[dmurray]

The "abstract Wikipedia" just seems like a solved problem with LLMs.

However sceptical of "AI" you are, "give me the information on this page in my preferred language" is the kind of task they excel at. (I won't use the word translate). It wouldn't even require prioritising the English Wikipedia: any agent today could one shot a task like "check the Wikipedia pages in all languages for X, summarize the results and note any disagreements between them".

[bawolff]

Abstract wikipedia is taking a symbolic AI approach instead of an LLM or other statistical approach. The hope is (as i understand it) that this will provide reliability, predictability and better extend to languages that don't have a large corpus of text to train things on.

Personally i think its a bit of a wild bet, that seems especially surprising in the modern context. Guess we'll have to see if it pans out.

[emodendroket]

I'd kind of expect that they do better with translation if one of the languages on either end is English due to the amount of input they get in it compared to this abstract language (even in the world of human translation, translating stuff into English as a "pivot language" and then doing every translation from the English translation rather than the original text is not an uncommon practice).

[internet_points]

> However sceptical of "AI" you are, "give me the information on this page in my preferred language" is the kind of task they excel at.

Except for the 90% or more of the world's 7000-ish languages which have barely any data online.

E.g. the huge CommonCrawl corpus has stats https://commoncrawl.github.io/cc-crawl-statistics/plots/lang... for only 160 languages. English takes up nearly half the corpus, and after the top 16 or so all languages have <1% of the corpus, over half of those 160 have <0.1% and the other 6000+ languages are distributed amongst the <unknown> category. The long tail is very long.

(You'll see people use the term "low-resource language" and then talk about Finnish or Macedonian – if you're not a linguist and you've heard of the language, it's most likely not low-resource ;-))

[ccgreg]

Common Crawl is working hard to improve diversity in our crawl.

[dotancohen]

  > give me the information on this page in my preferred language

I'm sure that works great for European languages and other languages with huge corpus. Those are not the target languages of the program in question.

[caturopath]

LLMs are great with minority languages compared to almost anything else. Including better than the by the natural language generation employed to use Abstract Wikipedia, which whiffs at relatively large languages like Zulu and Xhosa, let alone many of the rarer languages that popular LLMs speak fluently.

[dotancohen]

This program is aimed at getting actual humans to write their actual language. Nothing beats that.

At a minimum, it provides more material to train an LLM on.

[tovej]

LLMs are really bad at smaller European languages even, e.g. scandinavian ones, or Finnish. Much worse than the NLP situation before LLMs.

[caturopath]

My experience with their Norwegian has been fantastic, I'd be shocked if the other Scandanavian languages aren't at least as good.

[Wikipedianon]

It's not a good idea for common languages like German or English or French.

But it is a great idea for indigenous languages that aren't in the training data but many people speak, which was the original purpose.

I am hopeful that it'll create synthetic training data for those groups.

[danaris]

> "give me the information on this page in my preferred language" is the kind of task they excel at.

...So long as you don't mind it introducing random hallucinations into the information.

[archagon]

Biased by its training data to boot, the opposite of what you’d want in an encyclopedia.