AI is real. But the LLM-token economy still looks like a bubble. The money flow is simple: Enterprise → LLM → GPU → Memory Investors are moving upstream in search of certainty. No one knows which AI app will win, so they buy the LLM labs. No one knows whether OpenAI, Anthropic, Google, Meta, or xAI will win, so they buy NVIDIA. No one knows how durable GPU demand is, so they buy memory. Every layer looks safer than the one below it. But that certainty is an illusion. The entire chain is still funded by the layer closest to real ROI. Enterprises pay for tokens because they are trying to prove LLM adoption works. But so far, LLM productivity gains have not clearly translated into revenue growth. Consumers are not obviously buying more. Expenses are not obviously falling either. Most AI-related layoffs look more like companies using LLMs to rationalize previous overhiring. The layoffs truly driven by “AI efficiency” often create backlash, operational problems, or quality issues. When hallucination is still unsolved, critical work still needs human supervision. So the economics are awkward: no obvious revenue lift, no obvious expense reduction, and a new token bill on top. Meanwhile, customers are not receiving much of the surplus. They are not getting better products at lower prices. They are getting higher prices, worse content, and weaker job security. Personally, I hate this most when I see gaming consoles getting more expensive, PCs getting more expensive, Macs getting more expensive, and the internet filling up with lower-quality AI slop. If consumers do not spend more, the companies selling to them cannot justify ever-growing LLM expenses. And enterprises do not need to abandon LLMs for the chain to break. They only need to slow the growth of LLM spending.

Vibe coding is no longer an AI skill. Your agent team must autonomously identify opportunities, assemble a team, file issues in your repo, draft PR/MRs, review them, merge them, test them, and deploy them. If your workflow still requires human involvement, then you have a skill problem, and you can do better.

I actually think SpaceX has a fair price now (~$2T), based on simple math using publicly known information that everyone already has access to. And since I also did some secret research (lol), my take on the win rate of certain core projects might be different from the market’s take. As a result, personally, I think it is actually very undervalued and a good deal now.

Yes, Tesla is deploying Robotaxi very slowly for some reason. However, Tesla is constantly producing new Cybercabs, and those are filling parking lots at a stable rate. Recently, new Cybercabs reportedly no longer have steering wheels, according to @tesla2moon. If the large-scale deployment timeline is unclear to Elon, then why doesn’t he pause or slow down Cybercab production, or at least add steering wheels to them? In fact, if you think about Robotaxi long term, the scaling rate will probably be bottlenecked by the manufacturing rate for years. So I suggest that, in the early stages, we look at the production rate of Cybercabs, especially those without steering wheels, to estimate the actual “scale” of Robotaxi, instead of only looking at those running on the road. This might reduce short-term noise.

Even if we give them every possible benefit of the doubt and assume FSD was active, why didn’t the driver disengage or intervene? The car was clearly speeding, and that was an obvious warning sign that it was time to disengage. Was the driver asleep with his foot on the accelerator? If the driver were trying to make up a story, claiming the brakes failed would’ve been a better lie. Saying “self-driving was active” sounds more like an attempt to avoid responsibility.

My Cybertruck, on full FSD, in downtown Los Angeles, suddenly swerved to the left and narrowly avoided a collision with the white Audi car changing lanes very fast. If I were the driver, I wouldn’t have been able to prevent the accident. Views from three different cameras of the CT. @elonmusk @Tesla_AI @Tesla @MarioNawfal @dallasteslaclub #FSD# #Cybertruck# #Tesla# #Robotaxi#

FSD v14.3.4 just saved us from a maniac. The camera does not do the situation justice. I don’t believe any human could have done better than FSD here. This was superhuman driving skill.

I gave it more thought, and I think I found all the answers. I have to say SpaceX’s story is much harder to comprehend compared with Tesla’s. The short answer is that 1) inference is trending toward becoming the majority of future compute expense. One day, training in space might also make economic sense. 2) Sending AI chips to space does not make compute cheaper in absolute terms, but during a severe shortage, it can make adding new compute less expensive relative to ground-based alternatives. And 3) to make the AI story complete, Terafab must succeed in solving the production capacity issue around memory and packaging. Hopefully, faster iteration is the key to addressing that challenge, but the risk remains.

Upcoming releases of FSD will remember your parking preferences, so that the car goes to the right location at your home, office, school drop off, etc. Destination parking is by far the biggest reason people now intervene with FSD. Critical safety interventions are extremely rare.

Please correct me if I’m wrong, but here are my thoughts about AI satellite from SPCX: 1. Training/inference: They are kilometers away from each other. Due to physics limitations, the latency and bandwidth between satellites will make them better suited for inference only, not training. 2. Energy saving: Energy cost for data centers is usually less than 10%. The majority of data center cost today is still depreciation of AI chips. Therefore, even if you consider energy cost as zero, I still have trouble seeing significant cost efficiency when comparing on-ground data centers with in-space ones. 3. Production capacity: If AI chip manufacturing speed is the scaling bottleneck, then launching AI satellites does not unblock scaling, since the satellites still need to wait for the chips. Otherwise, the plan to have 10,000 Starship launches per year becomes the bottleneck, and at this speed, it still does not significantly outpace the expected compute scaling speed on the ground within five years.

Tesla Robotaxi was at least ~1.81x safer than humans in April 2026. With the latest NHTSA crash report, and thanks to Robotaxi Tracker, I can finally report that Tesla Robotaxi in Austin surpassed average human safety in April 2026 (last month). 1. According to NHTSA, I estimate the human crash rate, including both reported and non-reported crashes, is ~249K miles per accident. 2. I collected all Tesla-reported mileage, as well as 7-day active fleet size from Robotaxi Tracker, and found the mileage is pretty predictable. This allows me to accurately estimate monthly Robotaxi mileage in Austin with <5% error. 3. Using the latest NHTSA crash report, I calculated the 3-month rolling crash rate, since Tesla sometimes has 0 monthly crashes. It clearly shows that in April, the 3-month rolling crash rate surpassed humans for the first time.