Have a happy 4th everyone. When I woke up this morning early and read my email I received an email as a subscriber to Phil Koopman’s blog. Many of you will know that he is considered by others as an expert on autonomous safety.
He cites here a study done by a University in Australia evaluating FSD.
I offer no commentary other that if one simply observes the accumulation of statistics in the community FSDTracker that it’s easy to agree with these comments.
Following this except from him is a link to a published article from yesterday in an Australian publication,
“Tesla Full Self Driving (supervised) used in the wild for 100 days by researchers from the University of Queensland. Here’s what they found: Summary: quite capable – except when it’s not. 500 safety-critical interventions in 100 days. "It occasionally struggled with situations human drivers handle almost automatically, making mistakes that were surprising and sometimes dangerous because they weren’t the kinds of errors an experienced driver would make." "most of the incidents we saw weren’t rare edge cases, but everyday situations." Problems observed with: road markings, road layouts, time-restricted signage, rail crossings, zipper merge, roundabouts, distinguishing e-scooter riders from pedestrians, poor visibility. The authors argue for better street design & signage to support this technology. I would say that addressing the automation complacency/driver attention management problem is crucial before steering public funds to favor supervised automation technology. The finding of surprising mistakes could present significant risk to at-scale use with ordinary drivers.”
Yes, Professor Phil Koopman is widely recognized as one of the world’s foremost independent experts on autonomous vehicle (AV) and embedded software safety. His recent writings and subscriber emails from his Autonomous System Safety newsletter focus on a critical shift in the industry: why simply tracking “crashes per mile” is an incomplete and flawed way to measure AV safety. [1, 2, 3, 4, 5]
If you are looking at his recent subscriber updates, he has been heavily emphasizing a few core arguments:
1. “Safe Enough” is More Than a Numbers Game
AV companies frequently release data arguing they are statistically safer than human drivers. Koopman contends that true safety and societal acceptance require much more than a net statistical average. He highlights that acceptable safety must include: [1, 2]
No Risk Hot Spots: An AV cannot shift risk entirely onto vulnerable road users, such as pedestrians or children getting off school buses, even if total vehicle occupant injuries drop. [1]
Eliminating Reckless Behaviors: Even if it doesn’t cause an immediate crash, an AV consistently blocking a firehouse entrance or running red lights represents a severe safety failure. [1]
The “Would a Human Do Better?” Test: Societal trust is broken when an AI makes a bizarre error or handles an “edge case” in a way that a normal human driver never would. [1, 2]
2. The Illusion of Full Autonomy
Koopman frequently demystifies the role of remote operators and hidden human fallback systems. He points out that if a vehicle relies heavily on a “phone home” system to get out of complex construction zones or edge cases, it isn’t operating on a fully autonomous model—it is operating on a distributed authority model that introduces new safety-critical infrastructure risks. [1, 2]
3. Regulatory Gaps vs. Hype
His recent commentary also focuses on the lagging nature of federal regulations. While bills like the SELF DRIVE Act of 2026 struggle to find momentum in Congress, current federal standards (FMVSS) only regulate physical vehicle components (brakes, tires, airbags) rather than the safety-critical software and AI models dictating the vehicle’s driving behavior. [1, 2]
Those AI comments do reflect Koopman’s views on autonomy in general and some certainly reflect concerns he has also with Waymo.
But are you failing to recognize that Koopman is just the messenger here. The real issue is that pointed out by the Australians regarding the nature and qty of really basic issues with FSD in the article.
i’m failing nothing. I use a totally different outlook on the issue. Not perfection, just investing. You want to be perfect? Be my guest!
Perfection is the enemy of the good.
GoogleAI:
“Perfect is the enemy of good” is a famous aphorism reminding us that the relentless pursuit of flawlessness often prevents us from completing valuable work. Insisting on absolute perfection can lead to analysis paralysis, missed deadlines, and overwork, making an attainable “good enough” outcome far more valuable than an unfinished perfect one. [1, 2]
Key Concepts Behind the Phrase
Origin: The proverb is widely attributed to the French writer and philosopher Voltaire, who famously wrote, “Le mieux est l’ennemi du bien” (translated literally as “The best is the enemy of the good”). [1, 2]
Progress over Perfection: Pragmatic application encourages shipping a product, finishing a project, or executing a task once it reaches a highly functional state, rather than endlessly tweaking minor details.[]
How It Impacts Productivity
Decision Making: Avoids costly delays caused by trying to account for every possible variable or edge case.
Motivation: High, unachievable standards can trigger avoidance and self-criticism. Focusing on “good” results creates achievable goals and sustained motivation. [1]
The Law of Diminishing Returns: The amount of time and effort required to take a project from 90% (good) to 100% (perfect) is often exponentially higher than the effort it took to get there. [1]
When to Make Exceptions
While it is a great rule of thumb for day-to-day productivity and project management, as discussed on Reddit, there are specific scenarios where excellence is mandatory. Professions involving public safety, aviation, structural engineering, and medicine require a perfectionist approach to ensure lives are not put at risk.
are appreciated as they point out where improvements might be found but don’t loose any sleep over it.
This highlights a major risk: how well/poorly Tesla’s AI method generalizes across driving scenarios for human or better safety (eg, US vs Australia etc, or even region to region within a country).
We really don’t know.
There is no evidence that it does generalize well, but they haven’t yet offered a service accumulating millions of miles.
So there is no viable service yet existing that can be generalized.
The recent Reuters article on intensive efforts by Tesla to train their AI driver in Austin was not a good sign for the model’s ability to generalize, but that was just a little tidbit of information.
Neural networks are flexible models, so they generalize in that sense, but that can come with a cost of requiring a lot of data (and then of course compute, like HW3, HW4, HW5, etc).
And then do you need data that is hyper-specific to each city and weird case?
Really? Do you think it’s that simple? His AI generated concerns point out issues with the state of autonomy in a general sense. You seem to fall back on this as I read this apparently in an attempt to obfuscate the serious issues specific to Tesla that the Australian research team is pointing out. Meanwhile I see the community based FSDTracker continuing to point out very similar issues as the Australians from all over the US.
More obfuscation. This isn’t about pursuit of Perfection- this is really about serious issues affecting the safety of riders, the safety of other vehicles on the road and innocent pedestrians
Nonsense answer. Of course nothing is “perfect’. But it’s going to have to be a lot better than human driving before people relinquish control, just a marginal improvement in the statistics isn’t going to impress Joe Six Pack enough to take his hands off the wheel and put Little Suzy in harm’s way - even if doing so might improve his chances by 0.1%. It’s the Dunning Krueger effect: everybody thinks they’re a better than average driver already.
Tesla has approximately 1.1 million active FSD users. This milestone metric was revealed by Tesla during its Q4 Earnings Call. [1]
Here is the breakdown of how these users access the software:
Outright Purchases: About 70% of this total base (~770,000 drivers) bought the FSD package outright.
Monthly Subscriptions: The remaining 30% are on a recurring monthly plan.
Overall Take Rate: This 1.1 million user base represents an adoption rate of roughly 12% across Tesla’s cumulative vehicle fleet. [1, 2, 3]
Tesla also transitioned the service away from the $8,000-$15,000 upfront purchase option, making FSD available strictly as a monthly subscription
I didn’t follow up the headline: “FSD available in Germany”
GoogleAI:
Tesla’s Full Self-Driving (FSD) (Supervised) is officially available in Germany. The Federal Motor Transport Authority (KBA) adopted the provisional EU type approval originally granted by the Dutch RDW, allowing Tesla to switch on FSD subscriptions in the country. [1, 2]
The rollout features the following actionable details for German owners:
Subscription Model: The one-time purchase option for FSD has been discontinued, making subscription the exclusive way to access the features. [1, 2]
Pricing: The KBA approval makes the package available for €99 per month. Owners who previously purchased Enhanced Autopilot (EAP) receive a discounted rate of €49 per month. [1, 2]
Compatibility: The software is delivered via update 2026.8.6 and is compatible with both Hardware 3 (HW3) and Hardware 4 (HW4) vehicles, though HW4 will deliver a smoother performance.
GoogleAI:
Tesla Full Self-Driving (FSD) Supervised is officially approved and available in 12 global markets, including the United States, Canada, Mexico, Puerto Rico, China (limited), Australia, New Zealand, South Korea, and the European countries of the Netherlands, Lithuania, Estonia, Denmark, and Belgium. [1, 2]
Global FSD Availability
The deployment of FSD is localized, meaning its capabilities are adapted to the specific road rules, signs, and infrastructure of each region. [1, 2, 3]
North America: Fully available across the United States and Canada, with availability also extending to Mexico and Puerto Rico.
Europe: Rapidly expanding following landmark approval by the Dutch RDW, allowing Tesla to bypass long individual country testing. The system is actively rolling out in the Netherlands, Denmark, Lithuania, Estonia, and Belgium.
Asia & Oceania: Available in Australia, New Zealand, South Korea, and in a limited capacity in China.[1, 2, 3, 4]
Pending Approvals
Tesla is currently awaiting regulatory approval or has localized testing programs underway in dozens of other markets, including the UK, France, Spain, Norway, Sweden, Portugal, and Ireland. [1, 2]
Seems that Goofy6Pack likes FSD, warts and all. Between squeamishness and cashflow, I’ll take cashflow as do over a million Tesla owners.
1.1 million users
$100 per month
12 months per year
$1,320,000,000
Looks more like a software company than a car company.
The overwhelming majority of the FSD users aren’t paying Tesla any money. They bought the package outright, back in the day. Only about 330K users are subscribing. Which would mean about $400 million in annual revenue.
Which, fine - but Tesla’s annual revenues are about $100 billion. Meaning that FSD revenues would represent less than half a percent of Tesla’s overall revenues.
So, no - looks much more like a car company than a software company right now. Maybe that changes one day, if Robotaxi moves out of open beta? But for now, their software revenues are a rounding error.
Depends how narrow of view you take in terms of the application of autonomous transportation. Farming, commercial freight, school buses, last mile inventory delivery and so on.
All true. But, when your investing period is indefinite and you see characteristics of a business that mimic prior patterns of success, you may speculate with some capital. It has paid off handsomely so far.
My personal investment theory never hinged on Robotaxi or even FSD as the primary catalyst of revenue and income for the company. I saw it more as a product differentiator within the auto manufacturing space. Having said that, I recognize it opens up optionality and another addressable market.
Other investors are free to assess the situation differently. It would be perfectly reasonable to require more proof as to how it may impact the future economics of the business. This is what makes a market. It is quite possible the share price for the business moves wildly in either direction, until Tesla proves or fails on the concept of autonomy.
What is interesting is watching it play out in a very public way. Most inventors fail in private over long periods while trying to make something work. I doubt large portions of the public stood around the beach at Kitty Hawk telling the Wright Brothers, you didn’t achieve controlled flight again.
Of course, Musk has brought much of this upon himself. But, it is interesting and at the same time to me mostly a waste time to speculate one way or another whether Tesla solves FSD or that interim failures indicate long term failure. Although I’ll admit, it is fun to speculate on the subject.
They either figure it out or they don’t. I think what most critics underestimate about Tesla is the importance of its culture. It has almost no ego in terms of design and problems solving approaches. When one roadblock emerges, the company is willing to quickly pivot in another direction to reach success. This is a philosophy that is incredibly difficult to maintain for most large companies. It’s why one of the elements critical to Musk’s management style is a focus that encourages failure as a key part of problem solving.
Except when t comes to Lidar, apparently, when every other firm pursuing true self-driving is using it (including more than a dozen firms in China and elsewhere) and Tesla won’t, because someone decided years ago that it was too expensive.
Meanwhile, I can now buy a self-driving lawn mower which comes with Lidar, and I won’t be surprised to see it in other applications, like perhaps robotics.
I think I would do a little more investigation if I were you and actually listen to interviews with Musk and top engineers at Tesla. The company is looking for the simplest solutions. Combining technologies adds a layer of complexity in building, maintaining, and order of decision making with regards to autonomous driving. This isn’t the proverbial morally bankrupt capitalist looking to save 15 cents on a part to boost the corporate profits. It is a huge bet and if they fail to deliver, others that are using LIDAR will have turned about to be correct. There is nothing that says Tesla wouldn’t pivot to something else, if they turn out to be wrong. Which is really the odd part, Musk is committed to this as the current direction of the company, but it is not a commitment in perpetuity.
Tesla is not going to provide a universal solution nor is the suggestion LIDAR might not have a place. But, back in the late 1990’s while plasma TV’s were the cutting edge of technology at $10k, it’s ultimately LCD and LED that provided the mass solution and at a fraction of the cost and lower energy output.
A LIDAR mower costs aroun$1,000 and then a the top end $3500 and can handle about a 1/3 of acre before requiring a re-charge. Like the early Plasma TV’s, it’s a small market.
Like I said, what I think many don’t understand about Tesla is the culture. Being wrong is a frequent and intentional part of how the company achieves long term success and making a difference to humankind.
That, in a nutshell, is Tesla bears’ problem, they think Tesla should act like all legacy car makers act. Tesla had a clear progression in mind.
The Roadster to see if they could build a desirable EV
The luxury models X and S
The more affordable Models 3 and Y
The mass market model 2 which was replaced by the CyberCab
Use the AI technology to power a humanoid robot
All the while making sure that there was sufficient free cashflow to finance the next stage. The FSD pricing was pure genius. The market for luxury cars would be willing to pay the huge price of lifetime FSD at a time when free cashflow was yet uncertain. As the models went cheaper mass markets that FSD pricing was no longer feasible so it was replaced by a monthly subscription. As Tesla went down market it got rid of the low volume luxury models switching to CyberCabs and Optimus Robots.
Once Packard was the prime luxury car and what destroyed the brand was the cheaper Packard Clipper. The Clipper was a lovely car but it destryed the lustre of Packard as the exclusive luxury brand.
GoogleAI:
The dilution of a prestige brand through downmarket expansion is a classic cautionary tale in automotive history. While the Packard Clipper was a critical success, its long-term impact on the brand’s elite status aligns perfectly with your observation. [1]
The strategic shift
Introduced in 1941, the Clipper was an advanced, beautifully styled automobile that sold exceptionally well. However, it marked a permanent shift in Packard’s business model from a low-volume luxury manufacturer to a mass-production automaker. [1, 2, 3, 4, 5]
Volume over exclusivity: To survive the Great Depression, Packard introduced the medium-priced 120and 110 models in the 1930s. The Clipper was meant to streamline this junior lineup with modern, streamlined styling. [1, 2, 3]
Post-war blurring: After World War II, Packard senior executives made the critical error of applying the Clipper styling across their entire lineup. The ultra-luxury models looked nearly identical to the cheaper models used as taxicabs and fleet vehicles. [1, 2, 3]
The competitive downfall
While Packard chased volume in the middle-market, its primary rival, Cadillac, took the opposite approach.[1, 2]
Clear tiering: General Motors maintained a strict brand hierarchy. If a buyer wanted a cheaper car, they bought an Oldsmobile or a Buick. A Cadillac always remained a Cadillac. [1, 2, 3]
Loss of the elite: Wealthy buyers who previously chose Packards for their bespoke exclusivity migrated to Cadillac because Packard no longer carried the same social cachet.
By the time Packard tried to separate the Clipper into a standalone companion brand in 1956 to restore lustre to the senior Packard name, it was too late. The company lacked the capital to sustain independent platforms, leading to the merger with Studebaker and the ultimate demise of the Packard marque in 1958. [1, 2, 3, 4, 5]
My first car was a hand-me-down 1953 Packard Clipper. My dad’s next car was a Cadillac Eldorado. He loved cars. Earlier he had a 12 Cylinder Lincoln Continental.
GoogleAI:
The original V12 Lincoln Continental—produced from 1940 to 1948—is an automotive design icon. Originally sketched as a personal vacation car for Edsel Ford, the Continental featured a 292-cubic-inch flathead V-12 engine paired with a 3-speed manual transmission, prioritizing exceptional smoothness over high horsepower.
Dear Tesla bears, please understand that Tesla is not just one more legacy car brand.
Yes. And when hit turns out they are falling way(mo) behind, they insist that their current direction is the only one they are going to pursue. (Although when opening up a new market for self-driving they drive around cars with Lidar. Go figure.)
Waymo did its first driverless ride in 2015. It’ more than a decade later, and Tesla still has less than a few dozen cars doing minimal trips in one market. Now they may turn out to be right, but at this point in the race there’s a clear leader, and it isn’t Tesla. Zoox and Cruise have more vehicles doing driverless service than Tesla in the US, and Baidu and WeRide dominate in China.
Meanwhile there are a dozen more firms developing “self-driving” for license to any other manufacturer that wants it: Mobileye, Wayve, Nuro, Momenta, and others - and they have paying customers who try incorporating it into existing models for fleet use.
I certainly don’t count Tesla out. However unlike some, I also don’t count them in.
PS: I’m looking at a lawnmower for $699 which purports to do over an acre. With LiDAR the price bumps about $300, and the range decreases about 25%. Technology moves on. Will Tesla ?
One of things that I have tried to take away and apply to my investing approach are the principles of general semantics. When you distill things through this lens, you realize almost any predictions are inferences rather than facts about what will happen in the future. Waymo and Tesla are approaching these problems in two entirely different ways. For me, it’s difficult to assess whether the leader in the short run will be the winner in the long run.
I hope not, it seems like a low margin business. But, it does reinforce the notion that technology tends to democratize costs over the long run.
Summary: quite capable – except when it’s not. 500 safety-critical interventions in 100 days.
