Le mer. 29 janv. 2025, 13:17, John Clark <[email protected]> a écrit :
> On Tue, Jan 28, 2025 at 6:55 PM Brent Meeker <[email protected]> > wrote: > > >> *> I wonder how sure we are that there's no malicious code in Deepseek?* >> > > *Deepseek is completely open source and transparent, so anybody can check > its source code, however it's written in Nvidia's intermediate level > Parallel Thread Execution Programming language PTX, rather than Nvidia's > high level language CUDA that most other AIs are written in. Since PTX is a > lower level language and therefore closer to machine code than CUDA is, > it allows the Chinese to employ lots of low level optimization techniques > that improves performance, but that does make it a little more difficult to > check for malicious code although it's not impossible;* > Yes, just use an AI that you trust to check it ;) Surely that's what anthropic and openai do or will do. *and it's so small and bare boned there isn't enough room to fit in a lot of malicious stuff. And the Chinese published a journal article that goes in the considerable detail about how they made Deepseek so they are not hiding any technological secrets, anybody can make an AI similar to Deepseek now. * > > *DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement > Learning* <https://arxiv.org/pdf/2501.12948> > > *Also, anybody is free to completely retrain it, the first thing I'd do is > educate it about the Tiananmen Square Chinese uprising which Deepseek > currently knows absolutely nothing about.* > > *John K Clark See what's on my new list at Extropolis > <https://groups.google.com/g/extropolis>* > stw > > > > > > > > >> >> On 1/28/2025 12:21 PM, John Clark wrote: >> >> *I own some Nvidia stock and yesterday it dropped 17 percent, I also own >> TSMC and ASML stock and they also dropped by a lot, but I am not worried >> because the reason for the drop was the Chinese company released an Open >> Source small AI that was nevertheless as good as the best much larger >> western models, and they made it using a new method that only needed about >> 5% as much computing power to train as western models do; even better they >> also released a 22 page article explaining exactly how they did it:* >> >> *DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via >> Reinforcement Learning* <https://arxiv.org/pdf/2501.12948> >> >> *Hi tech billionaire Marc Andreessen described making this knowledge >> available by the Chinese as a "Profound gift to the world**", and I >> think he's right. The semiconductor stocks dropped dramatically yesterday >> because people figured that now that we know how to get more IQ points with >> fewer computer chips, people will be buying fewer computer chips. But they >> have forgotten The Jevons Effect >> <https://en.wikipedia.org/wiki/Jevons_paradox> , as the cost of a resource >> drops the demand for it increases, and we will never have too much >> intelligence therefore we will never have too much computing capacity. * >> >> *And I don't think we should be looking at this as a victory of China >> over the USA but rather as a victory of Open Source AI programs over >> proprietary AI models. I also think that due to recent developments Ray >> Kurzweil is going to need to radically change his prediction that the >> singularity won't happen till 2045; I think he will need to chop off at >> least 10 years off that estimate, probably more. Let's hope that Eliezer >> Yudkowsky is wrong about what will happen during the Singularity because he >> can't be very happy today. * >> >> *DeepSeek R1 Shocked The World* >> <https://www.youtube.com/watch?v=IMXRpbxifq8> >> >> >> lez >> >> >> >> >> >> >> -- > You received this message because you are subscribed to the Google Groups > "Everything List" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > To view this discussion visit > https://groups.google.com/d/msgid/everything-list/CAJPayv1kGNHRwoQmh_HLSeJvZ4WGcLQ5GBqwvK6w7hWP6unKaA%40mail.gmail.com > <https://groups.google.com/d/msgid/everything-list/CAJPayv1kGNHRwoQmh_HLSeJvZ4WGcLQ5GBqwvK6w7hWP6unKaA%40mail.gmail.com?utm_medium=email&utm_source=footer> > . > -- You received this message because you are subscribed to the Google Groups "Everything List" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion visit https://groups.google.com/d/msgid/everything-list/CAMW2kAoV8Ea4eShNSPCvw%2B-AukqEBcNnkBfyJJnd%2Bv_A5579%2Bg%40mail.gmail.com.
