yorwba

6 ms·

61.
by yorwba 1mo ago
Chinese ML researchers primarily publish in English and only secondarily in Chinese. For example, take the Qwen-3.8-Next blog post https://qwen.ai/blog?id=qwen3.8-flash-next (which apparently doesn't include the langua
62.
by yorwba 1mo ago
Their representation is a key-value store that they embed via a tensor product to reconstruct the vectors the model produces. The catch is that it works best if the key-value mapping is task-specific. So if you already know how to solve the
63.
by yorwba 1mo ago
Then I will attach my feedback about a not well-parsed instance here: https://moviescenemap.com/#title=1942 has 5 locations listed, but of the "narrative locations" on Wikidata, at least Saipan and Midway Atoll se
64.
by yorwba 1mo ago
A LLM should at the very least be a language model, i.e. be able to take human-readable text as input or produce it as output. Transformers are used for plenty of tasks that don't involve language, for example object detection or blind
65.
by yorwba 2mo ago
It's helpful for learning a foreign language to read a text while simultaneously listening to it. But when you're not very good at the language yet, you'll likely often fall behind. Word highlighting makes it easy to get back
66.
by yorwba 2mo ago
For PoW to work, you don't need scrapers to have less compute than users, you only need their compute advantage to be less than the load-generation advantage they would otherwise have. And unlike legitimate users, scrapers aren't
67.
by yorwba 2mo ago
China's biggest humanoid robot manufacturer is Unitree, who sold 5215 humanoid robots last year: https://mp.weixin.qq.com/s/R2t4zdvCZmWj_HpTRJAXzg According to the filing for their IPO this month, they had 175 R&a
68.
by yorwba 2mo ago
Are you sure they don't represent drum memory: https://en.wikipedia.org/wiki/Drum_memory
69.
by yorwba 2mo ago
The day/light line is a great circle. The "straight line paths" are great-circle segments.
70.
by yorwba 2mo ago
The complaint in that article is that "instead of creating a single European system, the PPWR preserves a fragmented national model. A business selling directly to customers across the EU must register and fulfil its obligations separa
71.
by yorwba 2mo ago
https://hnrss.org/frontpage?q=wikipedia.org&search_attrs=url
72.
by yorwba 2mo ago
So I took literally the first complexity theory textbook PDF I could find https://theory.cs.princeton.edu/complexity/book.pdf#page=324 where we have Lemma 16.43 Let ε > 0. For every n and k ≤ n there exists a
73.
by yorwba 2mo ago
There is a table of squares from ≈2500 BC https://cdli.earth/artifacts/10678 that predates the earliest known copies of the Sumerian King List by a few centuries. With completely wrong values for the numerals, the math
74.
by yorwba 2mo ago
The "lie" is that some Bézier-spline drawing programs let you restrict on-curve points to a "symmetric" configuration where the control points are at an equal distance on opposite sides of the on-curve point, but this do
75.
by yorwba 2mo ago
There's no shortage of sources seeking to deliver the most important news as quickly as possible. It's the niche stuff that few people care about that requires effort to seek out.
76.
by yorwba 2mo ago
The Weld-Blundell Prism https://en.wikipedia.org/wiki/Weld-Blundell_Prism doesn't use positional notation, so 5.50.0 is expressed as 𒐧𒐢 (five 3600, five 600) and 5.10.0 is expressed as 𒐧𒐞 (five 3600, one 600). (Li
77.
by yorwba 2mo ago
I think you're overindexing on government involvement and underindexing on FOMO driving investors to jump on any new trend. When Unitree briefly reached a $69B market cap after IPO earlier this month, that wasn't because of govern
78.
by yorwba 2mo ago
Yes.
79.
by yorwba 2mo ago
A patch was submitted, but apparently not merged. That was also my experience trying to submit a patch for https://trac.ffmpeg.org/ticket/8738 . Somebody on the bug tracker took note, but was apparently unable to effec
80.
by yorwba 2mo ago
They're trying to explain the phenomenon, where during the training of a deep neural network, sometimes it will spend a long time with suboptimal parameters and then suddenly improve a lot very quickly, which has been termed "grok
81.
by yorwba 2mo ago
Even if it weren't slopped together, 65% vs 80% on 10 tasks just isn't a significant difference. For 80% power to distinguish at a significance level of 0.05, you'd need more like 140 samples, if those were the true success p
82.
by yorwba 2mo ago
SEO juice, malware distribution, harvesting payment information, selling the domain to the real model provider... Lots of ways to make a quick buck. For example, last year when Lightricks released the LTX video model, somebody else register
83.
by yorwba 2mo ago
No, I mean the WebText corpus whose construction from 45 million Reddit post with at least 3 karma is described in section 2.1 of the PDF I linked. They did remove all Wikipedia documents. Anna's Archive didn't exist in 2019.
84.
by yorwba 2mo ago
I'm not aware of any successful accusations against OpenAI for illegally obtaining copyrighted material, in contrast to Anthropic, who settled for $3000 per work and then still had to buy legal copies to keep using them (likely for m
85.
by yorwba 2mo ago
In China, it has more to do with putting drug users into labor camps: https://madeinchinajournal.com/2019/10/25/punish-and-cure-fo...
86.
by yorwba 2mo ago
GPT-2 was trained using data scraped from the web ( https://cdn.openai.com/better-language-models/language_model... section 2.1), i.e. copyrighted data provided free of charge to anyone with an internet connection.
87.
China's Overhyped Embodied AI Sector(chinai.substack.com)
2 points by yorwba 2mo ago|0 comments
88.
by yorwba 2mo ago
There's an established network of intermediaries that can supply a large variety of books for a few dollars apiece, so no need to contact copyright holders directly.
89.
by yorwba 2mo ago
AI companies legally acquiring books have indeed been in the news: https://news.ycombinator.com/item?id=49330742 And where are you getting the idea that Mistral doesn't train on copyrighted data? There's not a lot
90.
by yorwba 2mo ago
You mean this one https://committees.parliament.uk/writtenevidence/126981/pdf/ where they write "it would be impossible to train today’s leading AI models without using copyrighted materials"? That

More ›