Tesla's Robotaxi Takes a Wrong Turn

Vision Quest

If a car drives itself down the street with no one behind the wheel, but it has a "safety monitor" sitting in the passenger seat and is only allowed to drive in good weather within a small, heavily-mapped area of Austin, is it really a robotaxi? And if so, does it matter if it occasionally gets confused and tries to drive on the wrong side of the road?

These are the questions Tesla asked the world this week as it launched its long-awaited, much-promised robotaxi service. For years, Elon Musk has been telling anyone who would listen that a fleet of a million fully autonomous Teslas was just around the corner. In 2019, he said they’d be on the road in 2020. They were not.

But on Sunday, a small fleet of about 10 specially marked Tesla Model Ys began picking up paying—or at least, invited—passengers in Austin, Texas. The rides cost a flat fee of "$4.20."¹ In one sense, this is a huge milestone for Tesla. But in another sense, the immediate aftermath was a bit messy. Here’s CNBC:

In the videos shared widely online, one Tesla robotaxi was spotted traveling the wrong way down a road, and another was shown braking hard in the middle of traffic, responding to “stationary police vehicles outside its driving path,” among several other examples.

A spokesperson for NHTSA said in an email that the agency “is aware of the referenced incidents and is in contact with the manufacturer to gather additional information.”

The whole thing is a wonderful illustration of the central fight in the autonomous vehicle world, which is a fight about philosophy. In one corner, you have the "belt-and-suspenders" approach favored by companies like Alphabet’s Waymo. Their cars are loaded up with expensive sensors—cameras, radar, and crucially, lidar—to create a redundant, 3D model of the world. They argue that to be safer than a human, a car needs to see better than a human, in more ways than a human. This is the cautious, methodical, "we are very serious about safety" approach. It is also very expensive.

In the other corner, you have Tesla, which is making a much stranger and more ambitious bet. Tesla’s cars have only cameras. The argument, more or less, is that the world is designed for beings who navigate with vision, so a car should be able to do the same. If you can solve vision with a sufficiently powerful artificial intelligence, you don’t need the expensive lidar crutch. And because every Tesla on the road is a rolling data-collection machine, you have a massive advantage in training that AI. This approach is cheaper, more scalable, and if it works, it could change everything. The question is, does it work?

Well, the Austin launch was supervised. The cars have human safety monitors. They are limited to a geofenced area. They are not supposed to drive in bad weather or through complex intersections. And they still do things that make regulators want to "gather additional information." This is not exactly the "unsupervised" fleet Musk once envisioned.

But if you are a Tesla bull, this is all part of the plan. You start small, you collect more data, you iterate the software, you get better. The fact that the cars are on the road at all, generating real-world data in a commercial (ish) setting, is the win. The path to a trillion-dollar robotaxi business, they would argue, is not a straight line; it involves some weird swerves in Austin. Waymo may be leading a quiet, successful, and comparatively small-scale service in a few cities, but Tesla is trying to solve the general problem all at once, which is a much harder, but ultimately more valuable, proposition.

And so you have a market that is trying to price two very different things. On the one hand, Waymo has driven tens of millions of commercial miles and has a peer-reviewed study suggesting its cars are much safer than human drivers. On the other hand, Tesla has a visionary CEO, a much cheaper hardware platform, and a stock that seems to trade on the assumption that full autonomy is not a question of if but when. And if you believe that, then a few chaotic videos from Austin are not a bug, but a feature. They’re just part of the data-collection process. This is not financial advice! But it does seem like a weird way to beta test a product on public streets.

[1] This is Musk's weed joke, obviously. We don't have to get into it.

A Tale of Two AI Stacks

For a few months, a Chinese AI startup called DeepSeek was the toast of Silicon Valley. It released models that were, by many benchmarks, as good or better than what the big American labs were producing, and it claimed to do so at a tiny fraction of the cost. Major US cloud providers raced to offer its technology. It was a fascinating story about scrappy innovation and the commoditization of AI.

Well, the other shoe has dropped, and it’s a big, heavy, military-grade boot. Here’s Reuters:

AI firm DeepSeek is aiding China's military and intelligence operations, a senior U.S. official told Reuters, adding that the Chinese tech startup sought to use Southeast Asian shell companies to access high-end semiconductors that cannot be shipped to China under U.S. rules. …

"We understand that DeepSeek has willingly provided and will likely continue to provide support to China's military and intelligence operations," a senior State Department official told Reuters in an interview. …

The official also said the company was employing workarounds to U.S. export controls to gain access to advanced U.S.-made chips.

This is...awkward? And it’s not just a story about one company allegedly being naughty. It’s the clearest sign yet that the global AI ecosystem is fracturing into two distinct, non-compatible, and mutually suspicious stacks. The idea of a single, global AI race is over. Now, it’s a tale of two entirely separate technology stacks.

The American AI Stack is the one we talk about most. At the bottom, you have the hardware: Nvidia’s GPUs, which are the engine of this whole revolution. Those chips are physically made by TSMC in Taiwan, using one-of-a-kind machines from ASML in the Netherlands—a set of chokepoints the US government is very keen on controlling. On top of that sits Nvidia’s software ecosystem, the programming language everyone has used for two decades, and it’s rather hard to switch. At the top, you have the big models from OpenAI, Google and Anthropic, distributed through the massive cloud platforms of Microsoft, Amazon, and Google.

The Chinese AI Stack, meanwhile, is being built out of necessity. With the US restricting access to its best gear, China has been forced to build its own. Its national hardware champion is Huawei, whose Ascend processors are, in the words of Nvidia’s own CEO, "formidable." China’s top chipmaker, SMIC, reportedly managed to produce 5-nanometer chips without ASML’s best machines, which is impressive. And to counter Nvidia's CUDA, Huawei has its own software called CANN. It’s a full stack, even if it’s a generation behind America’s.

And the US government seems determined to make this split permanent. The export controls become a strategic tool to slow China’s progress. The allegations against DeepSeek—that it’s accessing restricted H100 chips via shell companies and working with the PLA—are precisely what the controls were meant to stop. Washington is now policing this new border, and it’s a border that runs right through the heart of the tech industry.

The fallout from this is, of course, a mess. Global companies like Apple are caught in the middle. Allies like the Netherlands are forced to choose sides. But the most interesting consequence might be that this tech divergence creates different evolutionary paths for AI. The US, with access to the most powerful hardware and compute, is focused on building ever-larger, more capable models. China, forced into a state of relative GPU "poverty", is getting extremely clever about efficiency. We’ve already seen how DeepSeek’s models can achieve amazing results with less compute. It’s a fascinating, if unnerving, new reality: the race for AI supremacy is no longer a single event, but two different races, run on two different tracks, with two very different sets of rules.

The Scoreboard

• AI: Abridge, Whose AI App Takes Notes for Doctors, Valued at $5.3 Billion at Funding (WSJ)
• Semiconductor: How US Chip Policy Squeezes Key Allies (ARPU)
• Semiconductor: Chipmaker Melexis Bets on Malaysia’s ‘Neutrality’ to Power China Growth (Nikkei Asia)
• E-Commerce: Amazon to Invest $54 Billion in Britain Over Next Three Years (Reuters)