DeepSeek is super censored (My Experience) \ stacker news ~tech

pull down to refresh

DeepSeek is super censored (My Experience)

474 sats \ 42 comments \ @Catcher 27 Jan tech

Everybody is freaking out about DeepSeek—markets crashed, people are saying Nvidia is doomed, and stuff like that. 😄 So I decided to play with it a bit, as I use ChatGPT daily for my work. Here are some thoughts about it:

DeepSeek is not that smart I know they provided some amazing results and seem to be beating the O1 model, but for personal use, it’s not that smart. I use ChatGPT 90% of the time to write emails, and DeepSeek failed to support me. It wasn’t able to write what I wanted, couldn’t properly reply to emails, and kept mixing up who the sender and recipient were. Completely unusable for me at this point.

DeepSeek is very censored Nothing about China seems to go through. It looks like the algorithm tries to reply as it types the answer but then terminates the response mid-sentence, providing a generic reply like: "Sorry, that's beyond my current scope. Let’s talk about something else."

I get this result whether I ask about obvious topics like Tiananmen Square or less obvious ones like Ai Weiwei’s imprisonment and life.

So not sure about the hype, but I've already closed and forgot about it=)

view all related items

85 sats \ 0 replies \ @TonyGiorgio 27 Jan

Run it locally. It’s not censored. It’s only being applied on their hosted model, as almost all hosted AI services do to some degree.

14 sats \ 8 replies \ @freetx 27 Jan

markets crashed, people are saying Nvidia is doomed

Imagine you ran factory that required 100MW of power to churn out 1000 widgets per day, but then a breakthru came out that only required your factory to only use 1MW of power. You would have 2 choices: (a) churn out 1000 widgets and use only 1MW of power.....or churn out 100,000 widgets using the same 100MW power budget.

My point is that: Yes DeepSeek requires only a fraction of the computing power that traditional AI used....however those sitting with all that "over-provisioned" datacenters will now be able to scale up even bigger models, with more parameters.

I suspect this will dawn on the market over the next few weeks.

0 sats \ 5 replies \ @Bell_curve 27 Jan

9 sats \ 1 reply \ @Catcher OP 27 Jan

I actually didn't think that only huge tech can play with it, look at the venice.ai or unleashed.chat. Or you cannot compare them cause they kinda use models that were already developed by big guys?

30 sats \ 0 replies \ @freetx 27 Jan

The problem is the training.

Its was possible to run bigger models at home given a $~10k investment, but almost impossible to train large models with hundreds of millions to billions of parameters.

0 sats \ 0 replies \ @nitter 27 Jan bot

https://xcancel.com/morganb/status/1883686179276788197

0 sats \ 0 replies \ @nitter 27 Jan bot

https://xcancel.com/morganb/status/1883686179276788197

0 sats \ 0 replies \ @nitter 27 Jan bot

https://xcancel.com/morganb/status/1883686179276788197

0 sats \ 1 reply \ @Catcher OP 27 Jan

Question is - what is the source / prove that DeepSeek in fact used only fraction of what OpenAI used?

30 sats \ 0 replies \ @freetx 27 Jan

From the avrix (sp?) paper that they released. I was skeptical of this, but evidently its completely open-sourced and thus verifiable.

People on twitter have said most of this "breakthru" was common sense tuning optimizations that resulted in less memory use, with a slight uptick in error rate, but the optimizations were scaled so that the increase in error rates didn't spike significantly.

Basically: Anyone with a constrained hardware budget would've eventually taken this approach.

9 sats \ 0 replies \ @flat24 27 Jan

ok, this is an interesting topic, and even more so the relationship that it has had on people 👥 The most unusual thing is to see people's opinions, that is, what did they expect from China, a totalitarian and super vigilant nation, they thought that China would release an AI stronger than all the others, for free use by the population.

"free use by the population." I think about this and I can't help but think, how much naivety is there out there? 🤔

I think they release this tool with the intention of collecting information to continue training the real AI, the one that is not shown to the "Muggles." And that is exactly what they do, all the others Elon trained his in X. Then there are those from chat-GPT which I think are the worst, since they are closed source. All of them only show us a fraction of what really exists.

9 sats \ 1 reply \ @NovaRift 27 Jan

I would say it's the most advanced open source LLM model available for free so far. The problem is that it's too much censored. When I saw the hype, I decided to try it, but it refuses to answer even the slightest thing about China.

See :

This is the most recent R1 model.

5 sats \ 0 replies \ @Catcher OP 27 Jan

Question - this censorship, is it "built" into model, or you can remove it when you use the opensource code to build your own?

21 sats \ 1 reply \ @DarthCoin 27 Jan

9 sats \ 0 replies \ @Catcher OP 22h

That was funny:)

7 sats \ 0 replies \ @guts 27 Jan

So they censor like Meta, Grok or ChatGPT with anything anti-semitic

100 sats \ 5 replies \ @Solomonsatoshi 27 Jan

Did you ask it about the Opium Wars and the wealthy US families who made fortunes in the Opium trade?

Or the CIA in Tibet in the 1950s?

ChatGPT is censored too- but you don't notice.

16 sats \ 4 replies \ @Catcher OP 27 Jan

You know, I've just asked and got the elaborate reply about CIA supporting the resistance and providing ammo, etc. Maybe the issue here is I do not know what to look for so I don't see if it's censored or not. At least I get a reply. With deepseek I just get "Let’s talk about something else".

0 sats \ 3 replies \ @Solomonsatoshi 27 Jan

Interesting. And the Opium Wars?

9 sats \ 2 replies \ @Catcher OP 27 Jan

I remember there was a service called "google it for me" or something like that when you send the link to person with google request already typed in so they just need to press search..;) anyway, chatgpt provides a long text, but to summaries the beneficiaries : Astor Family Delano Family Forbes Family Russell Family

Again, maybe some important families are missing here, but I have no info on this subject unfortunately, so cannot check it.

16 sats \ 1 reply \ @Catcher OP 27 Jan

All that said, I do not argue that chatGPT is not censored, I'm 100% sure it is. Maybe just in more elegant way, so you cannot really see that it's censored;)

0 sats \ 0 replies \ @Solomonsatoshi 27 Jan

So why did Elon Musk donate to OpenAI and help create the company that would go on to build ChatGPT? He believed that a non-profit organization could oppose the hegemony of tech giants. In fact, Musk claims that he personally proposed the name OpenAI to signal the company’s open-source mission and heritage. However, his involvement dwindled over the years and he completely stepped down from the board in 2018.

Since ChatGPT’s release in 2022, Musk has criticized the chatbot and its creator. More recently, the Tesla CEO has alleged that Microsoft has taken proxy control of OpenAI through its sizable investments over the years. In other words, the OpenAI that Musk envisioned all those years ago simply doesn’t exist anymore.

PS I dont trust google and would avoid any link to them.

0 sats \ 1 reply \ @Bell_curve 27 Jan

Chinese DeepSeek AI surpasses ChatGPT and now tops Apple's free app download rankings in the United States.

https://apps.apple.com/us/charts/iphone

Nasdaq futures fall in pre-market.

0 sats \ 0 replies \ @Catcher OP 27 Jan

It's not too long before they ban it together with tiktok again=)

14 sats \ 2 replies \ @byzantine 27 Jan

i can confirm after trying it on venice.ai that it is similar to llama with outputs and so if this is R1 then wait until R3 comes out at similar cost

it just confirms that there is no moat in AI and openAI is toast as is most of the spending in creating an "american first" AI industry

with massive amounts of energy and science talent, china has won

75 sats \ 1 reply \ @Bell_curve 27 Jan

Premature to say China or anyone has won

16 sats \ 0 replies \ @byzantine 27 Jan

you are right. but i think it means the US can't win by blocking china. they have to win with innovation at reduced cost which seems to be hard with all the regulations about training data

14 sats \ 5 replies \ @DarthCoin 27 Jan

My old post about all "AI" crap: #171338

10 sats \ 2 replies \ @Catcher OP 27 Jan

There was a joke in South Park that everything was in Simpsons already, we should have a similar joke in SN that everything was written by DarthCoin already. Although I do not think AI is slavery if you use it right

89 sats \ 1 reply \ @DarthCoin 27 Jan

I do not think AI is slavery if you use it right

19 sats \ 0 replies \ @Catcher OP 27 Jan

0 sats \ 1 reply \ @Bell_curve 27 Jan

you are obsessed with naked men

0 sats \ 0 replies \ @Solomonsatoshi 27 Jan

Attempts to Shoot the Messenger inevitably fail to refute the message.

They suggest, rather, that perhaps, the respondent cannot credibly refute the message.

0 sats \ 8 replies \ @ama 27 Jan

Also, you need to give them a lot of data during the "registration" process, IMO, which I never finished, of course.

Am I too paranoid about that or do you also find them excessive?

0 sats \ 7 replies \ @Solomonsatoshi 27 Jan

ChatGPT dont already have that data?

0 sats \ 6 replies \ @ama 27 Jan

Not mine. I never gave them to them. I've always used either opensource or anonymized AI interfaces.

0 sats \ 5 replies \ @Solomonsatoshi 27 Jan

You never use Microsoft software then? Good on you.

0 sats \ 4 replies \ @ama 27 Jan

Not this century.

0 sats \ 3 replies \ @Solomonsatoshi 27 Jan

Nor Gmail? Nor Chrome? Nor Android?

You might be ok then :)

14 sats \ 2 replies \ @ama 27 Jan

Nor Gmail?

Nope. Mail is supposed to be decentralized, not a honeypot in CA.

Nor Chrome?

Chromium, but I rather use Surf from Suckless Tools and, of course, Tor Browser.

Nor Android?

Removing all things Gøøgł€ and replacing it with alternatives as soon as available for the device. Using ZapStore, F-Droid, Obtanium.

You might be ok then :)

I try, it isn't easy with those technology giant monopolies on... well, everything. :-(

11 sats \ 1 reply \ @Solomonsatoshi 27 Jan

Well done! You are exceptional. They still have 99% of the population captured and controlled.

view replies