Deepseek R1 scam? Could China use AI openAI for specific tasks?
I tested R1, both online and locally. I noticed significant differences in quality (although I was only able to run the mid-range model due to capacity constraints).
Just a theory: Could R1 delegate various tasks to openAI in the background and thus achieve these extreme values (level with low power consumption)? Has anyone achieved a similarly high level with a local installation as with the online version?
It would be the perfect plan to crash Chinese tech stocks.
(1 votes)
Loading...
This answers the question relatively immediately. 70B is only a fraction of 670B, accordingly the quality is also a fraction..because to think that it leads a few things over OpenAI, is relatively stupid.
The ding is open source. Absolutely everything you can see and check.
Such an API consumption for OpenAI would be very expensive. Since OpenAIs API is ultra expensive. by deepseak, about 20 times so much cheaper. This would make OpenAI notice that a user is asking for your MILLIONS on the day, and that would be an absolute loss business. Makes no sense at the front and back
the online version is of course not. And of course, it would be a loss business, but as an economic attack from China it would not be any more. It is also not a matter of directing everything in general, but of delegating only partial tasks or certain tasks.
It is also, and I say this explicitly, not a concrete suspicion, but an observation from a superficial comparative test and a thought that has arisen to me.
No. DeepSeek is completely open source, so you can see what the program does. Apart from that, it would be timelessly expensive to transfer data to ChatGPT and let them edit – would also notice quite quickly, especially with open source code.
You use a small local model, which is obviously much less powerful.
The local versions are finegetunete other LLMs such as Llama or Qwen who should do as if they were Deepseek
The 70B Llama Model does not come to the 671B R1 Model ran
That’s just the surface or not? You pull the Deepseek R1 model and not the Llama model right.
If you have half Terabyte RAM, you can run deepseak
The rest is not derpseek
so far clear, but still I find it is quite an extreme drop and a test task in the field of development was quite similar to that of r1 online. The R1 local was extremely far away. I would have to make sensible test series, that was just a spontaneous idea why the results seem so comparable and still so much less resources are needed.
They have used ChatGpt data for training so is not surprising
It’s not long. but also the mac mini setup
This was once, current can be up to 256GB
At some point, your CPU has reached the limit. Normal consumer CPUs go up to 64-128 GB RAM.
Yes rams get thrown behind, but I just don’t know if the part can be more than 256GB even if I buy dei bars. You’d have to look at that.
Ah ne Memory Controller says 256gb max. rip 671B
in principle yes, but this would also have to go into the direction in the local small training set.
Then make a few more bars in
No, I’ve got an old server from nem family member standing around here. Threadripper gedöns.
My desktop board has “only 64GB”
What system do you have with 250GB RAM? Is that a consumer mainboard?
Just a quarter, unfortunately. But a version smaller doess too.