AI-1: Let’s Make SuttaCentral 100% AI-free Forever

Essays
42

… And making it for independent research teams such as us impossible to provide translation services for, let’s say, Pali to Arabic or Suaheli based on SC data. That’s a pity. I think if there is one benefit for this type of technology, then it is in helping the marginalized communities that are off the radar of both large companies and academic mainstream to get access to the material in reasonable time.

I must think of the Newar Buddhists in Nepal who petitioned Google time and again to add their language to the system but it was always refused based on the fact that there are just not enough Newari speakers for Google to take any interest in this.

Now our Dharmamitra project has been very successful in empowering the Tibetan community to access Western material in their own language, and our work on Pali<>Tibetan (which now needs to be stopped since we have to remove SC data from the training process) gave us, for the first time in history, the situation that Tibetan monks don’t have to learn the colonizer’s language first in order to read the Pali canon in their own language. It was one of the most moving moments of my career when I saw Tibetan monks pasting Pali canon snippets into our system and getting a meaningful translation out of that. So, as people might have guessed, I’ll remain positive about this type of technology. :slight_smile:

I respect Bhante’s stance here and I will of course stick to his request. But I want to add the angle that his ban effectively hurts the smaller teams that actually care about his opinion, while the bigger teams that scrape the web indiscriminatively , will just continue. So, the academics lose, and big tech wins now.

8 Likes
43

You don’t really have to do that and are fundamentally and essentially free to do whatever you feel maximises the kusala potential of your own actions.

No one here is a sammasambuddha or above anyone to make your beautiful project a sin, offense or crime.

We are all puppets in Mara’s hands, propelled into this life by avijja, and hopefully trying to do the best we can with it as soon as we got in touch with the four noble and enobbling truths and its respective tasks.
:anjal:

3 Likes
44

I have good reasons to support your argument which you haven’t mentioned. This could easily be because, as you said, you are still writing content, so I am excited to see the rest.

I’ve been following the growth of AI since I was a kid (relatively not that long ago). I was originally majoring in data science, but I found their jobs to be boring in reality, so I did only computer science. I’ve still learned some about AI, and I’ve made AI’s, including my own neural networks. That doesn’t mean that I think AI is somehow good. It means I’ve seen very intimately what it exactly is and its potential. I see the advance of technology as oppressive and existentially disappointing in many cases. I’ve heard others say this, but they don’t really say why. I’ve seen in painful detail exactly how, down to the very code level, AI’s potential for good and evil. It’s use in the attention economy is especially insane (TikTok), but that’s not exactly your main point. This is one reason I’ve vowed to never work for a FAANG(-like) company.

By the way, AI is more powerful than the public knows since private businesses hoard the better models for competition purposes. I have seen this first hand although I shouldn’t prove it.

It’s good you specify AGI. Which excludes the AI which has done unimaginable wonders to medicine, but it’s also used to synthesize information in massive data tracking and selling schemes, which is also taken from its use in medicine itself. This will provide a lot more info about data privacy if its relevant to your argument Data and Goliath - Bruce Shneier.

This has information about the ethics of AI https://nickbostrom.com/ethics/artificial-intelligence.pdf. It discusses how there are many ways that AI is used to discriminate and hurt people unfairly. It also goes into a lot of depth about the (lack of) “morality” of an AI itself. Imagine an algorithm that serves you ads or even gives you opportunities which takes your race into account. Or, it could learn, based on user activity, if the user is suicidal, and it could easily learn to advertise guns to them. In fact, Facebook was able to detect which users were suicidal based on their post history, and they sent out help messages to those users, but imagine what else they could or did do with that data or the many organizations they could sell it to. From user activity, these models can predict, without asking the user anything, things like sexuality, race, age, and much wilder things, like if you’ve had an abortion, specific diseases, taking birth control, or even perfectly identify you (when you are not logged in / incognito).

Now, that is primarily about how this data is used for privacy, but I’m still assuming the primary problem for this thread’s topic is 1. LLM’s processing Buddhist texts/discourse and 2. People giving it prompts relating to Buddhism 3. How people use that data. Both the inputted data and the prompts are saved, tracked, synthesized etc. in all the ways I mentioned. It is harder to imagine how such data can be used to break privacy (note that the LLM isn’t entirely doing this step, it’s separate tracking), but AI is much more imaginative when it comes to that (discrimination by religion / belief / giving it your personal issues and struggles / data leaks).

3 Likes
45

Thanks Gabriel, we’re so lucky to have Michael’s work, he’s a careful scholar and a wonderful human being.

This is exactly the kind of problem that I am talking about. AI creates the illusion of understanding, which denigrates those who have taken the time to develop real understanding.

I understand what you’re saying, but to be clear, you’re working at Berkeley, so you are big tech.

4 Likes
46

Are you talking about using the English translations by Bhante Sujato to do this? Isn’t that a flawed method?

2 Likes
47

Agreed. I was looking to make an orthogonal point. One’s reaction to the phrase " AI is a tool" depends on one’s mental map of “tool” more than one’s opinion of AI. For some, tool conjureS up hammers and blenders and space heaters. For some, their mental map contains the full range of what the existence of the tool says about the society that makes it, the impacts of the tool, the revolutionary changes cause by the printing press, TV, and the worldwide web.

At the end of the day, I’m making a very small point. :slightly_smiling_face: That evaluating or reacting to the phrase “AI is a tool” might not be a very helpful activity before knowing what the speaker’s mental map is.

Note: I can think of lots of ways to challenge what I just said. I’m not going for air tight arguments. More just feeling playful about what is actually being said when people say “AI is just a tool.” Of course it is sometimes a way of dismissing certain useful discussions about why we built this tool, does the tool do more harm then good, etc. (E.g., guns don’t kill people, people kill people.) But I, FWIW, experienced the commenters on this thread who used the phrase as using it as an entrance point into those interesting discussions, not as a dismissal technique.

Anyway, fascinating topic. Loved your essay, Bhante. My Master’s was in Computer Science, and I studied AI. At that time - in the 90’s - a field with a lot of disappointed hopes. There were so many things that seemed easy for humans, that the first generation of AI theorists thought could be done, and were just really really hard or impossible to get computers to do. But with the re-focus from general intelligence (a thinking being like Data on Star Trek) towards using massive data sets to automate decisions and perform human-like acts and create human-like texts within limited scenarios, it has become massively powerful. And underneath the surface not a shred of judgment or morality. 100% stupid and 100% amoral.

There was a wonderful story I heard back on my AI class. They had created an AI using neural nets to detect if there were tanks and armored vehicles in images. They got to where the AI had near 100% accuracy with the training data. Shifted to new images and accuracy dropped to 50-50, no better than chance. Figured out that given original photos with the tanks and without the tanks were taken on different days, the AI had learned to distinguish sunny days from cloudy days. :rofl:

Despite the name, AI has zero intelligence. And it is absolutely immoral. So whenever we allow AIs to make decisions, we have no way to be sure what it learned to do is in line with our goals and values. And it has no intelligence or morality to check its actions against. So it is very hard to be sure its choices are aligned with what’s good for humans. And that’s when we’re trying to align our AIs with human values, let alone using AI to better disrupt democracies.

3 Likes
48

This reminds me of the “Guns don’t kill people, people kill people” discourse (which may be specifically American), with AI being the gun.

I suspect this happened with this “AI being able to detect autism from an eye picture” thing because all the kids with autism had the same glare of the doctor’s office in their eye.

3 Likes
49

Huh, amazing.

Imagine the mud on my face when I discussed AI briefly with Ven Juewei, and carefully asked if she was familiar with the concept of a language model, only to learn that she studied AI in the 80s in the US, in collaboration with Bell Labs. :flushed:

Ha ha, that’s useful too!

Alignment is such a thorny concept, I haven’t even been able to grok a way of critiquing it.

When setting up OpenAI, Musk and Altman basically said, “let’s make sure it is for the good of humanity. We’ll make a committee—you, me and three mates, that’ll do it.”

Money, investment, prestige. It’s just to big. You can’t put a billion dollars into it and then say, “maybe not?” That’s why the AI guys are always saying, “let’s do alignment” and then sacking people when anyone actually criticizes them.

5 Likes
50

This is part of the whole discourse around the ironically-named “OpenAI”. They’re not open at all, but then, should they be? I mean, I understand that a lot of it is smoke and mirrors, but there are obviously dangers involved. And yeah, corporations are not really the sharing type.

When they release their models, they say they want to get humans using them to get feedback. (Psst, it’s not the models being tested, it’s the people. We’re the experiment. They want us to tell them what we are willing to accept.)

Yeah, I agree absolutely with all this. The fact is, there are just so many problematic dimensions that I haven’t even focused on privacy much at all, but it is super major.

That’s true, I think people don’t really get how important this is for TikTok. As I understand it, it was originally based on AI (probably illegally) scraped from Chinese social media, which is where it got its world-beating recommendation algorithm from. Then the Chinese said, “This is terrible, it’s destroying the youth!” so they marketed it to the west and banned it in China. They have a much sanitized and censored version there.

An engineering safety report I saw on TikTok in the very early days—long before the current politicized atmosphere in the recent US Congress moves—concluded that it was data stealing malware thinly disguised as a social media app, an order of magnitude worse than any other social media, where the code was deliberately designed from the beginning to be as obscure as possible.

But that’s nothing. The real horror of TikTok, the thing that is ringing its death knell as we speak: it’s mostly millennials.

https://pluralistic.net/2024/03/21/involuntary-die-hards/#evacuate-the-platforms

The next big social media, which will capture Gen Alpha, will be built on AI from the ground up. It’ll make conversation and interaction much more fluid, generating memes and responses on the fly, altering conversations, condensing and expanding. It’ll decide what you want to see and what values you have.

https://twitter.com/YJernite/status/1777332127740092849

GJ0rAyNacAAWCwH
GJ0rAyNacAAWCwH1920×913 157 KB

2 Likes
51

Well then you use Big tech in a way that nobody in our world would use the term, which is okay.
The best hardware I can get my fingers on for training is 8x A6000, which is great compared to what I head in Germany (roughly 10x less) and much worse than even the smallest tech company will have. OpenAI can easily throw 1000-10000x more compute at a problem.
Its consensus that the universities, including MIT, Berkeley, Stanford, cannot keep up with the companies in this regard. Of course, students do internships at Deepmind, Openai etc. here and I know people who do that of course. But our project doesn’t have any links to these companies.

5 Likes
52

Maybe it was this Reverse Engineering Tiktok's VM Obfuscation (Part 1) | nullpt.rs - Someone painstakingly reverse engineered TikTok’s browser code to reveal the data it was collecting, and it was, well, basically everything it could have, even GPU data.

1 Like
53

But they could learn Pali, no?

We could also frame it as “For the first time in history Tibetan monks are using machine translations that have not been vetted by anyone qualified to judge if they are being accurately rendered in Tibetan.”

On a personal note, I do want to say that I appreciate you participating in this conversation. It’s valuable to hear how people are actually using this technology.

5 Likes
54

While you’re here, Sebastian, I wonder if you can tell us the energy consumption of running your model? Ven Vimala said it takes about a month to run it. That sounds like a lot of energy!

One thing I think AIs should do, and in fact should be legally compelled to do, is disclose their energy usage, both that for running the model, and for the individual queries.

Similar, but I don’t think it was the same article. I read it years ago and wish I’d kept it!

1 Like
55

Training a foundation model is expensive, yes, the good news is that once it is done its there, and requires little energy. So our energy budget is roughly 4kwh * 24 * 30 for the training. Not sure how this translates into CO2, but I guess an average flight will be worse than this. Such a model can handle all Buddhist languages, including Tibetan, Chinese, Sanskrit, in addition to Pali, so it servers multiple purposes at the same time.

Once that is done, inference is not visibly worse than hosting a website on a normal server (like SuttaCentral). We need about 8GB GPU memory to host the model, that’s easily done on a single processor that is shared with other tasks as well.

3 Likes
56

The thumb rule you can use for carbon intensity of power consumption is at most 1kgCO2 per kWh and at least 0.1 if you are in a 80% plus renewables grid (mostly where hydro is dominant).

Your footprint is not material if you compare to what some accept when flying here and there to attend retreats and seminars.

For further context, my electric kettle at home is rated 2,200 kW and I run it at least 6 times a day for 2-3 minutes. This adds to circa 3.3 kWh per month or 1,200 kWh per annum.

My airfryer is 1,600 kW and we use it 5 days a week for at least 30 minutes. You do the maths…

And to be clear, you dont need to rerun these trainings very often, right?

4 Likes
57

Yes! Thank you! I struggled so much before using that word! It kinda sorta pointed at the region of meaning I was going for, but it has become a bit of a weasly word that I fear sometimes obfuscates or sounds like an industry apologist.

And as you point out, do we have a good way to determine that good we are aligning with? Have we assembled the great scientific, religious, and artistic minds of our generation? Have we ensured a diversity of perspective and life experience? Or are we going with

:rofl:

1 Like
58

Thanks, good to know.

1 Like
59

To my knowledge the Berkeley AI compute infrastructure has a high share of renewables in the energy mix. If anybody is interested I can go and ask the actual numbers for this.

This long training run of one month was more an exception. When we are not doing that (which is most of the time), I train maybe for one day a week, try out some different data mixes etc. Every once in a while we train for a week,.

3 Likes
60

Let’s Make SuttaCentral 100% AI-free Forever

Got my vote.

1 Like
61

Indeed, may I second this.

Sebastian, may I circle back and follow up with this:

Indeed, yes, that’s the point. I think most people, and by most people I mean “me”, don’t really understand how interconnected these things are.

For example, you say:

Sure, I understand this. Still though, if I understand correctly, your main donor is Kurt, is that correct? Now, Kurt is a senior professor and a long term Buddhist. He also sold his self-driving DeepScale AI to Tesla. So your funding, if I am correct, has one degree of separation from Elon Musk.

My point is not that all AI projects are somehow conspiratorially a part of an agenda. It’s that there is a culture, and that culture has become inflated with vast sums of money, which attracts a certain kind of interest and ideology.

You’re saying that you don’t do the same thing as those guys. I get that. What I’m saying is that, regardless of what you say and do, the participation of prestigious institutions like Berkeley helps legitimize the very idea of AI, lending a virtuous halo to the work done by the OpenAIs and Googles.

3 Likes