Remote OK founder Pieter Levels argues Japan lacks proprietary foundational models, noting Rakuten AI is a fine-tune of China's DeepSeek

@levelsio about 20 years out of date

the cope, it never stops those innovative Japanese… what was the last Japanese innovation that made a splash? Ahegao?

@yayoi_mizuha None of these models rank high at all, no Japanese models even show up in the top. What's your point exactly?

I can fine tune DeepSeek on my GPU at home

@levelsio japan tried a few times but seemingly gave up after a few half-assed attempts

in general the people with money (like softbank) refuse to spend money on r&d, they want others to do the work and then they want to monetize it

@teortaxesTex apparently living in a country that isn’t china for three years makes you very knowledgeable about chinese innovation and industry

@levelsio “Doesn’t even have” is simply false. Japan has several orgs releasing LLMs trained from scratch, including PLaMo, Sarashina, llm-jp, and Stockmark. Some Japanese companies also fine-tune Qwen, DeepSeek, or GPT-OSS; Rakuten is just one example.

@yayoi_mizuha It literally does not have any LLM like DeepSeek, which Japanese LLM ranks same or better than DeepSeek in benchmarks then?

@teortaxesTex boob jiggle physics in anime.

@levelsio You’re shifting the argument. The claim was that Japan “doesn’t even have its own LLM like DeepSeek.” That is false.Nobody is saying Japan is on par with the US or China in frontier LLMs. It clearly isn’t.

@teortaxesTex Really surprises me that these guys still haven't figured out how to redact the quotation marks. Maybe it just doesn't make any difference at this subsistence tier of engagement farming?

@teortaxesTex Man, the French really are still salty about their beloved Rafales getting murked by the J-10C

@teortaxesTex The send cut send guy said it best, Japan has been stuck in the year 2000 since 1980.

@visakanv I went last year on December, had last been there 2018, it was completely different, it's a very exciting place to visit and see it change every time and generally really friendly people

@levelsio Japan is stuck in 1980s

@yayoi_mizuha @levelsio you can spin the situation however you like, before frontier model race, Japan was already known to be behind, what would tell you they are anywhere on pace to catch up?

First one learns what is good. Then one learns to produce what is good. Then one can innovate on top.

Most people have come to admit China is best at 1 to N.

In half a decade, it’ll be widely understood what is already clear: China is also as good at 0 to 1 as US, if not better.

@levelsio Japan really just does not do software well at all, you have major companies putting out software that would flunk a compsci class. It’s only the video game companies that have usable software

@levelsio Yeah you’re right. The AI landscape isn’t well developed here and there are hardly 2-3 companies like Rakuten, CyberAgent or NII that are releasing their models

Some are trained from scratch(with not many tokens). The SoTA ones are probably fine tuned.

@levelsio The point isn’t to claim that Japan is competing with DeepSeek on performance. The issue is whether Japan has the capability to train LLMs from scratch domestically.

@levelsio open source levels the field. fine-tuning is how we ship now anyway, robot vision or vibe coding. building from scratch is overrated, making it work matters more.