Get live statistics and analysis of Yun-Ta Tsai's profile on X / Twitter

Turn out you don’t need to tax citizens that much if the government is not ran by fraud.

I just got home, the first thing my wife asked me seriously to get a gun license. “Is that bad?” Then she showed me all the text messages from her friends in West LA with all the chaos right now.

I cannot think of any company hold the shareholder meeting at the factory. For most companies, they bring people to the design house offices while the factories fade into the background. But for Tesla, factories are the heart of our stories.

Many of my friends in AI companies thought we were just a car company. Yet, I personally have more compute quota than their entire teams. Thanks for all the ml-infra teams keeping it up 24/7.

Years ago, I did not bother to pay $1200 for Google Glasses that my team was working on. This year, I did not hesitate to buy the new Model Y on day one. If an employee willing to pay full price for their own product, it is usually a signal of a good product. See you guys at Fremont.

My wife asked me why I did not go to work on LLMs for more money. “I just want to save lives.” And the objective function of self-driving is saving lives and nothing else.

Designing an inference chip for robots is actually very difficult. In data centers each chip is bathed in jacuzzi and babysat by nannies. If they died it would be hot swapped by one of their clones. The fault rate of GPUs in datacenter is actually quite high. Industrial average annual fault rate of H100 is 9%. Ideal conditions could reduce it down to 2% but never below a single digit. The fault recovery of GPU nodes actually could take a while, from minutes to hours. It is not instantaneous. In robots, the chips are out in the cold and they need rapid self recovery. The fault tolerance is in a different league. It is not uncommon many robotic companies struggle to get the chip running more than a few hours without rebooting. For chip companies, this is great since they would tell robotic companies to buy more chips for hot swapping. For robotic companies, this is bad since it is obviously not a scaleable solution but they are stuck with endless back-and-forth JIRA tickets with vendors.

Sorry we are busy at launching Robotaxi and our devotion to the sustainable abundance.🙂

It is often overlooked that there is a 1.4T AI company with positive net income, operating multiple giga 🏭 in 🇺🇸, in-house chip design team, owning one of the largest and densest computing clusters backed by its 🔋, best in class AI team, iterating models faster than most frontier labs that has been deployed on actual 🤖 you could buy today. Most importantly, it is led by a CEO that is not afraid to answer questions, cares about retail shareholders and deeply love this country and humanity. Plus, it has the coolest annual shareholder meeting 😎, and you can vote!

A few days ago, I casually showed this meme to a person in the entertainment industry — jokingly begging they don’t ruin the next good film I love. The response was, “What’s Andrej credential to criticize this film?” I replied, “No, just people who used to enjoy the films.” Then I realized the whole movie industry was so deep in their ass that they have completely detached from their high margin paid customers — movie goers. Their Slack channels had a lot of “team building” for morale raising but very little on reflections why story was not good anymore. No wonder films sucks now.

There is no sci-fi movie shot as beautiful as hot staging.

The seed of Galatic Empire.

Movie poster vibe, except it’s real.

We have been working hard for this moment. Congratulations to the teams.

After years of dinners with FAANG friends, you experienced the progression of from being seen as nobody to a tier one AI company. They always joked, “Are you still working on Full Self-Driving?” Then you sat there quietly eating your own food. For many, top AI companies means hosting the biggest parties at CVPR, writing lots of papers, and have well known AI “god parents” in the org. We have none of these. Quietly, we just want to solve singular most difficult problem on Earth. This year they all have recognized Tesla is doing AI as good, if not better, as any other frontier companies, not because numbers of papers or podcasts being published, but a solid product that delivers the magic.

I cannot think of any company hold the shareholder meeting at the factory. For most companies, they bring people to the design house offices while the factories fade into the background. But for Tesla, factories are the heart of our stories.

I just got home, the first thing my wife asked me seriously to get a gun license. “Is that bad?” Then she showed me all the text messages from her friends in West LA with all the chaos right now.

Many of my friends in AI companies thought we were just a car company. Yet, I personally have more compute quota than their entire teams. Thanks for all the ml-infra teams keeping it up 24/7.

The seed of Galatic Empire.

My wife asked me why I did not go to work on LLMs for more money. “I just want to save lives.” And the objective function of self-driving is saving lives and nothing else.

Testing the same route, I am now actually driving 15 mins slower in average on manual than Full-self Driving. Full-self Driving not only reduces stress but saves me time. That’s 30 mins of saving per day.

Turn out you don’t need to tax citizens that much if the government is not ran by fraud.

It is often overlooked that there is a 1.4T AI company with positive net income, operating multiple giga 🏭 in 🇺🇸, in-house chip design team, owning one of the largest and densest computing clusters backed by its 🔋, best in class AI team, iterating models faster than most frontier labs that has been deployed on actual 🤖 you could buy today. Most importantly, it is led by a CEO that is not afraid to answer questions, cares about retail shareholders and deeply love this country and humanity. Plus, it has the coolest annual shareholder meeting 😎, and you can vote!

Years ago, I did not bother to pay $1200 for Google Glasses that my team was working on. This year, I did not hesitate to buy the new Model Y on day one. If an employee willing to pay full price for their own product, it is usually a signal of a good product. See you guys at Fremont.

Reality is like a spring. The more you displace it, the stronger it will coil back. F = -kx

I had a recent conversation with both Stanford professor and another professor from Tsinghua separately. The Stanford professor kept showing me the Ghibli and toy packaging gen from GPT that maximized his dopamine. To him, this is the pinnacle of technology. I asked him what he thought on the Full-self Driving, he was not amazed, and he told me his colleagues did not believe it’s going anywhere either. The professor from Tsinghua was the opposite end of the spectrum. His main research focus had been shifted toward robotics, and he was extremely impressed by the Full-self Driving videos on Weibo. What do you think how these two professors indirectly or directly influence their students on their research topics? And how big of the gap would that be many years from now? Every new PhD students would have the influence of technology breakthroughs 5-10 years in the future. If we are incubating our brightest and smartest PhD students at > $80K per year in elite schools, should we not encourage them to think bigger?

Levels of spouse tests: ✅Level 1: take my wife on FSD, while I am on the driver seat ✅Level 2: let my wife use FSD, while I am on the passenger seat ✅Level 3: my wife uses FSD without me, but only on high way ✅Level 4: my wife uses FSD from door to door ✅Level 5: my wife purchases FSD on a new car

On-demand chauffeur, powered by the real world AI.

Designing an inference chip for robots is actually very difficult. In data centers each chip is bathed in jacuzzi and babysat by nannies. If they died it would be hot swapped by one of their clones. The fault rate of GPUs in datacenter is actually quite high. Industrial average annual fault rate of H100 is 9%. Ideal conditions could reduce it down to 2% but never below a single digit. The fault recovery of GPU nodes actually could take a while, from minutes to hours. It is not instantaneous. In robots, the chips are out in the cold and they need rapid self recovery. The fault tolerance is in a different league. It is not uncommon many robotic companies struggle to get the chip running more than a few hours without rebooting. For chip companies, this is great since they would tell robotic companies to buy more chips for hot swapping. For robotic companies, this is bad since it is obviously not a scaleable solution but they are stuck with endless back-and-forth JIRA tickets with vendors.

Saving lives is our only objective function. Come work with us to save more lives.

Historically, it was difficult to reinvent a company at the height of its market cap. When I was in Nokia, I asked why my Finnish colleagues why the company did not reinvest its resources to build the actual smartphone — event its concept video predicted the vision correctly years ago. The answer was, it would take a long time to rewrite Symbians to support a new paradigm than changing the mechanical design and plastic molding of a phone. The risk was simply too high — new software, tooling and backend infrastructure, while it was under pressure to release new models every few months to fan off the competition from Asia. Quarters after quarters, Nokia enjoyed massive market shares that it was not in a hurry to change until it was too late. Looking back, the answer seems obvious. Yet, it was difficult to commit to the right path. Reinvention often requires repeatedly going from zero to one. That was hard to stomach to most people.