AI-3: There is no road from here to there

Essays
21

Haha, yes it sort of does! But they did eventually find other jobs (except for Hinton), and also spent a lot of time teaching others so they are generally well regarded by the A Icommunity. My point is that masses look up to them for trying to do the right thing so there must be a genuine drive in a lot of people to use this technology for good. Yann Le Cunn took another approach, he actually advocates open source models, because he thinks that monopoly over these technologies is more dangerous (although he is chief AI at Meta, so it’s better to take this with a grain of salt).

Indeed #nomoat but it’s entirely possible, that in the end only just a few large players will be left, similiarly to what happened to the internet.
<5B SOTA sounds amazing!

1 Like
22

Also an interesting point. There’s clearly a massive ethical component to the motivations of many people in the industry. Now, that’s not at all unproblematic, as the specific kinds of ethics they are drawn to are often dangerous (eg. longtermism).

It’s hard to judge, because a flimsy veil of ethics hiding a beating heart of rapacity is pretty much what capitalism does. My sense is that the field attracts a lot of idealists who are ultimately quite naive about where their work is heading.

I’ve no doubt people an build good things with it. Just 
 really hard to escape the feeling that a Butlerian Jihad would be pretty sweet about now.

1 Like
23

I think he’s doing two things here, first supporting AI research outside of Meta, which is of course a good thing. I do believe the term 'open source to misplaced however, as we don’t have the actual data or training routines, only the model weights. The other thing is to saturize the market with models that are ‘just good enough’ to get the job done. For 98% of the applications you actually don’t need the big Claude/GPT4 models, and a startup can build on top of LLAMA and be competitive easily. So there is also a strategic element for Meta for release these models.

I am actually not so surprised. Buddhist data is a lot to read for a human in a lifetime, but not a lot for machine learning standards. There must be some sort of saturation point, a finite amount of information can only be stored efficiently in a finite number of parameters, and once that point is reached, more parameters won’t help.

1 Like
24

Perhaps, but certainly achievable. If @sujato can translate the entire Sutta pitaka in a matter of a few years, it’s not beyond the scope of an individual to be reasonably familiar with large sections of it.

Surely it is the most important task one can possibly achieve in this lifetime, and may save countless future lifetimes.

At one time I thought this was unachievable, I was too busy, I don’t know Pali, there is too much to read, it is impossible to understand etc. etc. Excuses are easy to draft.

I didn’t seriously start reading the suttas until about a year ago. I managed to learn Pali in six months, and have spent at least six months reading suttas as well as various books, and so far it has been the best investment of my time I can possibly spend before I die.

I don’t need machine translation because it is just simply easier to read in Pali, then I don’t have to worry whether the translation is correct or not. And certainly easier than spending months fine tuning an LLM, doing cross checking and validation, etc.

I am not saying that AI is useless. But often, the effort to get it right in AI is not much less than relying on my imperfect mind. I was actually contemplating using AI to grade student assignments today, and went some way towards feeding it the assignment brief and the rubric, but then I realised I am better off just marking them the traditional way and it doesn’t take that much extra effort.

Perhaps the models will get better. They are certainly better than the GPT-2 type models I was deploying 5 years ago (a paid consulting assignment) but in some respects they haven’t improved that much. They just lie better and the output is less stilted and artificial but everything still requires validation and fact checking.

It’s a bit like self-driving. We are always promised it’s around the corner, yet there is always one more problem to solve, and ultimately a fully functional implementation needs to be aware of ethics and balancing between different degrees of harm. In other words, we need Asimov’s Three Laws of Robotics, but it is hard to implement this as a meaningful guardrail, but I live in hope.

1 Like
25

If you limit yourself to Pali you might be right, but the universe of Buddhist texts has a lot of material in different languages, and I yet have to see a human being that is 1) reading all of these fluently and 2) has the time to read all of these texts. Yes, some people exists who read some of these languages fluently, but they only manage that on restricted domains, like Analayo on Early Buddhist material, Poussin for Abhidharma, Schmithausen for Yogācāra etc. And these are the exceptionally skilled people that have achieved what the vast majority of “normal-gifted” people can’t.
The difference between self-driving cars and LLMs is that we see the latter now deployed in useful applications, and in combination with external knowledge sources they get certain jobs done. I am not doubting that they are overhyped and failing in some cases.
And to be fair, while self driving cars have failed on the ‘problem solved’ level, a lot of the technology has trickled down into normal cars and made them significantly safer.

1 Like
26

Yes, until I can read them while traveling in a self driving car, I’m not interested in self translating suttas. :wink:

2 Likes
27

That is true, but there is a lot of repetition in Buddhist texts, nor are they completely consistent. Whilst there are some in this forum who are doing comparative analyses across different canons, I don’t think it is necessarily the end goal for everyone, nor is it desirable.

At the end of the day, we need to reflect on the purpose of studying these texts. Is it for academic completeness, or to realise a soteriological goal? Arguably, if it is for a soteriological goal, one could say one should read the minimum number of texts to reach that goal, and it could be just one sutta (if one is “lucky” enough, or perhaps in the right disposition to realise the goal).

After one has achieved the soteriological goal, there is no need to read the remaining suttas. They can be disregarded, in the vein of the simile of the raft.

1 Like
28

This sounds remarkably similar to what someone told me in a Goenka Vipassana center while I was doing long term Dhamma servide there in 2016. Another jumped in and even questioned whether it is right livelihood to study multiple different Buddhist languages and traditions at the same time, as this can only lead to confusion and perhaps worse, distortion of the teaching in case such academic work gets published.
It was one of the final nails in the coffin for me that drove me to disengage with that community. But to be fair, I had some 30+ pleasant retreat experiences over almost 10 years.

For me, studying the texts from different traditions and eras of time has a soteriological effect on its own, but if somebody is happy with just reading one Sutta all the best to them, I am certainly not opposed to that.

That drives me back to the initial motivation of creating MT systems for these languages: The more tools a curious person has in the toolbox to interact with these texts, the better it is. Being able to verify something in a primary source yourself is extremely satisfying, and the years of studying Pali, Sanskrit, Chine, and Tibetan at the university certainly helped me to increase the trust in the teaching.

2 Likes
29

That is certainly a risk, but possibly overplayed.

The following book directly compares the relevant suttas in SN/SA - although it is clear there are discrepancies - as a whole, the teachings are consistent although there may be a change in emphasis and different formulations in some of them.

I would say it is possible to reach the soteriological goal no matter which canon one chooses to learn from. Again, it’s probably the most efficient as well as the most satisfying to read the suttas in the language they are presented in, rather than in a translated form. Eliminating a possible layer of misinterpretation is of course the primary benefit, a reduction in the number of symbolic representations is another.

But regardless, once one has reached the soteriological goal, one realises all the suttas essentially say the same thing, over and over again in various combinations. The Buddha’s teachings are relatively straightforward (despite the tendency for some to over interpret them) and can be grasped by someone suitably inclined relatively quickly, there are many cases documented in the suttas of people achieving realisation from a minimum amount of instruction. There can be no “confusion” or “distortion” once one can directly experience what he has taught.

I certainly wish you the best in your endeavours and your approach. I don’t disagree or oppose them, although I do empathise with @sujato 's thoughts on this matter and ultimately I respect that as a primary founder and contributor to SuttaCentral he has the right to request how the content should and shouldn’t be used, and we as moral beings have the ability to honour and respect his request.

2 Likes
30

Which I do. I wonder how people would argue on this forum around these questions if Bhante was not so opposed to AI technology. Just ask the question yourself, “how would I argue about this if Bhante was in favor of the technology, would I still hold the same opinion?” and let the answer speak for itself.

3 Likes
31

This is an excellent question and point of reflection.

I was actually opposed to MT being included on the site, including one that Bhante had already approved (which is now being removed).

And based on my experience with ChatGpt (which I know is not the latest technology) I would not be in favour of the SC data being fed into LLMs. Having used Google Translate for Sinhala, I am also very skeptical of these tools working for minority languages, especially those with specialized terms that may or may not already be established in the target languages.

I actually don’t share Bhante’s sentiments on many things related to AI, but I am indeed in agreement with his feelings about content on SC.

4 Likes
32

Yes, I think even the AI stans would agree that Machine-generated Translations (MT) shouldn’t be hosted statically. If the technology is really improving so fast, wouldn’t you always want it to generate the latest translation (rather than reading one generated by an old model)?

3 Likes
33

Yes I’d actually agree that hosting MTs on a static website is not a good idea unless there is a very good reason for that (i.e. tight integration into a search engine, or the MT performance has flattened out to such a degree that further improvements are not noticable anymore). While post-editing improves translation speed a lot, and having a translation tool available is generally great for the productivity of the users, static, unedited MT on a website isn’t nice.
I guess that’s the fine difference between “having a tool that is useful” vs. “claiming the problem is solved”. Machine translation has certainly passed the first bar, but the second is far far away.

3 Likes
34

This is a valid point, I am only here because I am interested what he and @Brahmali would have to say on these topics. This isn’t the first time they went against mainstream ideas and they were usually right so I am interested in what they have to say on this one. Even if they are not that tech-savy I can totally imagine scenairos where I am unable to see connections due to my defliements that they do.

I think what you do, the reason why you do it and the amount of progress you managed to achieve are all amazing. If the next step after the internet / youtube is LLMs, then it is an absolute must that Buddhism represents itself in this field. On the other hand, I can totally accept if monks feel like avoiding the technology altogether to ensure that Suttas are preserved 100% correctly without adding any ambiguity, although I feel all of AI’s negative results are exaggerated in order to emphasise this one point.
I would probably have been very upset had I spent several years working on a solution like that just to see it being shut down. One reason I’ve been here arguing for days is because I wan’t to figure out if it is indeed necessarry to give up all of my work in this field. So far I think it’s not, there are genuinly good use cases, but I am more aware of how closely even the good use-cases are to being used for war efforts.

Also: one thing I have not seen being brought up is that the first western translations had a lot of inaccuracies too, that western Buddhism still suffers from (we still use the word concentrate in Hungarian Vipassana meditation retreats as far as I know).

3 Likes
35

Thank you Richard. For me it is also interesting to see how people react to this technology on this board, and I guess its a rather unique blend of people and opinions.
To be fair, there are some very weird “translations” into Chinese such as the Sutra of the fourty-two sections, and I am pretty sure that even our modest engine would be able to render Buddhist thought more accurately into Chinese than what happened there (if it ever was a translation at all).
I guess its also important to note that the strong opinions on this board are not necessarily what I’d consider “Buddhist mainstream” even when restricting to Early Buddhism or “the southern transmission” as people in East Asia might call it. Outside of this forum there is a diversity of opinions, and the Bhikkhus I could talk with so far have a rather positive view on our work, not even to speak of the Tibetan tradition where this technology is very welcomed and officially supported.
Perhaps Buddhism needs a Yudkowsky, and if Bhante wants to have that hat on, why not. Its important to listen to these opinions, but regardless of that I think there is a very genuine opportunity for productive research here, and in many ways Buddhism is a very interesting case for the application of machine learning algorithms, something we already saw at buddhanexus.net back in 2019.
Its not all bad I guess, and I remain rather positive about the future.

36

But what if being active on youtube meant that Google wouldn’t let you post videos, but instead sucked them in, ground them up and spat back out something very different? Would you still want monks to be represented there? Because these LLMs are not just another content delivery method. At least not anything analogous to anything that has come before.

There is actually a diversity of opinions here on the forum, not sure if you were implying the opposite.

I agree with you very much on this point. One of my biggest concerns is a feedback loop where LLMs are being trained on MTs And since there is no way to stop this from happening, the only option I see is to not publish them.

Not everyone agrees with that. Two of the most active translators for SC have said the opposite. I’m sure for low stakes things (i.e. not suttas/sutras) it can save a professional translator time.

2 Likes
37

This is a very good point.

However, there is no gurantee that a Google search result, a Wikipedia article on Buddhism or a Facebook post of an image of a sunset with some fake Buddha quote on it would be true either. If you think of LLMs as a replacement for search engines that may run locally that users would rely on to ask day to day questions (aka. do search queries via natural language), then I would say it’s still better to have some sort of representation of Buddhism available on them. Most people would just ask basic questions anyway. So if a model that was fine-tuned to be somewhat morally aligned would come up with an ethically okay answer that does not contradict Buddhism, while still giving an URL to BSWA or SuttaCentral where the user can do their own research would still be more useful than not having this option at all.

Or at least that’s what I thought a year ago, now I am not that sure. It wasn’t so clear back than that the models would keep making stuff up in such magnitude, at the time it looked like we would only need more training data / better reinforcement learning techniques to mitigate these issues. :person_shrugging:

38

I respect your question and the line of thinking. For me, perusing Discuss & Discover – the highlight of which are the essays by the monastics – is part of my Buddhist practice for now. It’s where I’m learning dhamma through the lens of the pāli suttas in pāli. Particularly from monastics and lay-people Buddhist SMEs.

Peeling this onion a bit more, part of the practice is staying cognizant of how I’m responding to what I’m reading. It’s been the case ever sense I started practicing with Discuss & Discover about four years ago (I was trolling at first).

That aspect of cognizance is continuous and takes effort. I note, for example, how I’m relating to the material vis-a-vis who’s posting it. I check for tendencies not to think for myself (hence be deluded, among other things).

Importantly, I trust the majority of Discuss & Discover readers/posters to do the same. If they are practicing the noble 8fold path, I assume they are practicing this.

All to say that I appreciate where you’re coming from re: your statement (at least, I think I do). But I have confidence in those discussing anything on this forum that they are practicing like this.

:blush: :pray:t3:

2 Likes
39

To me as an outsider, that’s not really how it feels. My impression is that the ‘no no’ fraction here is dominating, with some group dynamics at play, and people in favor of the technology can now decide if they want to argue, or just withdraw their material from SC and move on. Maybe I am getting it totally wrong and things in reality are very different, which I would of course welcome.

That sounds a bit funny to me actually. For “low stake things” such as finance, medical domain etc. post editing is adapted on industrial scale. Of course I agree that Sutta translation is something that should be done with care, but I fail to see a clear characteristic in this setting that makes it, say, less risky to use post editing for large financial transactions in economy than for translating the word of the Buddha. I guess we are leaving the realm of serious argumentation and entering that of religious emotions at that point, where I don’t see a need to engage in a discussion.

2 Likes
40

Well, my main point was to counter your statement that so called per-translations save time for the translator. But on the topic of what is considered low-stakes, that’s hard to qualify.

I’d also push back against the notion that just because some kind of ai is being used on an industrial scale that it is free from problems and dangers, or that this some how justifies its adoption in other areas. In fact it may only be after this mass adaptation that we really begin to see the problems with the current and future ai systems. In the same way that it’s only after mass adoption of the physical industrialization techniques that we were able to see the dangers there.

I agree with this sentiment, although I don’t like the implication that religious arguments are more emotional and somehow less serious. Personally I don’t subscribe to the whole “magic of the blank page” thing. But I do believe that Buddhist scriptures relate not only to this life, but future lives and ultimately complete liberation. Therefore a position of extreme caution is (in my mind) a rational position.

And my experience has been that there are too many people who seem to be unquestioningly onboard with whatever technology rolls down the line.

But I do want to reiterate how grateful I am for your participation here, even if I disagree with your position on many things. We are coming from very different backgrounds I’m guessing and I am glad to have you as part of the discussion.

1 Like