Live data from Hacker News

Duolingo Max, a learning experience powered by GPT-4

blog.duolingo.com

61–70 of 250 posts

Re: Duolingo Max, a learning experience powered by GPT-4

#61
post #10

Just take a look at how poorly the bot does for Spanish: https://news.ycombinator.com/item?id=35155577

Garbage in, garbage out? Those examples are nonsense and I'm saying that as native speaker.

It does seem like the model picked up examples of pronunciations about which people are admonished, and presented them as tips of what to do rather than what not to do. This basic issue with negation also pops up a lot IME with Google's search infoboxes, so hopefully people are at least slightly inoculated.

Re: Duolingo Max, a learning experience powered by GPT-4

#62
Does it actually say how much Duolingo Max would cost? I do not see anything in that article or on their website. I don't have the app, maybe it's already live and pricing is available there.

I deleted Duolingo years ago when it felt like I knew everything it had to teach me, but was still far from fluent in the language. I assume it's gotten better, and would consider trying this new feature.

Re: Duolingo Max, a learning experience powered by GPT-4

#63

Earlier quoted context omitted.

My girlfriend is Spanish. I don’t think her family would be impressed if I were to pronounce “gracias” as “grassy ass” xD

I swear some people seem to say “grathios” in Spain but it seems to vary so much…

Yeah, Spaniards pronouncing a lot of 's' and 'c' sounds as English 'th' or 'z' is called the ceceo, and it's commonly incorrectly attributed to Spaniards copying a prior king's lisp.

Re: Duolingo Max, a learning experience powered by GPT-4

#64

Earlier quoted context omitted.

Irresponsible to give this to learners

My girlfriend is Spanish. I don’t think her family would be impressed if I were to pronounce “gracias” as “grassy ass” xD

Yeah, it sounds like Peggy Hill pronouncing Spanish.

https://www.youtube.com/watch?v=cByVeZhKmzo

Re: Duolingo Max, a learning experience powered by GPT-4

#67

This is interesting because my impression is that LLMs perform poorly on non-english languages[1]. This is likely 'just' a corpus size problem - but it's not like that is going to be fixed any time soon. That said, at least in language instruction there's a lot of meta material that will be helpful. I wonder how Duolingo is detecting when GPT-4 is hallucinating. [1] https://twitter.com/asmelashteka/status/16309662332…

For what it's worth, ChatGPT is able to hold a meaningful conversation in Czech (pretty difficult, latin script, ~11M speakers). I can almost tell it's a bot, but it could also be a PR person from a big corporation who can't even afford to talk like a normal person anymore.

Re: Duolingo Max, a learning experience powered by GPT-4

#68
post #16
post #8

Earlier quoted context omitted.

I haven't even noticed ads using uBO.

They're a lot more intrusive on the mobile app

I've had success blocking the ads using a pi-hole. You still see ads for Duolingo Super and other in-house offerings, but the obnoxious video ads go away.

Re: Duolingo Max, a learning experience powered by GPT-4

#69

This is interesting because my impression is that LLMs perform poorly on non-english languages[1]. This is likely 'just' a corpus size problem - but it's not like that is going to be fixed any time soon. That said, at least in language instruction there's a lot of meta material that will be helpful. I wonder how Duolingo is detecting when GPT-4 is hallucinating. [1] https://twitter.com/asmelashteka/status/16309662332…

ChatGPT's and Bing's German is flawless.

Re: Duolingo Max, a learning experience powered by GPT-4

#70

This is interesting because my impression is that LLMs perform poorly on non-english languages[1]. This is likely 'just' a corpus size problem - but it's not like that is going to be fixed any time soon. That said, at least in language instruction there's a lot of meta material that will be helpful. I wonder how Duolingo is detecting when GPT-4 is hallucinating. [1] https://twitter.com/asmelashteka/status/16309662332…

LLM outputs in Standard Chinese are fluent and coherent but the grammar style and choice of words are clearly non-native. I think the source training material just isn't as comprehensive as English, and contemporary Chinese Internet is packed full of latin abbreviations and obscure homophones/wordplay due to its evolution under severe censorship.

Awesome observation re: censorship. Last I looked into training sets, Chinese and English were the frontrunners by far in available corpus size, and benefit from a certain amount of government-enforced-homogeneity (a standard mode is taught in compulsory education).

Arabic is safe from automation due to its limited exposure to the open web and extreme variability of dialects, and Icelandic will remain obscure due to sample size and the fact that they mostly speak English online.

Post reply on HN