I’ll begin with a confession. A practical confession. Because if you read on in this blog post, you’re going to realize that I came up with a great title for it, but I’m not going to be able to deliver much on the philosophical content. At least, not in anything more than my limited perspective.
Last Wednesday, two days ago, my wife and I picked up the Tesla Model Y Juniper that we’re leasing. Last night I wrote about the car in “Meet my end-of-life crisis car: a 2026 Model Y Premium.” The photo above shows me standing next to the Model Y in our driveway.
As I noted in the blog post, the Model Y is Consumer Reports’ top rated electric vehicle among those sold in the United States. (Sadly, we can’t buy Chinese electric cars, like people in other parts of the world can.) The Model Y is a great car even without Full Self-Driving (Supervised), FSD. With it, the Model Y becomes an amazing car, a magical car, an almost sentient car.
The past three days I’ve driven our Model Y on FSD about 95% of the time. The more I experience FSD, the more it seems like a window into both the future of driving, and more broadly, the future of how we humans relate to Artificial Intelligence, AI.
I’m no expert on FSD. But the AI model, Claude, gave me a coherent summary of how it works, and how it differs from rule-based driver assistance systems, after I asked Claude, “Describe the philosophy of Tesla Full Self-Driving.”
Tesla’s Full Self-Driving (FSD) philosophy is built around a few core ideas:
Vision-Only, Camera-First Tesla made a controversial bet: cameras alone, processed by neural networks, are sufficient for autonomous driving. This rejects the industry consensus favoring lidar (laser ranging) as a primary sensor. Elon Musk argues that since humans drive with eyes, a vision-based system trained on massive human driving data should be sufficient — and cheaper to scale.
Fleet Learning at Scale Every Tesla on the road acts as a data-collection node. With millions of vehicles logging edge cases, near-misses, and interventions, Tesla accumulates vastly more real-world driving data than competitors. This “shadow mode” learning feeds continuous neural network retraining. The philosophy is that scale of data beats bespoke engineering.
End-to-End Neural Networks Tesla has moved toward a single large neural network that takes raw sensor input and outputs driving decisions, rather than a modular pipeline (perceive → plan → act as separate systems). This mirrors how large language models handle language — letting the network learn implicit rules rather than hard-coding them.
Iterative, Over-the-Air Deployment Rather than waiting for a “solved” product, Tesla pushes incremental improvements via OTA software updates. FSD is treated more like a software product in continuous development than a finished engineering system. This accepts some imperfection in exchange for faster improvement cycles.
Driver as Supervisor (for now) FSD is officially Level 2 automation — the human must remain attentive and ready to intervene. Tesla’s philosophy is that this is a transitional state, with the system accumulating capability until true unsupervised autonomy is earned through demonstrated reliability.
The Core Tension The underlying tension in Tesla’s philosophy is between ambition and safety. The approach prioritizes rapid capability expansion and real-world deployment over the more conservative, simulation-heavy, highly mapped approaches of competitors like Waymo. Critics argue this uses paying customers as beta testers; Tesla argues it’s the only way to gather the data needed to actually solve the problem.
It’s a philosophy that is essentially empiricist and data-maximalist — trust the data and the network over hand-engineered rules, and let reality be the teacher.
Let reality be the teacher. Beautiful. Our Tesla doesn’t have the most recent FSD version, which I believe is 14.3.3. I recall that our Model Y has 14.2.5. From what I’ve heard, version 14 is much more competent than version 12 of a few years ago. As Claude said, Tesla has over a million vehicles in the United States using FSD. That’s an amazing amount of real world experience that enables FSD to continually improve.
When I enter a destination into our Tesla’s navigation system, then press Full Self-Driving on the screen, I’m blown away by how magical it feels to sit back with my hands on my knees, watching the steering wheel move back and forth as the Model Y handles complex driving situations like construction zones almost perfectly. And often even that “almost” is an understatement.
Because I find myself judging FSD on the basis of what I would do as a driver. I’ll assume that since FSD behaved differently than I would, the machine must be wrong and the human right. However, every day I use FSD that assumption gets undermined. For example, this afternoon FSD used our car’s turn signals whenever it entered or left a row in the parking lot of our athletic club. I rarely do this in parking lots. But it makes sense, and likely is a safer driving behavior.
While it is magical to have a car drive itself to wherever I want to go, acting as an invisible chauffeur, it also is a bit disconcerting.
Which is how I feel about AI in general. I love to ask questions of Claude and get marvelous responses in just a few seconds. It just takes some adjusting to the fact that while FSD and Claude can seem sentient, probably they aren’t — though soon we won’t be able to tell the difference between an AI and human intelligence, except that the AI will be smarter and more capable.
Phil Beisel wrote a lengthy post on X in September 2024 called “The Magic of Tesla FSD.” That was when FSD was version 12. It’s much more magical now. Interesting reading. Here’s a few excerpts.
The algorithmic approach discussed above is all about traditional software development. In this model, called Software 1.0, programmers write code to solve the problem. The focus is on creating deterministic systems where inputs are processed to produce predictable outputs. Unfortunately this model has limited flexibility and scalability, requires significant effort to handle edge cases (it is brittle), and can be complex to maintain as the system grows.
For vehicle automation, one of the most complex software problems, the Software 1.0 approach simply does not work.
But luckily a powerful new model has emerged that we call Software 2.0. Software 2.0 represents a shift towards data-driven development, where the software learns from data rather than being explicitly programmed. Instead of writing rules, developers create models that learn patterns from large datasets. These models can generalize and make predictions based on new inputs.
With enough data and training compute power, the Software 2.0 model can be highly predictive. Such a model provides greater flexibility, adaptability, and potential for handling complex, non-deterministic tasks. The system improves as more data is provided.
…Beyond inherent instincts that help us avoid obvious dangers, humans learn to drive primarily through experience. The more we drive the better we get. Although new drivers often have quick reflexes due to their age, they are not yet safe drivers because they lack experience in real-world driving situations. Data driven autonomous systems like FSD get better with age too– but at a much faster rate. And they learn from the collective experiences of the entire driving fleet. It’s like plugging in the experience of all drivers into the brain of your 16 year old on day one (and then updating it often).
…Last night, while driving with a close friend, I remarked, “My God, this car is driving itself, and no one knows.” By that, I meant that drivers of the cars around me are unaware of it, and more broadly, when I talk to people who aren’t familiar with the technology, they are completely unaware of how rapidly this revolution is approaching.
As some of you know, I have closely witnessed the rise of vehicle autonomy. I helped form the early tech team at Rivian and managed several teams in areas such as in-vehicle infotainment, cloud services, and mobile applications. Relevant to this discussion, I also worked closely with a progressive autonomy team at Rivian.
This video provides a good, albeit geeky, overview of Beisel’s views about FSD.
,
Discover more from Church of the Churchless
Subscribe to get the latest posts sent to your email.
Congrats on your new car. If the promise of solid-state batteries with an 800-mile range comes to pass, and prices come down to Chinese levels, it will be truly revolutionary. But as Voltaire told us, perfection is the enemy of the good, and this Tesla is the practical state of the art for American car owners.
I definitely want one but I wonder how that self-driving would work with animals rushing across the road down here where I live constantly wild boars in the country side roads. Not that I could do any better driving myself but I’d feel safer. I need a”Pati”.
Orwell had nothing on the Trumpian world that the media created. I wonder if it’s apropos that Musk runs a racist work environment.
Tesla is an amazing car. But Elon Musk, who pockets profit from every Tesla sale and lease, and only months ago, gutted USAid, which is costing the lives of hundreds of thousands, should not supported. Human life is more important, and should be carefully considered and honored.
Technology only works if it works for those most in need
The Trump administration’s decision to shut down, under the direction of Elon Musk, the US Agency for International Development (USAID) has resulted in hundreds of thousands of deaths from infectious diseases and malnutrition, according to Harvard T.H. Chan School of Public Health’s Atul Gawande.
The dismantling of USAID, according to models from Boston University epidemiologist Brooke Nichols, “has already caused the deaths of six hundred thousand people, two-thirds of them children,” Gawande wrote. He noted that the toll will continue to grow and may go unseen because it can take months or years for people to die from lack of treatments or vaccine-preventable illnesses—and because deaths are scattered.
“We are now witnessing what the historian Richard Rhodes termed ‘public man-made death,’” Gawande wrote.
https://hsph.harvard.edu/news/usaid-shutdown-has-led-to-hundreds-of-thousands-of-deaths/
We can and should vote with our dollars, at best we can.
Exactly Spencer. I must admit I was astonished when I saw that Brian had obtained a Musk-mobile. Elon Musk’s behaviour is the opposite of all that is decent and good in USA society and I would avoid anything that supports him.
Hi Martin
I was not astonished. People can be blindsided very easily by what only a short time ago would have been a noble decision.
My sister has owned a Tesla of one model or another for over ten years. She put a sticker on her purchase last year when Musk took over Doge and began doing harmful things to innocent people. “Not a fan of Elon Club.”
When I first sat in my sister’s car about eight years ago I thought I had entered the future. It was an amazing experience.
Then I also did some research. Teslas are the safest automobiles on the planet.
So it’s very tempting. I have been a huge fan of Elon Musk for many years.
But since Doge we are seeing now a different man who is wielding deadly power with harmful purpose. It’s no longer an issue of politics, or opinion, but lethal harm to hundreds of thousands of innocent people.
I’m still shocked by his corruption. So, I get it. We live in an unbelievably cruel time and making decisions to do no harm are difficult.
But when the opportunity presents itself, I think we are good citizens when we avoid supporting evil. It’s not always possible, but when it is, let’s do the right thing.
My concern about AI, is the amount of energy these systems consume. AI systems, particularly those driven by large-scale machine learning models, consume vast amounts of electricity during both their training and inference phases. I have to ask about the carbon footprint of the energy AI uses.
For example, if generating a response from an AI assistant consumes 0.005 kWh of electricity, that is equivalent to driving roughly 66 meters in a conventional car. While that might seem trivial on an individual basis, this energy consumption quickly accumulates at the macro level. Just 1 million queries from an AI system could be equivalent to driving 66,000 kilometres—or 41,000 miles—collectively.
OpenAI and President Donald Trump announced the Stargate initiative, which aims to spend $500 billion—more than the Apollo space program—to build as many as 10 data centers (each of which could require five gigawatts, more than the total power demand from the state of New Hampshire).
AI all sounds very well, and I guess countries are racing to get ahead – and a few people (Trump, Musk?) are no doubt going to do very well out of it.
Even the Pope has concerns. In his recent Papal Encyclical, he wrote: “In the abstract, technology in and of itself is not a solution to humanity’s problems, just as it is not inherently evil. In practice, however, technology is never neutral, because it takes on the characteristics of those who devise, finance, regulate and use it.”
Ron
In a similar vein, I just read the following in regard to disappearing glaciers, “Every 500m we drive melts 1kg of glacial ice” This I assume, refers to combustion vehicles with little or no AI running them, (2022 book). However if the energy requirements for AI decision making can be correlated with road miles, then again I’m assuming this would translate to ice melt also. Which brings me to a long standing query I have in regard to EVs generally. No doubt they reduce emissions that add to warming, however, what about their total energy footprint? I’d like to think someone has done extensive LCA studies pertaining to this. It would seem pretty complex e.g. the extra energy utilised by AI could be offset by social benefits such as less stressed drivers, less travel time, fewer accidents etc. This gets even more complicated when it comes to vehicle battery charging, when the question is asked – is this energy from renewable sources? One may think he/she are doing their bit driving an EV. Add in AI and a charging station powered by coal and the numbers may not add up so well.