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

Essays
21

Honestly, I’d prefer to just burn the entire idea of machine learning to the ground. Before they burn us.

I think accessibility is a major possible area where AI can potentially be of use. For many people, having realistic, or at least listenable, machine-generated human voices makes knowledge available in a way that is simply not possible elsewhere.

And there’s a difference between doing a personal project and sponsoring a web platform. I think it’s crucial that we keep ML sutta translations off the web as much as we can.

On the other hand, “want” is doing a lot of work here. Maybe just be content with English? Or do the translations yourself, the human way?

We’ll have to agree to disagree on that one.

5 Likes
23

Hi Bhante,

I share some of your concerns, but I am not in full agreement as to the non-utility of AI. Like you, I am worried that with generative AI, we are fast approaching the fruition of what was once just a conspiracy theory: Dead Internet Theory. For this reason, I think there is a moral imperative to label any AI-generated content on the Internet as such and to never pass off any AI-generated content as otherwise.

That being said, some of the more capable AI models (GPT-4 Turbo or Claude Opus) have proven exceptionally useful as tools to help with translation, word interpretation, and making for helping to make connections among the large Buddhist compendium. These AI brains are going to become even better very quickly, and people are going to use them for all sorts of tasks where we like it or not. To that extent, we would want them to be trained with the most reliable datasets we can, including SuttaCentral’s works. In my view, forbidding the use of SuttaCentral content as training data is shortsighted. It would be like banning the site from being indexed by search engines and, in my opinion, will only serve to limit the spread and influence of the Dhamma.

7 Likes
24

That’s certainly a minimum standard, and easy enough if you’re just hosting your own content. But the internet is built from scrapings, and nothing is lost quicker than context.

This is something that even Sam Altman agrees with, BTW. He said that it would have been good to ensure that all AI generated material is permanently labelled as such, but we’ve missed the boat and there simply is no way to do that.

Ahh, if only we had Xanadu!

They’re not brains.

Sure, valid opinion. Really, who knows? We can just just try to do what we think is right. Personally, I think it’s longsighted? I think in a few years, AI-free sites will be treasured like nuggets of poure gold in a swamp of mud.

2 Likes
25

AI is a tool. The result of the work of any tool is the result of the work of the person who used this tool, not the result of the work of the tool. With AI, as with any more advanced tool, it will be easier and faster for a human to perform some tasks, and the quality of such work will be more stable and on average higher. What will change is that the product, which previously required a lot of individual manual labour of a specialist, will be possible to create with the help of a new tool for a non-specialist, faster and of higher quality on average. The product made by man “manually”, as history tells us, will not lose its value, but on the contrary, its value will increase, but only if it has a higher quality or special characteristics that distinguish it from the mass product.

2 Likes
26

The teachings of the Buddha should indeed be preserved in their original essence. A person can hardly, if at all, understand Dhamma solely through books, let alone learning Dhamma through AI-generated text. I’ve realized that LLMs’ misunderstandings about Dhamma are largely due to the misinformation interpreted by people about Buddhism, and LLMs are simply trained on these data. Whether AI-generated content is beneficial depends on the user. For instance, as someone who is not adept at articulating everything, I find AI impressively intelligent in expressing my ideas to others more effectively, including thoughts on Buddhism. This cannot overlook the capability of understanding and interpreting the profundities of the scriptures. Nevertheless, the Buddha’s teachings aim solely to enlighten individuals, not for blind acceptance. The Buddha’s teachings are meant for everyone, and handling them a bit more wisely with AI should harm no one. Individuals must still filter information for themselves in their quest for truth. Of course, I respect and am grateful for Bhante’s translations, which have enlightened me further about Buddhism. Bhante reserves the right not to permit others to use your translations for purposes you deem inappropriate.
This comment of mine is also generated by AI.

1 Like
27

Sasha, and for all others on this thread, thanks for commenting! I do have many posts to come on this topic, so I won’t answer everything here.

But very briefly: tools shape us just as we shape tools.

hueso-2001-odisea-espacio

The quality of something being “impressive” exists in your mind, not in the machine. Intelligence, too, is something you have and machines do not; it is fooling you into thinking it is intelligent: that is its purpose.

4 Likes
28

So we have the AI saying, “I find AI impressively intelligent”—what to make out of this?

4 Likes
29

Yes, as always has been - the world will change and is changing.

AI is already here, and its presence is already changing the world, perhaps even more than the invention of the microprocessor. What’s the scary part? The inevitability of change? For a Buddhist?

AI is a tool. The danger is not in the tool, the danger is in the intentions of its use by a particular person. Fighting against AI or products made with AI is like fighting against looms instead of fighting for changes in social security and legislation in line with the emergence and development of labour automation.

The danger is in acting blindly out of fear of change and the unknown. The danger is in denying the inevitability of change. The danger is in not realising that it is impossible to achieve true happiness and security from unhappiness by changing the world. The danger is in confusing the happiness that can be achieved by changing the world with such true happiness.

1 Like
30

Certainly. When I talk about intelligence, I’m discussing my own personal experience. AI-generated text that hasn’t been interacted with by humans is just plain text. Though this is a personal viewpoint, it’s undeniable that systems like chat GPT are not without “intelligence”. For instance, when comparing GPT-4 with GPT-3, users can easily notice improvements in language understanding and argumentative skills, among many other aspects.

As for Bhante’s concern that AI attempts to trick us into thinking it’s smart, well, it indeed could be, right? If so, this naturally serves as an example of “intelligence”.

31

Right, it’s extremely hard to talk about these things without anthropomorphizing them.

If it were the case that the machines are trying to trick us, then yes, that would mean they are intelligent.

I think that, as Buddhists, the situation is a bit like the trickiness of using “self” in a conventional sense: you have to stay on your guard!

However, the point I was actually making was that the people who built the AI want to trick us. This is a deeply internalized assumption that, in my view, is inextricable from the very existence of the field. It is, after all, called “artifical intelligence”.

As noted before, though, I have much more to say. I’ll be posting an article a day, probably, for the next week.

5 Likes
32

Bhante, to clarify specifically, my comment was written in Vietnamese without accents and translated into English by Chat GPT 4.

2 Likes
33

Venerable Bhante, this may be a tool of Mara. I indeed have a certain curiosity and interest in it, but not overly afraid or placing too much faith in it.

Glad to hear more about Bhante’s viewpoint on this topic.

3 Likes
34

This is not what Bhante is saying or doing at all.

This is not analogous. A search engine (at least originally) points you to an actual source of information. These LLMs are chewing up and spitting out new information, often wrong, without even the ability to show it’s work or point to the original material.

5 Likes
35

Well. I don’t expect everyone to calculate the mean squared error of their data with an abacus, nor am I deeply against sorting cucumbers using computer vision.

From a SuttaCentral policy point of view we’re talking primarily about large language models in regards to translation and answering questions. This doesn’t sound like major request to our community.

Personally I think generative AI will hit the top of its hype cycle soon enough. There is no obvious pathway from LLMs to artificial general intelligence. As Charles Stross said at the beginning of this cycle:

I’m sure it’s just a coincidence that training neural networks and mining cryptocurrencies are both applications that benefit from very large arrays of GPUs.

4 Likes
36

The insights I’ve gained from working with EBT-DeepL are not from the AI. The insights have come from manually reviewing EBT-DeepL output and finding errors in translation, consulting ES/PT Pali dictionaries and human dictionaries.

I’d like to share one of those insights that shows the value of multi-language Dhamma study. This insight arose in the context of translating AN1.31-40, “Wild”. AN1.31-40 is also the sutta we should all be reading when discussing AI. To illustrate this, we have:

AN1.31:1.1: “Mendicants, I do not see a single thing that, when it’s not tamed, is so very harmful as the mind.
AN1.31:1.2: A wild mind is very harmful.”

AI today is wild. It is therefore very dangerous and harmful. I think we can all agree on this.

AN1.32:1.1: “Mendicants, I do not see a single thing that, when it is tamed, is so very beneficial as the mind.
AN1.32:1.2: A tamed mind is very beneficial.”

Some of us agree with the above and infer that AI-“mind” can be tamed to be helpful. Others here are adamantly against AI in the sense of “AI is the evil taint that corrupts. AI is a tool for Mara that cannot be cleaned and should be burned to the ground.” Some of us think that AI isn’t really a mind at all even though AI is currently fooling people into thinking it is human (witness AI-bots making money with fake human avatars).

The thing is…we cannot escape AI. Every single phone you hold in your hand was made with AI. AI designs chips better than people. AI finds workable solutions faster than people. Now please all throw away and burn your phones.

The danger of AI is very real. We are headed towards an illiterate world fed by AI-pablum summoned on demand by voice command to direct AI factories in the automated depletion of the world’s resources.

What can we do?

I see few choices here:

  1. We can ignore AI and stick with “human-only”.
  2. We can try to get rid of it.
  3. We can look AI in the eye and tame it.

Realistically, we will probably have to do all of these. But let’s continue with AN1.31-40…

AN1.39:1.2: An untamed, unguarded, unprotected, and unrestrained mind is very harmful.”

The Buddha teaches us this danger.

The Buddha also teaches us:

AN1.40:1.2: A tamed, guarded, protected, and restrained mind is very beneficial.”

Giving the above to DeepL results in the rather clumsy:

Uma mente indomada, desprotegida, sem proteção e sem controlo é muito prejudicial.

Well “indomada” is “untamed”, “desprotegida” and “sem proteção” both mean “unprotected”, and “sem controlo” is an odd way of saying “unrestrained”. This is clumsy and not good. We have to fix this!

Consulting the Pali ES/PT dictionaries, we see something interesting. We see “vigilado”:

rakkhita protegido; guardado; vigilado; conservado.

And we realize that ES/PT both have a verb for the EN word “vigilance”. This is quite remarkable. It is remarkable because EN has lost the ability to say “watch with vigilance” as its own verb. Both ES and PT have the verb form of “vigilance”. :open_mouth:

Is this a trivial detail? I think not. The verb form of “vigilance” is quite important. To explain this, I will need an example. The example is heating milk for cheese. No, this is not a silly example. Pay attention carefully here.

When making cheese, one scalds the milk to transform the protein chains that make cheese cheese. When milk on the stove is left unattended, it will boil over and spread in a big mess all over the kitchen. Boiled milk will even keep spilling over when the heat is turned off. It’s a really really big mess. Because of this, heated milk needs to be watched with vigilance to take the right action at the right time.

The reason that milk is important here is not because we all need to make cheese. The reason milk is important here is that milk is a simile for the mind. When the mind is aroused it can boil over and create a really really big mess. Because of that, we need to watch the mind with vigilance so that we can take the right action at the right time. If we never turn on the stove, we never get cheese. If we turn on the stove and fall asleep, we get a really really big mess.

Because of this, it makes sense to use “vigiado” as one of the words used in translations of “tamed, guarded, protected, and restrained”. This can be done in ES/PT but not in EN itself, which has no verb form for vigilance. If you don’t believe this, please try to use any one of the EN words “tamed, guarded, protected, and restrained” with regard to the milk simile. Vigilance is required for milk, but there is no EN verb. However, there are ES and PT verbs for this one concept that applies well to the wild mind.

We all need to address AI as it infiltrates our lives and heats up our world (data centers under the ocean?). We all need to be vigilant. I really wish I could use the non-existent “vigilate” word here in the imperative. But we haven’t done so in EN since 1770

EN is a view. It is not the only view. That is why translating the Dhamma is important to me. I hear things in ES/PT that I cannot hear in EN. Any one language is a shadow of the Dhamma, a flat projection of the Dhamma.

For those of with more life ahead, please do look for and wait for human translations of the Dhamma. For those of you like me with much less time left, please forgive my Quixotic attempt to tame AI.

7 Likes
37

Is AI a tool?

Well, yes, in one sense of the word tool. The issue is what connotations you carry when you say the word “tool.” Do you picture something innocuous like a screwdriver? Or do you include in your mental map of tools that the Atomic Bomb is also a tool?

2 Likes
38

This is true. There is a substantial body of argument that shows that MLLs are linear/logarithmic, where the ability scales linearly and the data scales logarithmically. Available data sources peaked out in 2021, so now they are digging ever deeper to find more data. That means they are using more and more dodgy sources. But i think it’s only a matter of time, if not already, that they’ll be tapping real-time data, like video calls, web cams, biometrics from watches, and ultimately directly from a brain interface (i.e. Neuralink, etc.).

Depends on the kind of screwdriver, I guess.

0831068a8bf0116fea105749c2746b1b--sonic-screwdriver-doctor-who-stuff
0831068a8bf0116fea105749c2746b1b--sonic-screwdriver-doctor-who-stuff736×552 58.6 KB

The power of the tool is one thing. But the point I want to make is that a tool does not exist objectively: it is a product of desire. An Atomic Bomb doesn’t just happen to exist as a world-annihilating possibility, it exists because people wanted to to build it.

And once a tool exists, it opens up a certain range of possibilities, which narrows attention in a certain way, eliminating other possibilities. As they say, if your only tool is a hammer, all the world looks like nails. Today, it seems to me, the entire world has gone cuckoo bananas for a shiny hammer, and everyone is rushing around looking for nails to hit.

4 Likes
39

About 30 years ago or so while in graduate school, I freelanced as a fixer of software translations. It was very enjoyable work that could be hilarious at times. Fast forward 30 years and my first known contact with AI generated Suttas was incredible. Machine translation has come a long, long way.

But I agree with this assessment. I’ll grieve the decision to exclude AI translation from SC a little since I’m a little involved in what I only learned today was AI translation (I don’t keep up with tech) :slightly_smiling_face: and the translations have been surprisingly good. But after reading Bhante’s essay (thank you Bhante) and this post I agree with the reasons. I grieve for those who will be waiting for the benefit of reading multiple translations in their language beside the original Pali. Conditions …. I’ll keep moving along with manual translations which do offer a lot of rewards for the translator that I noticed AI didn’t offer.

5 Likes
40

I don’t think Bhante is proposing “escaping AI”. And just because something is useful in one domain doesn’t mean it is equally useful or appropriate in other domains.

It seems that many people are arguing some form of “this is reality, don’t try to avoid it.” This makes no sense to me. Why is something necessarily good just because its new?

Anyone can use the existing tools to translate whatever they like. What Bhante is proposing is that SuttaCentral not publish machine translations. He is also asking that people not use his translations as fodder for the LLMs, but I believe the horse is already out of the barn on that one.

2 Likes
41

Your example of Portuguese translation is not adequate.

The best translation is domada as Michael Beisert translation has it already.

It would be really good if you refrain from making posts involving Portuguese without knowing the language.

And, frankly, these hypothetical insights you’re looking could be more reliably acquired by learning Pali instead, or instead, sticking to translations made by those who do so.

That’s why the translations I do are based mostly on what venerable @sujato did and only partially informed by consideration of synonyms that best match the nuances and greater indoeuropean parallels between Pali and Portuguese.

I trust venerable Sujato best efforts to know Pali and English well enough to have rendered to a modern style what he understood. That informs my learning of the Pali in context and therefore my choice of translations.

I do use Deepl for testing synonyms or alternative phrasing when I get stuck with getting what is meant but forgetting how that is actually said in my mother language (I have been outside Brazil for 10 years know).

4 Likes