5 ms·

What most people don’t know about history & humanities is how much work there is to do. People get (rightly!) excited about decoding burnt scrolls in Herculane

What most people don’t know about history & humanities is how much work there is to do. People get (rightly!) excited about decoding burnt scrolls in Herculaneum. But most don’t realize that less than 10% of Neo-Latin texts from the renaissance to the early modern period have been translated to English. One example: Marsilio Ficino was a brilliant philosopher who was hired by Cosimo Medici to translate Greek texts (Plato, Plotinus, the Hermetica, etc) into Latin. This had a massive impact on the renaissance and European enlightenment. But many of Ficino’s own texts have not been translated to English! LLMs will, of course, have a massive impact on this… but so can students! Any student can make a meaningful contribution if they care to. There is so much to be discovered and unpacked. Unfortunately, so much of the humanities in high school is treated as an exercise rather than real discovery. I judge my students now based on how much I learn from them. Why not?

verisimi 1y ago
The history we have inevitably passes through bottlenecks. Much is left out, edited away, embellished. We have no idea what happened over 500 years ago, but the idea of a brilliant scholar translating the Greek texts on behalf of the Medicis, shouldn't simply be accepted as stated. If you are running the world (like the Medicis did) history would be a lever of control. It seems inevitable that the stories should be directed (carefully) towards whatever ends. History is actually a present day activity - it provides the backdrop to the present. Altering that backdrop has present value. I can't see how ai will help us the endeavour of trying to get a better handle on what happened in the past. I do see how it would provide modern-day Medicis a way to change the backdrop more quickly.
- dr_dshiv 1y ago
  > I can't see how ai will help us the endeavour of trying to get a better handle on what happened in the past. Well, AI democratizes the power of translation (and interrogating texts). Before only a few specialists could go directly to the source. Now anyone can try to make sense of it.
  - verisimi 1y ago
    We don't get to the source. Have you ever personally seen any historical sources? Eg the sources for classical writings or the Bible? At best you get high resolution imagery provided by suspicious folks such as RB Toth. If the source is corrupt - as I suggest and as is possible because the Medicis provided/sanctioned their version - all you have is interpretation based on flawed data. Endless production of information (by ai) based on flawed sources (this is our history) only serves to increase the haystack rather than helping you to converge on truth.
- UncleMeat 1y ago
  Paleography, transcription, and translation are hellishly difficult to learn and really make it difficult for interested people to explore the archive. LLMs have become pretty darn good at doing this out of the box. Prior to this you needed specially trained ML models and before that there was no automation whatsoever.
  - vintermann 1y ago
    I can't find any big AI model which can read historical Kurrent-family handwriting well out of the box. You still need specially trained models (i.e. transkribus) which generalize terribly.
    - UncleMeat 1y ago
      Transkribus is definitely still the best option in some cases. But there are a bunch of cases where it was necessary two years ago and isn't necessary anymore, which is pretty remarkable.
wizzwizz4 1y ago
The transformer architecture was originally developed for translation tasks, but massive, overfit generative models are awful at translation. (Seriously, part-of-speech classification + dictionary lookup + grammar mapping – an incredibly simplistic system with performance measurable in chapters per second – does a better job, and you even get confidence intervals.) If you want a translation tool, use something like Project Bergamot (https://browser.mt/ https://browser.mt/), not generative AI. > Unfortunately, so much of the humanities in high school is treated as an exercise rather than real discovery. I, too, find this exceptionally annoying.
- bonoboTP 1y ago
  That's an absurd claim. Show me the translation benchmark where you saw such a result.
  - wizzwizz4 1y ago
    Does "every time I've challenged people about this on Hacker News" count as a benchmark? See e.g. https://news.ycombinator.com/item?id=35531558 https://news.ycombinator.com/item?id=35531558.
    - fc417fc802 1y ago
      No, it doesn't. I don't find the linked thread remotely convincing regarding your (frankly preposterous) claim that preexisting non-deep-learning solutions are better at translation in the general case. Don't get me wrong. I'm sure that once fine tuned by a human for a specific language pair that such systems are better at performing literal translation. But the value proposition of deep learning here is that you don't need a large team of experts to laboriously train a given language pair, that the entire training process is largely unsupervised, and that the translations aren't hopelessly literal. The ML algorithms can pick up on idioms given a sufficiently large dataset.
      - wizzwizz4 1y ago
        Found the other good post in that thread: https://news.ycombinator.com/item?id=35532941 https://news.ycombinator.com/item?id=35532941. Google Translate and DeepL were capable of mostly-accurate Japanese translation two (or even 5) years ago (as seen in my previous link). Multiple people ran this text through GPT-4, and none of its translations were any good, as assessed by this Japanese-speaker. There's a reason we needed the Rosetta stone to give as an in for translating Egyptian hieroglyphics. You can't just throw Big Data at the problem and expect it to give accurate results. Dedicated translation setups can get good results, but ChatGPT's ilk doesn't. Convincing translations are not the same thing as accurate translations. RLHF optimises for convincing, not accurate.

verisimi 1y ago

The history we have inevitably passes through bottlenecks. Much is left out, edited away, embellished. We have no idea what happened over 500 years ago, but the idea of a brilliant scholar translating the Greek texts on behalf of the Medicis, shouldn't simply be accepted as stated. If you are running the world (like the Medicis did) history would be a lever of control. It seems inevitable that the stories should be directed (carefully) towards whatever ends. History is actually a present day activity - it provides the backdrop to the present. Altering that backdrop has present value. I can't see how ai will help us the endeavour of trying to get a better handle on what happened in the past. I do see how it would provide modern-day Medicis a way to change the backdrop more quickly.

dr_dshiv 1y ago
> I can't see how ai will help us the endeavour of trying to get a better handle on what happened in the past. Well, AI democratizes the power of translation (and interrogating texts). Before only a few specialists could go directly to the source. Now anyone can try to make sense of it.
- verisimi 1y ago
  We don't get to the source. Have you ever personally seen any historical sources? Eg the sources for classical writings or the Bible? At best you get high resolution imagery provided by suspicious folks such as RB Toth. If the source is corrupt - as I suggest and as is possible because the Medicis provided/sanctioned their version - all you have is interpretation based on flawed data. Endless production of information (by ai) based on flawed sources (this is our history) only serves to increase the haystack rather than helping you to converge on truth.
UncleMeat 1y ago
Paleography, transcription, and translation are hellishly difficult to learn and really make it difficult for interested people to explore the archive. LLMs have become pretty darn good at doing this out of the box. Prior to this you needed specially trained ML models and before that there was no automation whatsoever.
- vintermann 1y ago
  I can't find any big AI model which can read historical Kurrent-family handwriting well out of the box. You still need specially trained models (i.e. transkribus) which generalize terribly.
  - UncleMeat 1y ago
    Transkribus is definitely still the best option in some cases. But there are a bunch of cases where it was necessary two years ago and isn't necessary anymore, which is pretty remarkable.

wizzwizz4 1y ago

The transformer architecture was originally developed for translation tasks, but massive, overfit generative models are awful at translation. (Seriously, part-of-speech classification + dictionary lookup + grammar mapping – an incredibly simplistic system with performance measurable in chapters per second – does a better job, and you even get confidence intervals.) If you want a translation tool, use something like Project Bergamot (https://browser.mt/ https://browser.mt/), not generative AI. > Unfortunately, so much of the humanities in high school is treated as an exercise rather than real discovery. I, too, find this exceptionally annoying.

bonoboTP 1y ago
That's an absurd claim. Show me the translation benchmark where you saw such a result.
- wizzwizz4 1y ago
  Does "every time I've challenged people about this on Hacker News" count as a benchmark? See e.g. https://news.ycombinator.com/item?id=35531558 https://news.ycombinator.com/item?id=35531558.
  - fc417fc802 1y ago
    No, it doesn't. I don't find the linked thread remotely convincing regarding your (frankly preposterous) claim that preexisting non-deep-learning solutions are better at translation in the general case. Don't get me wrong. I'm sure that once fine tuned by a human for a specific language pair that such systems are better at performing literal translation. But the value proposition of deep learning here is that you don't need a large team of experts to laboriously train a given language pair, that the entire training process is largely unsupervised, and that the translations aren't hopelessly literal. The ML algorithms can pick up on idioms given a sufficiently large dataset.
    - wizzwizz4 1y ago
      Found the other good post in that thread: https://news.ycombinator.com/item?id=35532941 https://news.ycombinator.com/item?id=35532941. Google Translate and DeepL were capable of mostly-accurate Japanese translation two (or even 5) years ago (as seen in my previous link). Multiple people ran this text through GPT-4, and none of its translations were any good, as assessed by this Japanese-speaker. There's a reason we needed the Rosetta stone to give as an in for translating Egyptian hieroglyphics. You can't just throw Big Data at the problem and expect it to give accurate results. Dedicated translation setups can get good results, but ChatGPT's ilk doesn't. Convincing translations are not the same thing as accurate translations. RLHF optimises for convincing, not accurate.