I have a similar experience using Gemini for quick Catalan translations for iOS apps given enough context.
I once asked it to summarize The Hobbit in Catalan to explain it to my daughter before sleep. I was expecting a lot of mistakes as I see regularly if I ask anything in my native language when using GPT or Claude, but it was surprisingly good. I was going just to kind of skim ahead and retell it my own way, but ended up almost saying it verbatim because it was good already.
She loves Zelda so I asked it to explain the story of Breath of The Wild keeping the original names, and to make it fun, etc.. I was surprised again. I did retell some bits in my own style and taste but it is very convincing.
I haven't tried Catalan on newer models like GTP-6 Astra or Fable tho. We have all these benchmarks based on software development, and AGI, etc.. but it would be cool to have some language benchmarks for different communities.
As I work in english and use them in english, I wonder if using LLMs in a different language to code renders a different result as well. Like, if some of these benchmarks were made in other languages, would the result be similar.
> As I work in english and use them in english, I wonder if using LLMs in a different language to code renders a different result as well. Like, if some of these benchmarks were made in other languages, would the result be similar.
I am not a Chinese speaker but my understanding is that all of the models have substantially different behavior in Chinese, to the degree that it's kind of like a second model. Would be interested in hearing more if anyone has direct experience.
Fwiw, it's been very easy to nudge Qwen 27B/35B to think in Chinese. No need for <think> prefills like "思考：" ("Think:") or such.
Relatedly, when chatting with frontier models about science education content design, I've found it very helpful to mix in Chinese education terms. In English, for example, NGSS is such a massive attractor, discussing nearby topics often yields NGSS "slop". And "estimation" (educational) in the US means one (dysfunctional:) thing, which similarly distracts. Perhaps if AIs become increasingly multilingual, but remain weak at deep conceptual reasoning, it may be fruitful to have multilingual thesauruses, to use language-associated cultural conceptual differences as a way to convey conceptual nuances with which LLMs otherwise struggle?
"Estimation" as a skill in US education, primary school on, is pervasively about point estimates - single numbers. Bounding, a range of safe/reasonable/likely/possible numbers, is very rare. In Fermi questions/problems also.
Even in such goodness as Sanjoy Mahajan's Art of Insight and Street-Fighting[1]... the only only "bound" mentioned is "Printed and bound in [USA]", and there is no "round".[2]
In contrast, IIUC, China is pervasively about "Da Gu" and "Xiao Gu", "round up" and "round down". Point estimates are secondary. Lower and upper bounds are apparently even a cultural concept: an individual's degree vs capability/fate; policy equity vs excellence; economic targets.
In education, my understanding is point estimates have much poorer conversational/collaborative dynamics than range estimates. Bounding lends itself to incremental collaboration. "Can anyone suggest another low or high bound?". Discussions around point estimates seem usually merely about the collection of individual estimates.
> NGSS
Ah, NGSS isn't bad particularly. I meant that discussion/mention of NGSS appears so very much in English training data, that AI chat on topics related to NGSS seem to draw regurgitations of NGSS. AI "slop" in the sense that "ignore NGSS, just reason from a superset of its underlying concepts"... isn't an LLM strength.
ERRATA: Those google book links now do have hits for "bound" and "round". Also "range" and "interval". The searches really didn't before, thus the reference links. Perhaps the absence should have failed sanity check, but given that one hit... partial failure of Google Books search wasn't something I recall noting previously. Oops.
Yes, and that brings in a shit load of slop to the problem space from the creationists.
Like how the word life will have a strong correlation with death, the word animate, while it also means alive will have a much less strong correlation with death.
Does European public dialog do much causal comparisons among states? Does somewhere else?
Or do you mean professionally... Instead of merely searching for "ways to do X", one might translate the query into a couple of languages, and translate the results, to get a broader culture sampling. Hmm... That seems something an AI-based search could integrate smoothly - cross cultural insights as a normal part of search results?
While there's a lot of diversity in policy and culture among US regions and states, contrasting them isn't leveraged much. Perhaps failing a "that's analysis, not news" threshold for being part of public dialog? So maybe it's not "non-English cultures" neglected as much as comparative discussion in general?
One "criminally underutilized" which really struck me, was Japan having a far more sensible and powerful COVID safety one-liner than the US managed - "Three C's" (Closed spaces, Crowded places, and Close-range contact - poorly-ventilated / crowd / conversation-contact - avoid overlapping all three) vs "6-foot rule".
As others already commented a lot of content creators are getting copyright claims for just 5 seconds humming of a song and this allows them to edit that part and change it.
It’s funny how people want to edit their tweets when they misstype something but don’t want others to do so. In a video of 10min there are a lot of mistakes to be made and things that can be changed.
> people want to edit their tweets when they misstype something but don’t want others to do so
Those are actually two different (though somewhat overlapping) kinds of edits. People want "I wrote tyop instead of typo and that makes me sound like a idiot, so imma fix that". They don't want "@jswift1729 tweeted that we should kill children for food, but then edited it so everyone compaining about it sounds delusional".
This problem could be (mostly) solved easily by showing the most-revised version by default, but including a text to the effect of "edited 5 times ; last edited 1970-01-02 04:30:21" linking to a list of all revisions.
I moved to the UK but I still know many friends from Spain that work from 9am-1pm -> lunch break -> 3pm-7pm. Which means that in reality they can end up leaving the office at 7:30pm.
By the other hand I worked for the Catalan TV for many years and they had very flexible hours and I could leave by 5pm, but it's not that usual in my experience (and that was 12 years ago).
For commercial apps one needs to buy data. Personal projects can use free data from some websites. I have ideas (and others' suggestions) to improve the UI of the app, but decided to pursue other directions.
I once asked it to summarize The Hobbit in Catalan to explain it to my daughter before sleep. I was expecting a lot of mistakes as I see regularly if I ask anything in my native language when using GPT or Claude, but it was surprisingly good. I was going just to kind of skim ahead and retell it my own way, but ended up almost saying it verbatim because it was good already.
She loves Zelda so I asked it to explain the story of Breath of The Wild keeping the original names, and to make it fun, etc.. I was surprised again. I did retell some bits in my own style and taste but it is very convincing.
I haven't tried Catalan on newer models like GTP-6 Astra or Fable tho. We have all these benchmarks based on software development, and AGI, etc.. but it would be cool to have some language benchmarks for different communities.
As I work in english and use them in english, I wonder if using LLMs in a different language to code renders a different result as well. Like, if some of these benchmarks were made in other languages, would the result be similar.
reply