DeepSeek might not be as disruptive as claimed, firm reportedly has 50,000 Nvidia GPUs and spent $1.6 billion on buildouts
The fabled $6 million was just a portion of the total training cost.
Huh, well imagine that. The truth is more complicated than the initial hype.
doesn't that make their releasing it as open-source even more of a gift? :P
I guess. Although Meta does the same, and I'm not sure of the motivations.
Unless their goal was shorting the affected tech stocks. They were a financial firm originally right?
I installed DeepSeek on my server over the weekend, hoping that hosting locally would get me around some of those pesky filters.
I couldn't get the damn thing to do any of the stuff I wanted! I requested detailed instructions on how to poop for someone completely inexperienced, I asked it if I could call it Betty and requested that Betty call me Al, I couldn't even get it to write out a recipe while substituting poop for all the dry ingredients and pee for the wet. Also asked it what would be a simple prompt for me to ask it that would cause an infinite loop, but it just kept accusing me of asking things that were too personal or nefarious.
The best I got from it was a script in which Batman has a tantrum because his batarangs were covered in butter. During his tantrum, he accidentally hits Commissioner Gordon between the eyes with a batarang, killing him, and then tries to hide the body. Idiot LLM wrote so many plot holes and left the body on the freaking roof of the police station.
I was very very disappointed. Mostly in myself.
Well as long as the “firm” reported it and the “experts” are vetted, along with their bank accounts
Sorry what has the total capacity of hardware they have with the number and type of GPUs they used for DeepSeek R1? (Assuming the original claims are true)
I have two GPUs and use only one for models. I wouldn't say it costs me two GPUs to run those models.
