about DeepSeek, but

It seems a little noisy about DeepSeek, but as I mentioned below, let me explain easily why LLM cannot be an AI killer technology. This is my personal opinion. Basically, LLM focuses more on whether or not the basic direction is to produce results for various questions rather than on accuracy. So, focus on finding the right answer for various scenarios during training. LLM is basically about 70-90% of accuracy in targeting, and only needs to be Max 90%. Because it's not directly related to the real world, and it doesn't hurt or fail with this result. That means if you have data, you develop some programming and train it, and it will quickly follow anyone until maturity, as in the graph below. Then, for LLM to survive, it must be available in the real world of the person without going through something human hands when it enters maturity. You need to have true reasoning and creation skills. It requires convergence with other services. So, there's no point in saying that you've been caught up with LLM technologies that are not 90% accurate and not related to other services. It's like, "I made an electric car, but I can't get out of the red." And now, they're all less accurate, so of course, they might like cheap ones. The usability is bound to be temporary and limited, and when you get a 99% accuracy service, people pay more for it. The reason why I appreciate xAI's potential is that it can converge services with Elon's companies as if they were one company.


This phenomenon can also be seen in FSD, where the initial version of V11 to V12 took off in the graph above. Then, from v12.3 to v13, it went to a stage where it became more mature. In the meantime, the FSD utilization rate (based on my Stats) has developed to less than 50% when v11, 90% when v12.3, 96-99% when v12.5, and 99.8% when v13. China's FSD also seems to have developed rapidly with the introduction of E2E, but most importantly, going from 90% to 99% and 99% to 100% is another level. To do this, we need to increase the accuracy, and the accuracy is achieved through optimization. Optimization of "sensing", "judgment", and "execution". You can never survive without an optimized process without latency because you have to judge something with real-time.

Both LLM and FSD are easy for anyone to reach 90% accuracy. However, the real fight is who gets 90% to 100%. FSD is already at 99.8%. What's interesting is that this is a winner-take-all fight, so once you attract customers, the game is over. Because trust is the most important thing.

This is just Jay's brain official, so just listen. I'm a Tesla guy. 🤣