I asked it which island is north of the Philippines, and it said this:
The island located directly north of the Philippines is Taiwan (officially the Republic of China). Taiwan is situated in the Luzon Strait, which separates it from the northernmost part of the Philippines, specifically the Batanes Islands. Taiwan is an island nation with its own government, though its political status is a subject of international dispute.
However, I also asked it about world events in 1989, and it started answering the Berlin Wall, then when it got to Tianenmen Square the whole response disappeared and it said “Sorry, that’s beyond my current scope. Let’s talk about something else.”
Edit: it’s happy to talk about the great famine and the genocide of Uyghurs though, and even other student protests. Only Tianenmen Square is censored.
Another edit: but if you ask in Spanish it’ll tell you all about it. Ask it to answer in Mandarin Chinese and it won’t even try to answer, it’ll go right to the “beyond my current scope” message (in English).
pearsaltchocolatebar@discuss.online
on 02 Jan 2025 01:13
collapse
Tianenmen Square isn’t political. It’s an event that happened.
queermunist@lemmy.ml
on 02 Jan 2025 01:14
collapse
Have you ever heard of “political events”?
Viri4thus@feddit.org
on 01 Jan 2025 19:24
collapse
If they can really train a model at 600b parameters for 6M usd it’s an absolutely mind boggling achievement. No wonder it’s getting downvoted, it puts the capital intensive US firms to shame for lack of innovation. I’d love to see some benchmarks.
As a side note, this would open up the market for non English language LLMs with less requirements than the mammoth necessities of current models.
The trick that makes it possible is the mixture-of-experts approach. While it has 671 billion parameters overall, it only uses 37 billion at a time, making it very efficient. For comparison, Meta’s Llama3.1 uses 405 billion parameters used all at once. It also has 128K token context window means it can process and understand very long documents, and processes text at 60 tokens per second, twice as fast as GPT-4o.
Viri4thus@feddit.org
on 01 Jan 2025 20:36
collapse
Ty for the benchmarks and extra info. Much appreciated!
threaded - newest
We already met it.
It said that Taiwan was a part of One China, and it even used the phrase “we” when discussing it, as it agreed with the CCP policy.
I asked it which island is north of the Philippines, and it said this:
However, I also asked it about world events in 1989, and it started answering the Berlin Wall, then when it got to Tianenmen Square the whole response disappeared and it said “Sorry, that’s beyond my current scope. Let’s talk about something else.”
Here, try it yourself: chat.deepseek.com
Edit: it’s happy to talk about the great famine and the genocide of Uyghurs though, and even other student protests. Only Tianenmen Square is censored.
Another edit: but if you ask in Spanish it’ll tell you all about it. Ask it to answer in Mandarin Chinese and it won’t even try to answer, it’ll go right to the “beyond my current scope” message (in English).
Much like Gemini does whenever you ask it anything about US politics
Is that due to government censorship, or Google not wanting to potentially feed misinformation?
Exact same question can be asked about SeepSeek.
Tianenmen Square isn’t political. It’s an event that happened.
Have you ever heard of “political events”?
If they can really train a model at 600b parameters for 6M usd it’s an absolutely mind boggling achievement. No wonder it’s getting downvoted, it puts the capital intensive US firms to shame for lack of innovation. I’d love to see some benchmarks.
As a side note, this would open up the market for non English language LLMs with less requirements than the mammoth necessities of current models.
there’s some more info with benchmarks here, it does as well and in some cases better than top tier commercial models www.analyticsvidhya.com/blog/2024/…/deepseek-v3/
The trick that makes it possible is the mixture-of-experts approach. While it has 671 billion parameters overall, it only uses 37 billion at a time, making it very efficient. For comparison, Meta’s Llama3.1 uses 405 billion parameters used all at once. It also has 128K token context window means it can process and understand very long documents, and processes text at 60 tokens per second, twice as fast as GPT-4o.
Ty for the benchmarks and extra info. Much appreciated!
no prob