There’s no way I believe that Deepseek was made for the $5m figure I’ve seen floating around.
But that doesn’t matter. If it cost $15m, $50m, $500m, or even more than that, it’s probably worth it to take a dump in Sam Altman’s morning coffee.
DeepSeek claimed the model training took 2,788 thousand H800 GPU hours, which, at a cost of $2/GPU hour, comes out to a mere $5.576 million.
That seems impossibly low.
DeepSeek is clear that these costs are only for the final training run, and exclude all other expenses
There would have been many other runs before the release version.
There is no downside to lying these days. Yet the public seems surprised that all they see is lying.
Nevertheless, like the funding-hungry CEO he is, Altman quickly turned the thread around to OpenAI promising jam tomorrow, with the execution of the firm’s roadmap, amazing next-gen AI models, and “bringing you all AGI and beyond.”
AGI and beyond?
I kind of suspect this is as much about A.I. progress hitting a wall as anything else. It doesn’t seem like any of the LLMs are improving much between versions anymore. The U.S. companies were just throwing more compute (and money/electricity) at the problem and seeing small gains but it’ll be awhile before the next breakthrough.
Kind of like self-driving cars during their hype cycle. They felt tantalizingly close 10 years ago or so but then progress stalled and it’s been a slow grind ever since.
I hope that normal people will now realize how full of sh*t he is. They won’t, but DON’T TAKE THIS FROM ME
Does the elephant call the ant hopeless?
One of them is threatened with extinction. ;-)
