BlackRock's Larry Fink urges trillions in AI investment to keep US ahead, citing national security risks if China advances. His proposal to use retirement savings sparks social media uproar.
DeepSeek is one of if not the most popular Chinese AI and is open source and requires a small amount of computing compared to others. Its used in numerous Chinese car brands, smart phones, and even government services throughout China.
China isn’t competing for Third they are leading the world in AI development and have already integrated it in many areas.
Again, at the very least, no, they’re not open source. Open-weights means anyone can download, use, and tinker with them a bit, but there is no access to their code, training data, or process.
It’s just as limiting as closed source software with some modding allowed, but not as limiting as an online-api-only model, as many of the most powerful modern models are.
There are no heroes in the Global Powers’ race. The USA is a comically cartoonish villain in real life, yes, but all the biggest Chinese data centres for all that training, are still built in poor areas (Inner Mongolia and the bullshit that China has apparently inflicted on them), and still fucking over those who live there.
You mean all of this code that is clearly on their github: https://github.com/deepseek-ai? They release both their model weights as well as the source code for their AI. You can literally take what they have provided to create your own LLM if you would like to and get a good understanding of their AI. Sure you can’t see the training data but that would be like putting the entirety of the internet in a github repo and just isn’t feasible, but you can contribute your own training data to a local setup of deepseek and shape it in a way you want to.
The training data is as important as the source code here to replicate the end result. The weights are more like a binary distribution. You can run the model and you can technically edit it just like you can technically edit a binary file.
They also only release some libraries and tools for running the model if you have a set of weights (which they do graciously provide), but they do NOT release the source code for their training pipeline itself. That’s up to you to reverse engineer from the whitepapers. Right now even if you had the exact training data and the compute available, you could not train your own Deepseek V3.2, let alone V4.
Haven’t they clearly documented how they did it and what they used so that anyone can replicate it? Anyone with the compute power, which of course few have. But universities could do it.
So how is it not open source in this specific domain of problems? What would a LLM model need to do to be open source then? Duplicate the whole training dataset in a big zipfile for you to download?
From what I understand you could even replicate deepseek by replacing the “cold start” with latest deepseek instead.
Haven’t they clearly documented how they did it and what they used so that anyone can replicate it?
They don’t put up the actual code for their training pipeline though. It’s more of a “if you have enough engineers, you can do this too” whitepaper, because they wouldn’t want any rando training their own model.
Right now, even if you had the exact training set (which is a CRUCIAL part of an LLM and you can NOT replicate it without it), you couldn’t rebuild the thing exactly, you’d need to do a whole lot of extra work.
So how is it not open source in this specific domain of problems?
You could call all proprietary software open source then. The UI and user manual describe what it does, you can do your own engineering to duplicate the functionality.
DeepSeek is one of if not the most popular Chinese AI and is open source and requires a small amount of computing compared to others. Its used in numerous Chinese car brands, smart phones, and even government services throughout China.
China isn’t competing for Third they are leading the world in AI development and have already integrated it in many areas.
Again, at the very least, no, they’re not open source. Open-weights means anyone can download, use, and tinker with them a bit, but there is no access to their code, training data, or process.
It’s just as limiting as closed source software with some modding allowed, but not as limiting as an online-api-only model, as many of the most powerful modern models are.
There are no heroes in the Global Powers’ race. The USA is a comically cartoonish villain in real life, yes, but all the biggest Chinese data centres for all that training, are still built in poor areas (Inner Mongolia and the bullshit that China has apparently inflicted on them), and still fucking over those who live there.
It’s abuse all the way down.
You mean all of this code that is clearly on their github: https://github.com/deepseek-ai? They release both their model weights as well as the source code for their AI. You can literally take what they have provided to create your own LLM if you would like to and get a good understanding of their AI. Sure you can’t see the training data but that would be like putting the entirety of the internet in a github repo and just isn’t feasible, but you can contribute your own training data to a local setup of deepseek and shape it in a way you want to.
You’re talking like you know what you’re talking about, but you clearly are guessing. Knock it off. Don’t mask conjecture as fact.
The training data is as important as the source code here to replicate the end result. The weights are more like a binary distribution. You can run the model and you can technically edit it just like you can technically edit a binary file.
They also only release some libraries and tools for running the model if you have a set of weights (which they do graciously provide), but they do NOT release the source code for their training pipeline itself. That’s up to you to reverse engineer from the whitepapers. Right now even if you had the exact training data and the compute available, you could not train your own Deepseek V3.2, let alone V4.
If people on Lemmy can’t understand this I have no hope for the average person.
Haven’t they clearly documented how they did it and what they used so that anyone can replicate it? Anyone with the compute power, which of course few have. But universities could do it.
So how is it not open source in this specific domain of problems? What would a LLM model need to do to be open source then? Duplicate the whole training dataset in a big zipfile for you to download?
From what I understand you could even replicate deepseek by replacing the “cold start” with latest deepseek instead.
They don’t put up the actual code for their training pipeline though. It’s more of a “if you have enough engineers, you can do this too” whitepaper, because they wouldn’t want any rando training their own model.
Right now, even if you had the exact training set (which is a CRUCIAL part of an LLM and you can NOT replicate it without it), you couldn’t rebuild the thing exactly, you’d need to do a whole lot of extra work.
You could call all proprietary software open source then. The UI and user manual describe what it does, you can do your own engineering to duplicate the functionality.
Why even have this discussion? Self-learning algorithms appeared more than ten years ago. AI is being used very effectively in countless areas.
The idea that there is some sort of prize waiting for whomever gets the most computing power, is highly dubious.