Even there, the “natural” equivalents have often changed meaning dramatically in the two millennia since the EBTs
But good enough compared to what? If people who speak e.g. Icelandic are curious about Buddhism, would we rather they read a bad machine translation of some Pāḷi suttas or a beautiful, human translation of Chogyam Trungpa? These things don’t exist in a vacuum.
This is exactly the mistake we should not be making. Machines are extremely charismatic. They exude an aura of objectivity. Their cold and calculated outputs appear rational and unassailable. But they are every bit as biased as any other human endeavor… because at the end of the day artificial intelligence is a human creation. Even if their biases are hidden away in a training set, in a million small decisions about weights and hyperparameters… we should absolutely not lose sight of the fact that AIs are every bit as biased as the humans that made them.
Which is why I feel the friends of Norbu Buddhist bot AI group on facebook, who helps to use and troubleshoot any issues with it is doing a great job in making AI safer for usage in Buddhism.
Strictly speaking in a verse setting, yes, which is why “Mass of Darkness” is a good translation, but one that misses the prose wordplay. So “Aggregates of Ignorance/Darkness” might be even more correct, right.
Well, if they choose to read CT, at leas they can do a quick google search to find out he was a raging alcoholic and make their own judgement. They can’t do the same research and find out about the problems of the MT they are reading.
I would be perfectly fine with someone using a state of the art tool that could translate from English into Icelandic. Because that transformation will always be the best there is for Icelandic. And I’m guessing that they would have experience using that same tool to translate other things into Icelandic, so they would have a good sense of the quality of the tool they are using. And if it came back with wacky things they would blame the tool, not the Buddha.
They would be able to judge the quality of the output based on their experience with that tool. So they have their own level of expertise, albeit not with Pali or even Buddhism.
And not to pick apart your argument on specifics, but a quick search seems to indicate that Icelanders have one of the highest levels of English proficiency in Europe. So I’m guessing they would choose to read suttas by a qualified translator in English than roll the dice with a MT. But it’s just a guess.
But if you ask me how I feel about publishing LLM translations from Pali into Icelandic, I’m going to say not at all.
I think the sticking point is if there should be these AI tools that go from Pali into target languages and if translators should allow their translations to be used for this. I guess my question is, who is going to be evaluating the quality of the translation from Pali into Icelandic? Because once you put it out there that tool X is made for translating Buddhist texts, then implicit in that is going to be some confidence for the user, no?
For the purposes of the essay series, though – in particular essay #1, we aren’t looking at the larger picture.
My point is that the purpose of the essay series was established right up front: to introduce and then rationalize the new policy of an AI-free SuttaCentral. Speaking here as a reader only; I can only surmise the author’s intent.
The comments throughout the essay threads ballooned IMO to discussing the absolutely largest picture that imagines an AI-free universe – and how absurd that is. It’s understood by just about anyone who has been reading the threads that this is not feasible on multiple levels. It can only be regulated for purposes of risk mitigation. However, going back to smaller picture, small instantiations like SuttaCentral can go beyond risk mitigation to full-blown risk removal.
I’m very familiar with how that goes. It’s always a vulnerable place to be as a researcher. It appears your team launched the project with a key assumption that turned out to be false: that the project would have unfettered access to the SuttaCentral data without terms and conditions. Is that what happened? Of course I don’t know the actual project details, but it appears this was a key assumption that may not have been well-vetted in advance of funding commitments. I may be totally wrong.
This touches on the rationale related to volume and speed for getting translations done, right? It either presupposes that the current approach has fundamental issues or that there is a compelling need to increase volume and speed (or both). Are either of these (or both) correct?
Yes, you’ve mentioned this before. I don’t expect everyone who disagrees with something in a public forum to “go public” on the forum and express that. Public exposure is not for everyone. (I mean, I trolled this site for two years before signing up with a profile.) And I think the SuttaCentral messaging feature is great for that reason alone.
I saw your recent X post that said something similar. In these threads I’ve seen people comparing this to the printing press, for example. Or airplanes. Or nuclear energy – maybe I haven’t actually seen that on the threads. (And I don’t consider facsimile technology anywhere close to an analogy.)
Actually, nuclear energy may be, for me, the closest AI analogy when we talk about trying to regulate technological progress. It is spot-used in medicine and it’s a viable alternative to fossil fuels for some people’s carbon-based needs. Nevertheless, it’s all about containment in the big picture for all the obvious reasons.
And, at the end of the day, there’s fundamental disagreement among people on these threads as to whether any plausible analogy exists at all. I tend to think No. But I don’t see any reconciliation of that. No one who thinks one way on that will change their mind because of the way someone else is arguing on that.
OK, a strong opinion says something. But in systems/organizations people ultimately have to make decisions. When it comes to data, for certain. Well, that ends up being policy.
I’ve spent years working with people – many of whom I reported to – who simply didn’t want to make a decision because of the blowback. I made decisions. Without making the decision (whatever it was), conditions persisted (whatever they were).
the larger picture I wanted to talk about here is ‘what is going to happen with SCs data and what implications does that have for us’. So maybe a misunderstanding of the term ‘larger picture’ here?
In a nutshell: The change of heart of a certain key individual a few months ago blew over what was believed to be a year-long collaboration in good faith that started with BuddhaNexus. If you want to be 100% AI free, good for you. I will respect that. The reality here is of course very complex. First of all, key individuals cannot speak for all of SC, and also not for all translators, and just because one person doesn’t want X to happen, doesn’t mean that other people agree with that. So we go through this one by one instead of writing it all off. Also, the Pali canon is just a fraction of the textual material that Buddhist traditions have produced, and a very small one looking at the larger picture. Certainly an important one, yes, but there is enough room to argue whether it is even very central or not. I guess that boils down to personal interest and preference.
Lastly, the Pali canon doesn’t suffer from a lack of translations, even very recent ones. So for us, replacing the data of translator X with that of translator Y is unnecessary work that nobody wants to do but it is not, ultimately, putting the project at risk, not at all. If somebody decides their work should not be relevant for the development of AI technology, the consequences of that decision is something they have to sort out with themselves, I am not going to give advice on that.
Our funding doesn’t depend on any of these assumptions here.
I get the idea that ‘look, the suttas are translated into English, nobody needs AI for that, and the commentaries are too difficult so your stochastic parrot won’t ever do something useful on that’.
Regardless of whether I agree with this premise (I don’t), even if that was the case, translating Pali into languages other than English, which make up the majority of what people use to communicate on this planet, is very important. I myself prefer to read content in German, even after living outside of Germany for quite some years.
For example, our machine translation system is currently successfully applied in a translation project to render the Pali canon into Tibetan, and its something that due to lack of staff power and funding wouldn’t be feasible without this technology. I.e. the level of funding that would be needed to educate enough people to carry out this task is just not there; with the help of machine translation, you can cut down the number of needed people by 50% or even more and make something possible that was not possible without (and since the Sutta’s language is generally highly repetitive, low-perplexity and with frequent predictable patterns, something that ML algorithms excel at, that is not difficult to achieve).
Another thing where we got unexpected positive feedback is rendering Sanskrit, Pali and Tibetan into Korean. A country with millions of Buddhists. The next step for us is to explore how we can facilitate translations directly into the major languages of South and South East Asia, empowering large groups of Buddhists to interact with the source texts with less barriers.
I don’t think its a good analogy because I just don’t see a major risk around generative language technology. ChatGPT is out for almost 2 years now, with no regulation, and the world hasn’t ended yet. Generative AI systems have millions of daily users and I yet have to see the big bad consequences that some people promised will happen. There might be some spam bots on facebook, and the one or other lazy student certainly tried to submit a gpt-written essay as a homework but apart from that I am just not seeing the problem.
Can it be misused to do something bad? For sure, the same applies to pretty much any other technology that mankind ever came up with. If at all, I think its more meaningful to compare this to how cars and planes were developed in the first place. I do think that generative AI will need much less regulation, than that technology. In the state that technology is right now I don’t see the need for regulation at all, and I am very happy that we have open source language models that anybody can download and start working with. And here as well, I yet have to see the evil applications that are empowered by that and that endanger humanity.
Once we get close to "real AI’, that might change, but I don’t find it meaningful to talk about something that is not going to happen anytime soon.
Right now we are likely where we where in the early 20st century regarding transportation. The new technology is here, but people will make absurd claims to defend sticking to what they are used to. “LLMs won’t ever be useful for translating commentaries, they can only work on text they already know” is as meaningful as saying "I can ride with my horse in the forest without restrictions, anywhere, while these stupid cars have to stick to roads that are man-made and predefined, there will never be a good use case for them’. Yes but that is just not the point. In the same way as cars didn’t make legs obsolete, since you still need your legs to get to the top of the mountain, I don’t see why the claim that LLMs can’t do complex commentaries says anything about whether they can be used productively in translation. I see a lot of applications that urgently need to be developed.
Interesting article today - a group of current and ex OpenAI employees together with Google DeepMind researchers released an open letter warning about the dangers of AI:
Interesting quote:
They also assert that AI companies possess substantial non-public information about their systems’ capabilities, limitations, and risk levels, but currently have only weak obligations to share this information with governments and none with civil society.
To me, what you’ve said here is, quite frankly, chilling. With all that has been said in this series, not having and enforcing strict policies has the most serious grave repercussions.
The way I see it, your larger picture is BethL’s smaller picture.
I am genuinly concerned about the “AGI-like” models that will be able to do most human jobs within 3-8 years and their immidiate effect on society. I do not buy the bullshit, that universal basic income will simply solve this issue.
Again, the models would contain ALL opinions / possible solutions at the same time for finishing a text snippet, and whichever MODE is returned depends on fine-tuning, the initial prompt and randomization.
One thing that could be interesting (and I think you mentioned something similar to it) is to have multiple bots generate solutions at the same time. We can then compare them or have other bots to compare them. Whenever there is a huge difference in translation / disagreement between bots, that part is probably difficult and would require a human’s attention. (This does not mean that no disagreement would gurantee a correct translation, or that disagreements couldn’t ever happen on easy examples).
Not sure if any of this is useful, but the way the code generation LLMs are used is also different: they are hyped as replacements for software developers, but what really happens (for now) is that the LLMs are good with common patterns, leet code questions, converting between languages, writing boilerplate code, creating summaries and tickets, but still make mistakes. GitHub’s Copilot can automatically resolve issues in a repository in around 40%+ of the cases. But you still need a senior engineer to be able to figure out if the code generated is indeed within the 40% and to be able to fix the generetaed code if there is a mistake. Note that some of the errors are due to the limitations of tokenization (like not being able to correctly represent offsets in line numbers) and the quality of the results depend a lot on the initial prompt and the code snippets selected for context! The discussion there is also very polarized: some people hype the technology and worry that programmers will become obsolete, some claim that programmers will never be replaced and bring up examples of other cases where AI fails.
PS:
By the way, where is Bhante, why is he not responding to any of this?
When “The Lords of AI” talk about universal basic income as solution, they simply push the responsibility of masses losing their jobs becuase of their software to governments.
I also find it cynical when Demis Hassabis brings it up, after Google laying off 30 000+ people. Maybe this year they were just unliucky, but next year, the shareholders will finally have enough and they decide to distribute the rest of their wealth instead! Maybe by that time, even Amazon will finally start paying taxes that governments can redistribute!
Stimulus checks after COVID were sort of take on unviersal basic income that prevented depression. However, it also made the rich richer, because people have been spending their extra money on stuff they usually didn’t need or put it into questionable investments. Even if we do expect everyeone to be able to use money wisely, the things in demand will simply become more expensive. Housing is already unaffordable, so throwing free money at it will likely further inflate prices.
And there is also the fact the people need something to do and some also have the need to find meaning in their jobs and move ahead in hierarchies. I can imagine scenairos where people would either endulge in meaningless nihilism as a result and / or join far-right groups as a final resort to feel somewhat special. And the AI arms race (both chips and algorithms) is also partly a war against China. I can’t imagine the general distribution of surplus wealth during wartime either.
(Note that I am not aruging about the need for social benefits. I live in the EU and I find it brutal how people in the US can pretty much lose everything overnight due to lack of safety nets.)
That’s where they can adopt a yogi lifestyle full time (should they wish to).
When people are no longer concerned with work to survive, likely after some hedonism, and nihilism, they would start to go to religions.
Indeed, DN26 said:
Do not let injustice prevail in the realm.
Mā ca te, tāta, vijite adhammakāro pavattittha.
Provide money to the penniless in the realm.
Ye ca te, tāta, vijite adhanā assu, tesañca dhanamanuppadeyyāsi.
and we see the poverty is the root cause of crime. So it’s another plus of eliminating world hunger plus elimination of crime due to poverty.
Not really a good argument to deny the elimination of world hunger.
they simply push the responsibility of masses losing their jobs becuase of their software to governments.
Well, DN26 already said it’s the duty of kings. Anyway, the most fair way is to have an AI tax to tax companies which uses AI to lay off people, so the work done by machines is actually given back to all humans.
Plus, so many pilot programmes of UBI around the world simply show that it’s not true that people will stop working once they got UBI. They will choose better jobs, not lousy jobs. Likely prostitution would be reduced in size as well as child slave labour to produce goods, who will be replaced by robots anyway.
I 100% agree with you that reducing poverty will also help people to live a more moral life, as they less likely have to make hard choices. And getting richer in general will also make the rich more rich.
We have been pulling out masses from poverty incredibly fast thanks to automation and globalistaion.
I believe that now that everything will be automated, human touch (having a person listen to you) will increase in value. We already have perfect solutions to generate coffee, yet people still got to coffee houses to drink the very same thing but overpriced. Similiarly, some roles will pay more when done by a person even if there are better solutions available.
It is also great that we now have the time to think about Buddhism and to resolve traumas transmited through multiple generations in everyone’s families. I think we are extremely lucky.
I am not doubting or agianst any of that. I just simply don’t believe that the solution is as simple as UBI.
