September 11, 2023, According to Forbes and the news outlets, Morgan Stanley analyst Adam Jonas upgraded Tesla to overweight with a target price of $400. https://www.forbes.com/sites/dereksaul/2023/09/11/tesla-stages-80-billion-rally-after-morgan-stanley-ups-price-target-on-ai-optimism
There is no news coming from Tesla. Tesla spiked up 10.09% $25.08 to $273.58. It appears this reaction is mainly from the upgrade. The report from Adam Jonas focused on the AI potential of Tesla. Elon Musk had previously publicly announced project Dojo, AI training chip to power its video AI training for Full Self Driving (FSD).
The market had seen the potential for AI from results of ChatGPT and is now rewarding companies more that are using AI to create better or new products. Tesla is building FSD using AI techniques.
Tesla’s FSD V12 will be done using AI from an excerpts on an upcoming Elon Musk biography by Walter Issacson published on cnbc.com: https://www.cnbc.com/2023/09/09/ai-for-cars-walter-isaacson-biography-of-elon-musk-excerpt.html
FSD V11 had been implemented using AI for recognition of street signals and signs and cars with rules written by developers. The rules in FSD V12 will be replaced with AI Neuron Network trained on a large collection of video from Tesla’s fleet of Model 3, Y, S and X.
With the large population of owners of Tesla cars, there is range of driver skills. If Tesla had used all the video, the AI produced would micmic the skill of the average Tesla driver. Tesla can be better than average by being more selective with what video to train its AI. It can do this because of its expanding fleet of HW3 enabled cars that can supply specific video clips. These clips would have situations where a driver using FSD Beta had a disengagement. Tesla will send a request to the fleet of cars for video with similar situation. Through this method, Tesla and build a collection of useful and necessary video clips to train and close any gaps in the AI’s capabilities. The neat trick based on the cnbc.com excerpt is Tesla uses human video labelers to review the 10 million clips of video to find the video where a particular situation was handled well. With this higher quality driver bias, the AI produced would be an above average driver. This method of training driving is similar to ChatGPT chatbot for the writing written text. The AI will drive like what a human driver would do in any given arbitrary situation once it’s trained with a sufficient target video data set that covers a wide range of driving situations. The video collection over time has the potential to cover more situations than one person may encounter in a lifetime. That can lead to AI driving safer than a human. This method will get better faster than the prior FSD version based on the rules. Finding missing rules and then implementing them in C++ programming language takes skill and effort. The AI method finds the corner case video and train the AI. No new code is needed and is very low effort. The programming effort is replaced with computation. Project dojo and the new Nvidia H100s they have recently installed will provide the computation Tesla needed.