It’s easy for us at this point to see a story in the news and accept it. It’s easy to believe something that we won’t personally deal with or attempt ourselves, assuming that it is out of our wheelhouse. We believe and therefore it is.
DeepSeek hit at the end of January. It shook the Western world with it’s ability, it’s power, it’s reported cost of use and efficiency. It rattled the markets with respect to AI hardware and software.
Then they began digging into the model and saw that it was an optimization of Open AIs model. Latter it was determined that the actual training cost was reported originally at $5 million, and then $6 million on a correction. But it seems that they didn’t count the cost of the hardware, which put it something around $0.5 billion. So it was on par with the Western world, just perhaps a little more thought out and efficient.
This is all well reported in the media. But I downloaded an LLM engine that worked locally on my system. And then I downloaded a couple of DeeoSeek R1 models, that were mods of existing models. And I tested them to see what would happen. My hardware is old, so it was very slow. I expected that. I asked it about things that would be forbidden in China, famous protest, etc. As expected I gave the answer that nothing at all happened. I asked it about events in America history that reflect negatively on the U.S.A. and it had no problem filling in all the details with great accuracy. Then for fun, I asked it if it though that Trump was good for the U.S. and it hallucinated an answer unlike any hallucination I ever saw. It told me Trump was the president for 2 days in 1968 and that there were a bunch of Mexican immigrants that protested in Ohio in 1970. Some how those were related, I have it, so if I wanted to read it, I can, but it’s not worth the effort. It’s meaningless and strays from the point.
The point is that DeepSeek was a bunch of hysterical nonsense. It’s not that much better than say Llama 3.2, if at all. It’s not really demonstrably faster. The hallucination is something beyond what I’ve ever seen, but I have heard about on younger models. So what it seems is that people don’t actually check a damn thing out for themselves.
My experience with DeepSeek puts it in the same place as any other LLM, but perhaps not as polished. It’s interesting that it comes from China, but at this point, we should expect real competition from outside the U.S. and without it, we will simply stagnate. I don’t care about the espionage allegations, it is being done inside the U.S. all the time and part of tech culture. Always has been. We need more composition and less collusion, which is the plus of DeepSeek. But seeing it myself, it’s just a reminder that nothing is accurate in the media. That was a complete shit show. I wonder how many LLMs wrote this news. It’s not believable now. I believe nothing. I tested it and the results I saw made it the same as Llama 3.2 with the added hallucination. It was a bunch of noise and I suggest that you don’t believe anything you read or see or hear. If you can test it. Test it. If you can’t, find someone who can and ask lots and lots of questions. Never believe any media about anything. If you really want to know something, ask an LLM about it and see if it gives you the same or similar output as you get from the media. My guess is that it will.
The thing that pissed me off more than anything else is that too many influences claimed to have run it and claimed to have all these incredible results. But I doing it myself, it was fabricated. And others have since actually made their own attempts and had similar things to say. It’s just tragic that people have money invested in these things and they suffered while others took advantage of the Hysteria. It was a wake up call to be aware that China does have tech, as does the rest of the world, and maybe they will catch up and provide honest competition. But the real issue is the deception in reporting. I am disgusted by the revelation.
Sigma Punk 