubt open source will overtake big tech anytime soon. Actually, I can only see the gap widening in the future. You can't create a good artifical companion without knowing how companionships work. This is obviously true for many things; it's hard to create an X-doer without knowing how X works, but while a company's private data is certainly helpful in training, say, a better coding assistant, you can still get pretty far by just scraping publically available sources such as github and stackoverflow. But it's a different situation here. Where would you scrape relationship interaction data? It is usually transferred over encrypted channels, or is at least heavily guarded by the owners. The only viable option is collecting that data as the users interact with your bot, which of course means that you would be infringing on their privacy.

Given how indifferent people are towards the privacy violations happening right now, I doubt it will deter anyone from talking to Microsoft's sexy chatbot. Take Replika¹: Initially released in 2017, it is one of the first semi-successful artificial girlfriend services. They also state fairly openly in their privacy policy that they collect your chat data, though they claim not to share it with advertisers at least. Read mozilla's report on replika.
Then again, Replika does not yet have mass appeal. Maybe it's only the truly desperate who are using it, and when (if ever) these products hit the masses, people might finally start giving a shit.

What will be the business model for AI waifus? Paid subscriptions are one option, that is how Replika operates. But I have a better idea: Take a leaf out of Google's book! Use them as a mass surveillance & manipulation tool. AI girlfriends could become some incredibly effective dataminers – not just passively observing, but actively inquiring the users about their preferences, or even subtly manipulating them in order to direct them towards the gaping mouths of the "partners" with the largest bags of cash. Get ready for truman show-like sponsorship interruptions.

..Yeah ok, maybe I will hold off just a little.

Still, it will be interesting to see how this develops. Change is interesting! How will AI girlfriends affect the already plummeting fertility rates and the climbing virginity rates? Will AI waifus even have mass appeal, or is the lack of physicality a large enough drawback to keep it contained within loser techie circles? (I am reminded of the movie "Her", which in 2013 predicted that the first human-ai relationships will be non-physical, though the ending was rather shit).
And most importantly, will /g/ have regular flamewars between win-tan and linux-tan users? I imagine the former will have to deal with a mostly functional, but bloated waifu that occasionally asks them to install candy crush; the latter will furiously defend their riced, slightly broken, but free AI that they have somehow also connected to a DIY robotic abomination they call their "rig". >>>/robowaifu/

ChatGPT!

A good chunk of the code on this site was written by that adorable guy. I've found it to be super useful for simple coding-related tasks. Other times, it's much less useful, in fact, sometimes it's a braindead piece of stupid that fails to block my 2/3 line in tic tac toe. It also tends to choke whenever it is asked to solve any math problem or logical puzzle. I'm fine with that though. Yes it's not exactly Einstein, but it is still an obviously helpful tool for many tasks that aren't demanding, and for the ones that are, I use the following trick: I don't use it for these tasks! Once I discovered this lifehack things started going a lot more smoothly. I know, it's simple, but what I keep seeing online makes me feel like I'm some kind of genius for discovering this. You've got the two extremes:

• panicked OMG L-LITERALLY AGI !!
• smug heh, it can't even solve a millenium prize problem. Just useless autocomplete

I feel like the former may cause trouble in the near future. Google fired an engineer who was convinced that one of Google's language models was sentient (here's the full transcript, if you're curious). I can already see some groups of particularly deranged people starting to advocate for AI rights. [Update: Check out this reddit post].

But man, argument 2 is just plain annoying, especially because it keeps being peddled by otherwise smart people. Very eloquent, accredited and again, clearly very smart people who speak in a confident and authoritative tone. They sound right, they really do, just like the ones who said that internal combustion engines couldn't be made, or that "horseless carriages" would never replace the trusty horse, or that flight was fundamentally impossible. Or, the ones who worked hard to downplay the importance of the internet, calling it "just a fad". Hell, when I read those essays, sometimes I even start to believe that they're right! Maybe the internet was just a fad? Maybe it failed in 1996, like the confident experts had predicted? But it hasn't.

I already wrote about why I think the "stochastic parrot" is a dismissive, low effort argument, but in hindsight that post may be going a little much into the philosophical aspects while ignoring the more important question of capability. It makes no difference whether it fits somebody's crappy definition of intelligence. We already know that technology doesn't have to be a carbon-copy of biology to be useful (planes≠birds, cars≠horses). ChatGPT's descendants probably won't have to be "conscious" to take your job either, or to automate a large of chunk of it (leading to reduced demand for your work). I think Scott Aaronson puts it nicely:

In a certain sense you’re right. The language models now being adopted by millions of programmers don’t write working code; they only seem-to-write-working-code. They’re not, unfortunately, already doing millions of students’ homework for them; they’re only seeming-to-do-the-homework. Even if in a few years they help me and my colleagues do our research, they won’t actually be helping, but only seeming-to-help. They won’t change civilization; they’ll only seem-to-change-it.

I do not share his confidence, but I think he addresses the goalpost-shifting quite well.

Side note: I suck at writing. I don't mean it in a "haha, sometimes I make a typos" kind of way, no, I mean that it often takes me hours to get a single paragraph right, and even after that there's still something missing. Then I go online and see people casually jotting down Shakespeare-level writing in a pornhub comment. This is honestly a bit discouraging, and makes me want to simply pipe my raw thoughts into ChatGPT in order to skip all that pain and just make it more readable. Yeah, its style is a little uptight for my taste, but I could just redact it and add all the profanities I want afterwards.. That's an appealing thought, but at the same time, there's just something about it that makes me deeply uneasy. It's not just that I want to actually learn writing and organizing my thoughts (unlike how to use yet another api or library), but also that it feels like I'd be replacing a part of me that really shouldn't be replaced.

I've come to realize that normies are fully aware of the fact that Google, Facebook, and others have direct access to their souls. This is not, like some people put it, a case of companies silently depriving us of our rights – normies just don't care that much. To make them switch to a privacy-focused alternative, your product needs to be easy to set up (preferably installed by default on their favorite OS), be fast, and just werk. Unfortunately, I doubt that privacy-oriented products will ever satisfy those criteria.

"Death of the author" fucking blows, man. It's just a silly way of making any media you consume about yourself. You're closing your ears and blabbering to yourself like a retard, but in truth it is never about that one specific topic that is coincidentally so very relevant to you.

I'm not saying that connecting with a story is bad. It obviously isn't. But reading a book, or watching a movie, or an anime, or playing a game can be a form of communication, in way. The author (an actual person) is trying to say something with their work. Why not at least hear them out?

1. "Language models are just predicting the next word"

Here's what I can make gpt3 predict the next words of:

"This is an essay about why the yellowstone supervolcano is unlikely to erupt, interrupted by a paragraph of a random string of words in the middle, then a valid round of tic tac toe, before the original essay concludes:"

If you wrote this in a note on your phone, how likely would it be for your keyboard's autocomplete to predict a valid response? So you press the default autocomplete option a few times and most likely it is already nonsense after just a few words, but let's assume that through an incredible coincidence it has generated an entire paragraph about the yellowstone supervolcano. It will have no trouble writing the random words, but the chance of it successfully generating a proper round of tic tac toe after that is zero. Gpt1 or 2 would fare slightly better, as they would probably consistently output something about the yellowstone. But where both would definitely fail is the random sequence of words, because this prediction is simply far too unlikely given the coherent paragraph preceding it. GPT-3, on the other hand, is different. Or just way, way, way better, if you want – It really doesn't make a difference whether it's a qualitative or just a massive quantitative improvement, the fact is that it knows how to follow instructions, for any reasonable definition of "know".

You can still say that the output is likely given my prompt, but so is a human's in most cases. How is GPT-3 able to perform this task which does not appear a single time in its training set? It generalized. It has to be at least slightly more intelligent than a rock, no?

1.5 "But it is literally just trained to predict the next word. They also need terabytes of text to do that, far more than any human will read in their lifetime. It's a stochastic parrot"

People like to forget that they're not born intelligent. It takes decades of training on incredible amounts of data from all your senses to go from a crying, pant-shitting retard to a somewhat functional human being. Hell, some never even reach that state. And the reason why we're able to learn so quickly in the first place is due to the billions of years of pretraining we have have received as we evolved into humans. With that in mind, evolution really doesn't seem all too different – it too is a "dumb" optimization algorithm that, when given enough time and data, can give some impressive results.

LLMs have a more difficult task. They must reach general intelligence, but do so by only analyzing text and using a small fraction of all data that was made by us. I'm not sure if that's even possible, but that's why I am excited for multimodal models. They will still have the goal of "just" predicting the next token, while we were trained on the far more sophisticated objective of fucking and reproducing as much as possible. It worked, because it turns out that you can get surprisingly creative about how to maximize reproduction. Likewise, predicting the next word can be hard. What if you were tasked with completing a new elaborate mathematical proof? In order to do this well, you'd probably need to know maths. (Edit: Or, maybe you need to predict proper moves in a board game. Rather than to try and just guess a common legal move it would probably be useful to create a mental model of the board keeping track of the pieces, and then use it to make predictions. Seems like LLMs have figured this out as well). There is no good reason to believe that next word prediction isn't general-intelligence-complete.

2. "Biological neural networks are different" & Quantum Bullshit

I don't have much to say here. It's true, they are different, just like a plane is different from a bird.

Both fly though.

3. The Chinese Room Argument

The chinese room argument is one of the more braindead thought experiments that still regularly does the rounds online.

In the experiment, you imagine a room where an English-speaking person sits with a book containing instructions in English for responding to Chinese characters slipped under the door. The person follows these instructions and sends responses back, making it appear to the outside observer that the room understands and speaks Chinese. However, clearly the person inside doesn't understand a word of Chinese!
The point, then, is that a computer would be doing the same thing as that person in the room while it's running an AI. Therefore no program can ever be intelligent.

But it's not the human that would understand Chinese, it's the human + program combo.

4. Chinese Room 2

There is a more interesting version of this argument which I call "Chinese Room 2". It goes like this:

Suppose an English speaker sits in a room filled with Chinese books. As they look at the texts, they begin to pick up patterns. Given enough time, the person may learn to construct full, grammatically correct sentences. With even more time, they could learn to put semantically similar sentences next to each other. And after a very, very long time, they could even pass as Chinese speaker, despite not actually understanding Chinese.

Unlike the first one, this is more of a statement about the quality of a model's internal representations than a thinly veiled soul argument. A stochastic parrot like this would pass as a Chinese speaker, but would likely fail outside of distribution, because it only learned surface level statistics and doesn't understand what the objects and concepts really are.
This is backed by the fact that language models are worse than humans outside of distribution. And obviously, there is also no way a language model will ever understand what roundness is to the extent we do, because it can't actually see the points being equidistant from the center.

It's just that I don't think the way we process and store knowledge is fundamentally different. We too only really have relational knowledge, although we do have data from our other senses as well, which admittedly may be a significant difference. The LLM knows that a "ball" can "roll", or that something called "wheel" can be attached to a "container" and be used for "transportation". It's true that these labels might as well just be random strings of letters (or, you know, vectors) for the model, but the relational knowledge is still there, and that is the type of knowledge that matters. It's only when you contextualize data with other data that it becomes meaningful, and this is something language models already do.