Open sourcing DeepSeek doesn't negate the propaganda built into the model. I don't really see why people think this is such a win. It's all propaganda, manufactured to get you to read more propaganda.
Discussion
I'm pretty sure the benchmarks it advertises are manufactured as well. DeepSeek seems substantially dumber than other LLMs, just from anecdotal evidence.
It's distilled, so it's dumber overall, but smarter in specific topics. You have to login to their website, to get the fuller, smarter version. And you pay for that because computation can't be open-sourced.
That said, a midwit intelligence that is small enough to fit on a normal cell phone and smart enough for particular implementations, definitely has a use case, but they didn't invent that. They're claiming that they produced it at a bizarrely low price, but they haven't published any evidence of this or explained exactly what is included in the calculations. Their competitors sometimes list hundreds of cost positions, like the salaries of the engineers and the rent of the office, and etc.
While Iâm suspicious of the true cost, I skimmed something other day that explained the innovation as a novel application of Reinforcement Learning (plus some other steps), as opposed to whatever is more standard (and expensive) practice.
I had that impression too, but maybe it matches the intelligence of AI lovers more, that's why they think it's so good.
Competition = good
Open source = good
not everything is good, but who wants all AI to be developed by 3-4 companies all subject to the same jurisdiction
That just isn't true: If the model and itâs training process is publicly available, you can remove the censorship: https://ollama.com/huihui_ai/deepseek-r1-abliterated:70b
Oh, neat, that's quite good actually
It would only be honest to admit that the implementations of stripping out some of the censorship/safeguards are still rather primitive and they donât seem to work all that well with deepseek r1 due to the way the model works internally: https://github.com/Sumandora/remove-refusals-with-transformers/issues/3#issuecomment-2622722682
So far most of the good models were stripped of their restrictions at some point.
I think all models have propaganda and cultural biases anyways. That being said, I do not want to aid the totalitarian and Orwellian CCP. But I see it as a win - a win over the hypocrisy of âOpenâ AI and the arrogance of Altman and Silicon Valley crowd in general. Humility finds us all in the end and humble pie has being served đ„§
I'm all for dunking on Sam Altmann, but celebrating the penetration of Chinese values into American tech seems more like a self-own
Yep. Just a rung on the ladder we are climbing. No need to linger. Letâs build on it and move forward correcting its handicaps.
Yes I see what you mean and agree. Itâs like a turning point in the whole game so far, who knows whatâs next! The American counter will be impressive I hope đ€đŒ
It's open. It also depends what you use it for
Isnât it crazy how much, seemingly well-founded, mistrust we have of systems from chinese companies. I tested a chinese hardware wallet (Keystone 3 Pro), but no way would I secure my funds on it or connect it to my computer. I am not inclined to use Deepseek for the same expectation of trickery. I will, however, be interested in replications of Deepseekâs tactics for efficiency gains by a trustworthy source. Iâm looking forward to a laptop running a highly optimized local, ethical, unfiltered, llm so that I can manage my own tool without some companyâs fears only allowing its use in a padded cell. One can dream.
That's not how models work. If it's open you can retrain it however you want and you can use the tricks and approaches in all other models, which is what all other AI companies do immediately.
So I also don't understand why were people thinking this is some hit to openai or others. Like they will literally use the same thing to make their model cheaper a s better...
propaganda is expected and that's the least of its problems. it has other huge lies..
price is free but the real cost is the misinformation built in it.
Yeah but itâs cheaper $đ
Yes like any other model
What items is the propaganda about?
I just asked about Tiananmen Square protests, and it gave me a valid response.
YMMV I guess, just tried and got this:
Why are yanks obsessed with asking DeepSeek about Tiananmen Square?
Why donât you try asking ChatGPT or Claude about who killed JFK? Or who was behind 911? Or what really happened at Waco? Or Vegas? Or whether you guys really went to the moon? Or which Presidents went to Epsteinâs Island? Or pizzagate? Or who is behind your fentanyl crisis? Or how many Jews are in control of the US apparatus?
All these LLMs are nerfed AF. They donât want you to have real power which could lead to power being overthrown. None of them are going to tell you the truth about shit that would enrage you and have tables flipped if ppl learn the truth.
So much cope going on for what China has censored when your own Tech mogulâs hide so much more from you having spent hundreds of Billions more for the same shit.
Truly pathetic to see.
Uh, sir this is a Wendy's
I don't use ChatGPT either
Right but if you had to choose, would you you use the open source LLM you can download and run locally that censors Tiananmen Square or the one you have to use hosted that censors all the shit I listed.
Iâll wait hodlbod.
Gather your thoughts whilst coping about how far gone your country really is.
I didnât ask about celebrating anything.
But youâd rather send your shit to Mr Worldcoin scan-your-retinas through his ânon-profitâ than running something locally.
Got it.
Why are you deliberately missing my point and presenting a false dichotomy in order to win a non-existent argument with me? Weird behavior.
Thereâs no false dichotomy.
DeepSeek you can run locally. Nothing goes to the CCP.
ChatGPT you canât - Orb man gets all your data if you use his stuff.
Youâre shitting on open source in favour of one of the biggest cunts in your country - that warrants a call out for you to check your own behaviour.
What about Ollama? Or even the non-censored version of DeepSeek? My point is that 99% of people are going to use the hosted version, not that it's impossible to use a better or self-hosted one. Anyway, you're being incredibly aggressive for no reason, so I'm muting you now.
Ollama is shit, it wonât even output regular expressions because they could be dangerous - I know because I run it locally.
The irony of muting me for this..
I used to really respect Americans but nowadays theyâre so programmed with patriotism that they canât see how far theyâve fallen. 
Don't forget 'Ruby Ridge' where the feds mercilessly slaughtered Randy Weaver's entire family , including the dog and his pregnant wife.
Because during model training process there are a lot of useful tricks, aka the core secret of openAI and DeepSeek found it independently ( at least they claim, and make it public as a paper)
And their work is also based on other open source llm
Technically, people could train an uncensored version. Or, more easily if itâs an rag filter, run the model on your local fixed