AI-16: What to do?

Essays
41

I think this is an amazing perspective so overlooked.

Just as printing press was a disruptive information technology that was first pioneered and used by less than honorable people, so is AI. It is our impetus to use AI wisely, to show what this amazing approach can help us with (again, already done amazingly by SCVoice, Buddha Nexus, and so on).

I try to be cordial and friendly with LLMs, not even caring if they’re capable of understanding in the same way as I do or not. I treat my items, pets, friends, family with same respect and diligence, AI is no different.

It is much more practical and skillful to figure out how to wield ML algorithms to shape a better future, than to bury our head in the sand and hope it goes away.

Seemingly ironic, it’s the spearheads of AI like Sam Altman and Elon Musk who scaremonger against AI like no other, casually talking about destruction of human race and not.

There’s probably a good reason AI overlords wanting us to be scared of AI. After all, afraid populations are easier to control.

Starting off “AI works great for killing people” is not a very nuanced introduction now is it.

It’s kind of like saying “Planes work great for killing people” when they were first invented, because that’s what we used them for mostly.

There’s a growing violent anti-AI sentiment bubbling, some people are having their lives threatened for drawing a few pictures with Midjourney. Such perspectives from our religious institutions lacking nuance and substance (even when they do employ Machine Learning algos in various projects they do endorse) is at best, unsympathetic and unskillful.

There are people working in CS that are working with big data and ML that have nothing to do with the things Bhante’s being critical of, but they’re all labeled under AI and so everyone’s under fire. Meteorology. Infrastructure. Revival of ancient languages. Healthcare. But they’re all “AI” so they’re all great for killing humans?

This topic deserves a greater nuance from our spiritual guides.

1 Like
42

One should be called an AI overlord if one had true control over AI. ASI is beyond human control.

If any humans wanted to be the overlord of humanity, it is based on Plato’s notion of philosopher king. Those who knows the danger of future are best to hold power to steer humanity away from such danger. It’s not so much about greed for money or power for its own sake, but to avoid humanity screwing up ASI.

We basically have just one chance to get this right, or else humanity may not get the chance to tell ASI what we want it to do with us. So it’s good that people are concerned and move themselves to help avoid the worst timeline. Although if their morality side (psychopath?) is a concern, then one of our lay Buddhists should step up and be the new leader and influencer in this AI field.

Practically speaking, AI will cause job loss and is already doing that. Especially with Sora, text to video generator AI, entire video making industry will lay off millions of humans.

This is a golden opportunity for Buddhism. We should tailor talks and advices to the public facing unemployment and being unemployable, to learn how to fill their time constructively.

Just like the start of internet allows for giants like google to spring up, so too this start of AI would bring different power shift dynamics to spring up. It’s a beginning of new things and if Buddhism gets in the wave right, we could make great contribution onto the world’s cultural scene and be more relevant to the world at large, not just to Buddhists. Any ideas?

2 Likes
43

I’m already using ChatGPT to translate Gandhari to Pāli, for example (Unsurprisingly, ChatGPT is amazing with grammar, seeing that it’s primary function). This could be extended greater. We could cross-translate so many material and analyze texts like never before. This would dawn a new age in historical-critical text analysis.

Furthermore, NikayaBots could serve as impartial Sutta search functions, for example “Give me all the suttas about meditation” and voila. ChatGPT is notoriously terrible when asked about suttas but LLM modules can be tailored specifically to answer on a specific set of texts.

DhammaBot is a bit more problematic, I think. Even if it were tailored only to give suttas (or a similar grammar / wording) to answer general spiritual questions, there’s all sorts of legality and responsibility issues that are hard to address properly. MeditationBot would fall under this as well, I think. Potentially incredible, but potentially terrible.

First step to all these is to get suttas in the database though. Once the neural network is intimate with the body of texts as to reliably pull out (or at least hyperlink) suttas, then it’s just a matter of tailoring it to end-user in a responsible way.

44

Buddhist AI bot already exist.

Edit: Sorry, I mistaken this with r/Buddhism. Here’s the posts:

They give different answers.

The generic Buddhism one gives deep Jhāna and something after parinibbāna, the EBT ones give Jhāna lite and nothing after parinibbāna.

1 Like
45

It sounds to me like you haven’t read this series of essays on AI by Venerable Sujato and the subsequent discussions about using ChatGPT and LLMs to translate and handle Buddhist scripture. You might have a different perspective if you do.

46

I have read all 16 articles. I respectfully disagree with Bhante. Again, to me it reads a lot like how Christian priests thought printing press would be a disaster to their faith because it would allow Christ’s words to proliferate without licensed church members to expound on them.

2 Likes
47

I made that argument a while ago: AI-1: Let’s Make SuttaCentral 100% AI-free Forever - #114 by SebastianN
Alas, this forum is not the best place to discuss positive outlooks on machine learning approaches to Buddhist material. At dharmamitra.org we develop these applications, and a search engine that does what you described is going to launch in the coming months!
:slight_smile:

1 Like
48

Because it’s an important topic. Shouldn’t we be investigating things deeply?

If people are posting things that aren’t true, then by all means I think it’s legit to challenge those things specifically. But I don’t see the broad generalizations to be very helpful to the discussion.

It’s easy to make a caricature of both sides of the discussion.

I’m personally sorry you feel this way. From my observation there is abundant support in discussions of AI, LLM etc. from many posters. But there is indeed going to be push back on the hype and hope around AI. So to me that makes it a very good place to discuss these things.

2 Likes
49

I think Bhante Sujato’s completely unnuanced opinion (he went on to say he wished he could destroy all ML applications everywhere in one of the replies) makes it seem like this is the stance official and there’s no discussion appreciated on the matter (except with regards on how to further the butlerian jihad).

There’s a saying in here “If imam farts, community shats”. So it goes. I personally discarded many long posts replying to Bhante’s concerns one by one, because I thought there’s just no point. It doesn’t seem like a “Let’s discuss ups and downs of AI and how to navigate it skillfully” atmosphere, it’s more like "“AI works great for killing people, let’s discuss how to stop it”.

I would love to be wrong on this impression.

This is amazing, is there going to be Pāli support as well?

I would love to talk more about your projects.

50

Actually, it happened that people reached out to me in the course of this debate telling me how much they agree with my stance but don’t want to raise their opinion in public due to the harsh respones that positive replies to AI will cause. Now I am a nobody here, an external entity who is amazed by SuttaCentrals achievements but is not spiritually or emotionally invested in the community save for a number of important individuals that I grew close to over the years, independent from this discussion board. So I can say what I want here and don’t have to worry too much, but I totally see the problem that other people might end up in who care more about how they are perceived on this discussion board than I do.

4 Likes
51

But this is obviously not true. Have any of your long, pro AI (etc) been banned? There have been countless posts of people disagreeing with Bhante’s opinions.

1 Like
52

This is true. I still feel uneasy and weird. And I’m guessing (from Sebastian’s post it seems) so do other people. Perceptions are a funny thing.

Then again, I remember in one of the posts, one user was trying to show something with ML translations to make a point and Bhante in no uncertain terms asked him to refrain from posting AI generated garbage. So, people are not banned perhaps, but when a venerable is scolding you for trying to make a point it points in a certain direction.

Anyway, perhaps we should move on from this meta-discussion, and move on to how we could benefit from accessing SC database for good dhamma related projects then. That’d be a positive direction for a discussion.

1 Like
53

I don’t deny there are power dynamics at play. But for feeling uneasy, you seem to have posted a lot (which I am not criticizing or discouraging, just saying that it don’t appear to have stopped you from posting)

And yes, Bhante did ask that an LLM translation of the commentaries be removed from a public post.

The problem with MT of ancient texts is not that it gives new people access to something (like the printing press did) but it gives people inaccurate access to something, which is (in some people’s opinion) worse than no access. And because these LLMs are being built/trained on material from the internet, posting AI generated text creates a feedback loop where new models will be trained on the content generated by the old models.

I really don’t see how people could accuse Bhante Sujato or anyone involved with SC or the forum of being against giving people access to things. That’s their whole point.

The problem with someone posting a MT of the commentaries as an example of how great LLM translation is, is that most people are unqualified to judge it’s accuracy. The only thing we are qualified to do is say, “Wow, that sounds like natural, human generated English.” And for me, that’s the crux of the issue.

Technology that can produce accurate sounding translation is not identical with technology that can produce actually accurate translation.

3 Likes
54

The argument was in one of the earlier posts that unethical companies would and do use this data, and conscientious buddhists are sitting in the sidelines. In due time, such applications are going to go live, with or without permission, and without any oversight.

Someone like Bhante Sujato for example is in a unique position to oversee a language model fine tuned to Dhamma translations. Both for his proficiency and ethical considerations. Our senior monks can spearhead next generation of applications for textual analysis and Pāli tutelage. In their absence, only low quality, irresponsible programs will emerge, and will result in more confusion.

My argument is, someone is going to do it, it would be better our leaders than otherwise.

55

I guess its not necessary to ban things if you foster a discussion atmosphere that dissuades people from speaking up.

But that’s precisely what BJ asked for – not to do any AI applications on top of his translations. For team members on our end, this meant quite some work the last months that we would have liked to spend on other things. It caused overhead, both time-wise and also resource-wise as we had to phase out earlier models that included certain sections of SC data in the training and retrain these. Thankfully, other relevant translators are not as harsh in their stance as BS is, so right now he is the only individual that demanded this from us. I am happy when we will release our next updated model without BS’s data.

1 Like
56

Of course we can’t know what you are not posting. But as I said, I see many, many people posting in opposition to Bhante’s position.

Any way, I’ll try to refrain from meta conversation.

And I guess I would say it’s better for them to spend time doing actual translation. It’s not their job to clean up other people’s messes.

On the Pali side, it’s really just the commentaries that remain without translations. And the Pali in the root texts is different from the Pali of the Suttas. So I don’t see how an LLM trained on the suttas will produce something good on the commentaries.

One problem, as I see it, is that we now have people who are not experts in translation of Buddhist texts creating tools to translate Buddhist texts. With previous technology (as I understand NLP) the people creating the tools had to be experts in the languages because they were coding in the vocabulary and the grammar. This is not the case with the LLM tools.

Just a basic look at the footnotes by translators (I’m thinking of Bhante Bodhi, Ajahn Thanissaro, Bhante Sujato, Bhante Ānandajoti) will show how difficult it is to translate some passages in the root texts. This is not because they are lacking in computational power. It’s because translation is difficult and translating religious texts is even more difficult.

As I understand the workings of LLMs, they are good at generating accurate sounding translations based on statistical probability determined by the existing texts that they are trained on. I just don’t see how people can consider this a great way to create quality translations of difficult, yet to be translated texts.

1 Like
57

I wonder why you feel the need to narrow down the possible application space of AI technology to the translation of difficult commentaries here, that is not what dogen was talking about.
I agree that LLM technology is unlikely to be very useful in the precise schema that you describe. But I also don’t see anybody suggesting to do what you are rightfully questioning here.
What we at dharmamitra.org do is to create better search systems based on machine learning algorithms. And that is something that very well can benefit somebody who is translating difficult commentaries, as they embedd material from other parts of the canon, perhaps preserved even in different languages.
LLMs can come in handy in different stages of the translation as well. For me, that would be mostly for post-editing draft translations that I create myself, simply because English is not my mother language and LLMs are better at producing fluent English than I am. So again, its not about replacing something that can’t be replaced – I am not claiming that LLMs are solving translation, or philology, or whatever else makes research difficult. But used in the correct way, they can be a great aid. I understand that the current leadership of SC doesn’t want anything to do with that, which is sad but we move on.

1 Like
58

Just to correct some error. I made a mistake.

1 Like
59

This assertion needs citations to back its claim and thus we need to discuss actual examples. Using ChatGPT to translate suttas, I don’t see this to be the case. In almost every case, the suttas are translated in a pretty consistent condition compared to Bhante Sujato’s. We can go over examples if you want.

And I should remind you, ChatGPT is not trained specifically on Pāli, but it just picked it up amidst all the data, and this is the very bare rough result that I suggest we could work on, and it’s already working great for the most part (for the rest, please read on):

I’m talking about actually overviewing the bot itself, not the output.

A bot can analyze extant suttas, bring up all common phrases and expressions, fine tuned to understand how sometimes some words are used for x but same word means y in other context. (it already does this to a large extent) And then this bot can be tuned to hundreds of languages. Even this fine tuning process can be repeated (now automatically) for those hundreds of languages. And with a negligible amount of work (about 1 day for most important 100 phrasings or so. perhaps a week for entire catalogue) fine tuned to perfection for those languages.

This is not about just translating commentaries. This is about bringing dhamma to untold amount of languages as possible. We could bring entire Pāli canon to say, Turkish, in less than a month’s work (and that’s with immense headspace taking into consideration that monastics could perhaps devote an hour of their day or so of their time at max). Once this process is established, it would only take a skilled practitioner in X language to go over top 100-500 phrasings and manually tune the bot to those parameters, and we can have a new set of suttas in a week in any new language.

At this point it’s prudent to ask, is it not irresponsible to not wield such a tool for proliferation of dhamma. As I said, this is where printing press analogy comes into play. LLMs aren’t just dumb tools we have no control over. They can be fine tuned to painstaking precisions, something we all somehow forget in these conversations.

Saying LLM produce accurate sounding translations is a semantic error. LLMs produce accurate translations, based on their training data. If the training data is inconsistent and haphazard, then the result is likewise. This is why early transformer models resulted in arcane gibberish that reflected the chaos of our collective consciousness. And this fine tuning process is how ChatGPT can pass the medical exams with great proficiency.

And as Sebastian said, this is just the tip of the iceberg.

60

I didn’t really want to wade into the discussion, but maybe it’s worthwhile putting in a different perspective.

First of all, I am not anti-AI in general, although I have strong concerns that mirror @sujato so I have read most of his articles with interest and agree with a lot of his views.

I first used LLMs about 5 years ago (I wrote code in Python that referenced the models, and deployed them on virtual machines that cost hundreds/thousands of dollars per hour), so I am not unfamiliar with them. They have advanced a lot since then. I have just purchased a top of the line MacBook Pro with 128GB memory so that I can do some transformations using LLM, and I have already determined for some use cases, they do a better job than what I could do, so they can be very useful.

However, @Snowbird is right - at the end of the day, these models are stochastic, and doing transformations based on statistical probability. That is just how the models work. The fact that they give an illusion of intelligence is just that - an illusion (although some AI scientists claim the higher levels of the model emulate inference and higher order thinking processes). However, statistical text generation and transformation can produce useful outputs, no doubt about that.

In terms of using them for analysing, translating or interpreting Buddhist texts, I would not do that.

From my perspective, I believe that the existing suttas are corrupt and no longer faithfully represent the Buddha’s teaching. I am not going to defend this, I have written an article about it in the past, and anyway, I am not the only one who has made this observation. Let’s just say it’s a hypothesis and leave it at that.

I also believe existing translations done by humans are inaccurate and further distort the teachings. I don’t mean to disrespect @sujato and others, I believe they translated to the best of their abilities and have tried to be as accurate as possible. But, there are plenty of places where I don’t agree with existing translations. I am not saying I am better, but there is room for doubt and alternative interpretations are possible.

I believe studying the suttas require a highly attentive mind, critical thinking, direct experience, and a willingness to question established interpretations. Otherwise, we may never reach the goal that we seek.

Because of this, I believe using AI will just add noise to the process, rather than provide clarity or new insights. I don’t mean that there are no useful applications. For example, using AI to generate code that analyses the suttas to derive patterns may be useful, because AI is not directly involved in the analysis, and it’s up to a human to interpret the analysis.

At the end of the day, we are all individually responsible for reaching the soteriological goal. I also believe what the Buddha taught is quite straightforward and should be graspable by anyone with moderate intelligence. To directly experience what he has taught requires some effort though. But still, thousands of people did it during his lifetime and I believe it is still possible today, provided one is dedicated enough and willing to do the critical analysis to separate what he taught from all the corruption.

I wish you the best in your own journey towards this, as well as anyone else reading this.

3 Likes