Deepseek took the training of foundation models from a billionaire’s game to a millionaire’s game. If Elon wants an AI monopoly, it’ll have to be done through litigation. Which, ya know, they’re also trying.

Flames burn and smoke asphyxiates, perfectly highlighting why relying on fire is a bad idea.

I’d say the egregore of a company does have a personality, but you’re right - it’s not human. We’re not great at understanding non-human intelligence even in other mammals. Another example being the discourse surrounding AI.

o1/o3 use a smaller model to summarize the reasoning, but they don’t show the actual CoT generation the way deepseek does.

It was a free o1/o3 equivalent at a time when there were only paid options. But in the short interim, Google’s made their r model free to use.

*Plus, deepseek doesn’t hide its internal monologue the way o1/o3 do. It’s fun to watch it go back and forth with itself.

True, but they also released a paper that detailed their training methods. Is the paper sufficiently detailed such that others could reproduce those methods? Beats me.

deleted by creator

My final semester in American Sign Language was “Sex, Drugs, and Profanity,” and most of the signs are just exactly what you’d guess. (I held on to those textbooks.) Plus, facial expressions are a big part of the grammar of the language. I don’t recognize this scene, but assuming it’s from a comedy - it’s probably also not far off from accurate.

deleted by creator

To run the 671B parameter R1, my napkin math was something like 3/4 of a million dollars in hardware. But that (plus the much lower training cost) made this a millionaire’s game rather than a billionaire’s. Plus the distillations do seem better than anything else we have at the smaller sizes at the moment. That said, I’m more looking forward to the first use of deepseek’s methods with google’s Titan architectures.