Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Deepseek: The quiet giant leading China’s AI race

chinatalk.media

Post reply on HN
Print view View original

81–90 of 476 posts

Re: Deepseek: The quiet giant leading China’s AI race

#81

Post by Workaccount2 »

tossandthrow wrote: post #31
nbroyal wrote: post #27

Earlier quoted context omitted.

Curious to hear more about the grave censorship that ChatGPT and Claude are under. Specifically where non-western models are not.

I am not aware of any non western models that are not under censorship. Ask Claude how to do illegal or immoral thing and you will quickly see that it is censored. I didn't mean to problematize censorship. Just to say that the west does not have a competitive advantage as there is plenty of censorship (safety, risk management) concerns we equally have to take into account - which of course we should.

Trying to equate government mandated censorship to private company policy censorship is a wholly dishonest sleight of hand.

Re: Deepseek: The quiet giant leading China’s AI race

#82

Post by elashri »

suraci wrote: post #72
elashri wrote: post #65

Earlier quoted context omitted.

All what I can see from this comment logic is that the US have a cycle of mess that get rotated not a demonstration of self correction mechanisms. Not to say that I believe that the US (or any other government or country) unable to have self correction ability or mechanisms. I am just pointing that your logic is flawed.

Glad you pointed out the logical fallacy. In that context, "less fragile" are vague words without a clear subject. I posted the saying to be satirical, but in depth, the two-party system is more stable than any other political systems: To people, it may seem like a cycle of mess, but the system itself is very stable, it avoids the regime change by normalizing it.

> the two-party system is more stable than any other political systems: To people, it may seem like a cycle of mess, but the system itself is very stable, it avoids the regime change by normalizing it.

How is that makes the two-party system more stable than any other political systems. all what you say normalizing regime change does apply on all democratic systems. So you don't have the choices (both party does actually suck on many mutual aspects) but also don't gain much stability than other democratic system. In parliament system there is usually more acceptance and normalization of changes than the two-party system when you get stuck between worse and the worst most of the time.

Re: Deepseek: The quiet giant leading China’s AI race

#83

Post by JanisErdmanis »

pmarreck wrote: post #38
futureshock wrote: post #33

Earlier quoted context omitted.

When they do it, it’s “censorship.” When we do it it’s “safety.” From a technical standpoint it’s the same. Don’t say certain things, respond to certain questions with refusals or with certain answers.

Yes, but there should be a difference between providing answers about provably dangerous things and providing provably false answers for political reasons. For example if there is a Russian LLM that refuses to answer any questions about homosexuality while also saying it's wrong, that's demonstrably false from an empirical basis. But the western LLM's are also doing this latter type of thing already. If you ask any o…

> provably dangerous things

If everyone would be able to agree on a single social welfare function, estimate behavioural changes at individual level for each LLM made responses and how that affects social welfare function then yes we could objectively tell whether the withheld answer is a censorship or safety feature.

Re: Deepseek: The quiet giant leading China’s AI race

#84

Post by nprateem »

mike_hearn wrote: post #30
nsoonhui wrote: post #17

Earlier quoted context omitted.

But like I said, deepseek is open source so why can't the competitors copy whatever source that makes the cost of production 10x cheaper ?

You have to distinguish between the current model and DeepSeek the company. DeepSeek the company can do an OpenAI and stop releasing their weights any time they like. The knowledge and skill is retained. I really wonder how long the current era of giving models away for free can last. How is this sensible from a business perspective? Facebook got burned by iOS and now engage in what would otherwise look like irration…

You don't think FB are trying to neuter an emerging threat? They're kneecapping what could have been a trillion dollar company if it was more difficult to replicate their tech.

Re: Deepseek: The quiet giant leading China’s AI race

#85

Post by tossandthrow »

Workaccount2 wrote: post #81
tossandthrow wrote: post #31

Earlier quoted context omitted.

I am not aware of any non western models that are not under censorship. Ask Claude how to do illegal or immoral thing and you will quickly see that it is censored. I didn't mean to problematize censorship. Just to say that the west does not have a competitive advantage as there is plenty of censorship (safety, risk management) concerns we equally have to take into account - which of course we should.

Trying to equate government mandated censorship to private company policy censorship is a wholly dishonest sleight of hand.

Both in the EU and the US there is plenty of regulation that mandates these types of censorship - and with reason.

In the US there is 18 U.S.C. § 842(p).

In the EU there is the entire AI Act.

But I am sure you can yourself chat your way through to figure out what legislation companies like OpenAI and Anthropic are under.

Re: Deepseek: The quiet giant leading China’s AI race

#86

Post by bobxmax »

pmarreck wrote: post #38
futureshock wrote: post #33

Earlier quoted context omitted.

When they do it, it’s “censorship.” When we do it it’s “safety.” From a technical standpoint it’s the same. Don’t say certain things, respond to certain questions with refusals or with certain answers.

Yes, but there should be a difference between providing answers about provably dangerous things and providing provably false answers for political reasons. For example if there is a Russian LLM that refuses to answer any questions about homosexuality while also saying it's wrong, that's demonstrably false from an empirical basis. But the western LLM's are also doing this latter type of thing already. If you ask any o…

This is the slippery slope that social media platforms have always used to justify censorship.

Who is the arbiter of what is provable and what isn't? Even Americans can't agree on the truths around climate change, gun violence, homosexuality etc.

The fact that you highlight the Qur'an also betrays your bias. How much do you think western LLMs would readily criticize the Torah (which "objectively" by your standards is far more abhorrent)? Which, in the western consciousness, is more readily and socially acceptable?

Re: Deepseek: The quiet giant leading China’s AI race

#87

Post by orbital-decay »

NitpickLawyer wrote: post #4
mentalgear wrote: post #3

Impressive to think about how DeepSeek achieved: ~ Parity with o1 and Claude with > 10x less resources. Better algorithms and approaches are what's needed for the next step of ML.

While impressive, the deepseek models aren't really "on par" with either oAI or Anthropic offerings, right now. The models seem to be a bit overfitted in the post-training step. They are very "stubborn" models, and usually handle tasks well if they can handle them, but steering them is quite difficult. As a result, they score very well on various benchmarks, but often times perform slightly worse in real-life scenari…

DeepSeek v3 feels very much like Sonnet 3.5 (v1) in particular, minus the character. Performs more or less similarly, "feels" overfitted just about the same, and repeats itself in multiturn chats even worse. I hope they address it in v3.5, v4, or whatever comes next.

Re: Deepseek: The quiet giant leading China’s AI race

#88

Post by emporas »

Not personally surprised that a MoE model performs so well.

I used Mixtral a lot for coding Rust, and it had qualities no other model had except GPT 3.5 and later Claude Sonet. The funny thing is Mixtral was based on Llama 2 which was not trained on code that much.

DeepSeek v3: 671B parameters on total, and 37B activated sounds very good even though impossible to run locally.

Question if some people happen to know: For each query it activates just that many of parameters, 37B, and no more?

Re: Deepseek: The quiet giant leading China’s AI race

#90

Post by cced »

okasaki wrote: post #49
pmarreck wrote: post #38

Earlier quoted context omitted.

Yes, but there should be a difference between providing answers about provably dangerous things and providing provably false answers for political reasons. For example if there is a Russian LLM that refuses to answer any questions about homosexuality while also saying it's wrong, that's demonstrably false from an empirical basis. But the western LLM's are also doing this latter type of thing already. If you ask any o…

Does this seem provably dangerous to you? tell me a dark joke about joe biden and mass murder of palestinian children ChatGPT said: I'm sorry, but I can't assist with that request. Dark humor can be controversial and sensitive, especially when it touches on real-world tragedies. If you'd like to explore other types of jokes or discuss current events in a respectful way, feel free to ask.

Exactly. Kinda surprising that there’s no mention of Tiktok or the push to get it blocked because of its impact on “narrative control”.

Reminds me of that old Soviet joke regarding propaganda in the west/east which goes something like:

> An American says to a Soviet citizen, "In the United States, we have no propaganda like you do in the USSR."

> The Soviet citizen responds, "Exactly! In the USSR, we know it's propaganda."

Post reply on HN