Tesla’s FSD Is A Half-Baked, Janky Mess, But It Should Improve (NASDAQ:TSLA)

Tesla, Inc.’s (NASDAQ:TSLA) full self-driving (“FSD”) feature has been made available to 400,000 customers and has driven roughly 100 million miles. Yet it remains an awkward bundle of strengths and weaknesses. As a toy for enthusiasts, it might be the best thing on wheels. But for the typical automotive consumer, there is a real question of whether it’s worth the trouble.

A video from YouTuber AI DRIVR showcases a wide sampling of both good and bad behaviors from FSD. This video is fairly representative of what I’ve seen from the many vloggers documenting their experience with FSD.

I have a personal theory as to why FSD struggles so much with many basic driving tasks and why, moreover, I believe FSD will improve its performance on these driving tasks in the not-too-distant future. Currently, FSD’s behavior is mostly hand-coded. For each type of behavior the car needs to perform (e.g., signaling, turning, reacting to speed bumps), an engineer has come up with a rule of thumb that can be expressed in computer code. In my opinion, this approach is doomed to fail.

Throughout academia and industry, a consistent and recurring lesson has been that hand-coded AI behavior is brittle, dim-witted, and ultimately doomed to be obsolesced by an AI system whose behavior is powered by neural networks. Currently, Tesla’s FSD AI is that brittle, dim-witted kind. But that will soon begin to change.

For years, Tesla has been very gradually transitioning small elements of FSD and Autopilot behavior from standard code to neural networks. The AI team’s philosophy toward these transitions has been pragmatic and piecemeal. Rather than try to build a pure neural network system from the ground up, the engineers have built a hand-coded system and swapped in neural network parts when they see an opportunity to improve performance.

That trend continues now. An update that will scrap many hand-coded elements for neural networks should be coming soon:

The next opportunity for any technical details from Tesla is March 1, when the company will be hosting an Investor Day. It is as yet unclear if any new details about the company’s AI tech will be revealed.

The exact timeframe for transitioning FSD’s behavior-generating software from hand coding to neural networks is unknown. However, we have good reason to think this will improve performance. ChatGPT is only the latest modern marvel that shows off the uncanny capabilities of large neural networks trained on large datasets. My personal favorite example of modern AI marvels is AlphaStar. However, to a general audience, this example is esoteric, since fully understanding it requires at least a passing familiarity with the game of StarCraft II. Suffice it to say that it’s a complex, difficult game and the AI is very good at it. In either the case of ChatGPT or AlphaStar, neural networks have surpassed many people’s wildest dreams of what AI can do.

ChatGPT and AlphaStar are also both victories of long-term planning, whether it’s planning a sequence of words that forms a cogent paragraph or executing strategies over a twenty-minute game of StarCraft. Already, Tesla has incorporated some of the same sort of AI technology used in ChatGPT in its “language of lanes” (in which natural human language and grammar serve as a metaphor for lanes on streets).

Tesla is the first company to attempt to apply this sort of cutting-edge AI research and development to a consumer robotics product. Given the uncharted terrain, progress will likely fumble along, like a person wading through a dark swamp. It seems assured, however, that Tesla is heading in the right direction and stands the best chance to have a breakout success in the field of autonomous driving. No other company has the combination of data, engineers, money, and technical approach that Tesla has.

What would a breakout success in the field of autonomous driving look like for Tesla? In the maximal case, robotaxis. In the minimal case, a premium software feature that evolves past the status of enthusiast toy and becomes a trusted co-pilot to attentive and sober drivers. Of course, the minimal case will likely precede the maximal case — if the maximal case ever comes to pass.

Projections for Tesla’s automotive revenue and gross margins need to account for growth in FSD as a product. It is practically inevitable that FSD’s rough edges will be sanded down and it will become a fan favorite feature like Autopilot. In the sketchiest outline, the technical path for Tesla to get there has already been laid out. It involves surrendering more and more control to neural networks that will be able to drive better than anything anyone can program by hand.

Why the competition fails to impress

The products that compete with Tesla’s FSD fall into two broad categories. First, there are the higher-tech, smaller-scale products like the Waymo Driver (GOOG, GOOGL). The Waymo Driver is capable of learning from experience and is updated frequently to leverage the latest in machine learning technology. However, its scale is restricted to less than 1,000 vehicles.

Second, there are the lower-tech, larger-scale product like Mercedes-Benz’s (OTCPK:MBGAF) Drive Pilot. Unlike the Waymo Driver, Mercedes-Benz’s product will be available in a consumer vehicle. Also unlike the Waymo Driver, Drive Pilot is not capable of learning from experience, nor will it be updated frequently to leverage the cutting edge in machine learning.

Drive Pilot is similar to other highway driving assistance systems like General Motor’s (GM) SuperCruise. It merits special mention only because it will have the unique feature of allowing the driver to remove their hands from the wheel and their eyes from the road under special conditions. Drive Pilot has a maximum speed of 40 miles per hour and will only operate in select geofenced sections of freeways. The intended use case is heavy, slow-moving traffic where risk is relatively low and where drivers already tend to disengage their attention, regardless of what technology their car has. Although much hype has been made about Drive Pilot, in truth it offers at best an incremental improvement over similar existing products.

Tesla’s FSD is singular in that it combines the high tech of a product like the Waymo Driver with the large scale and consumer availability of a product like Drive Pilot or SuperCruise. Other automakers could, in theory, copy Tesla’s playbook for their own ADAS systems, but, so far, none has. In practice, a lack of core competency in software and AI makes that prospect seem unlikely.

TSLA in the long term

Over the long term, I expect TSLA to continue to outperform the market. (I also expect TSLA to exhibit high levels of short-term volatility.) Tesla is leading the way in the evolution of driving into a software-defined experience. Until now, the role that software has played in the experience of driving has been quite minor. Once Tesla works out the kinks in FSD and polishes it into a reasonably mature product, as happened with highway Autopilot before it, FSD will become an attractive and absolutely unique selling point for Tesla’s vehicles. It will also push Tesla’s thick-margined software revenue much higher, thereby significantly increasing Tesla’s automotive gross margin, even as the company begins to sell more affordable vehicles.

That said, I could be wrong for one or more of the following reasons:

• Unlike Autopilot, FSD never attains enough polish to become a useful consumer product.
• Consumers turn out to be wary of or just apathetic toward FSD, even in the long term.
• Another company or alliance of companies releases a true FSD competitor, which hurts Tesla’s unique appeal and its pricing power with regards to its software.

I invite readers to consider which outcome they think is most likely.

Editor’s Note: This article discusses one or more securities that do not trade on a major U.S. exchange. Please be aware of the risks associated with these stocks.