LaQue Research. 15-Sep-2023. Twitter: @seth1620 $h*t Show at the Giga Factory The CEO has no Robotaxis Adam Jonas of Morgan Stanley wants you to believe that he’s discovered $500B in new value in TSLA stock—not due to any new public information from Tesla—but based on his new “analysis” of Tesla’s custom-silicon supercomputer Dojo. Actually the majority of this newfound value is not really about Dojo at all. In fact, it’s just a repackaging of a familiar bull case: that Tesla will be a leader in robotaxis and true autonomous driving. Dojo—according to Jonas—is just the tool that will accelerate Tesla to “achieve competitive asymmetry in the multi-trillion-$ autonomy market.” Jonas has not attached Dojo’s value to Tesla directly via cost savings—rather he posits that the cost savings will create additional value by allowing Tesla to reach the level of AI training compute (in the datacenter) needed for true autonomy sooner. Jonas uses 100 ExaFLOPs—based on Tesla’s own goal—as the apparent required level. Morgan Stanley estimates it would currently cost ~$7.5B to reach this with Nvidia hardware or ~$1.3B with Dojo. The purported cost savings of using Dojo are vast and prudent if real—but if this is truly “a $10trn TAM'' opportunity, and speed is the key to “competitive asymmetry,” then $7.5B would be a drop in the bucket and Tesla can afford it with over $15B in cash as of Q2 2023. Note that Morgan Stanley’s estimates are based on channel checks and so account for currently inflated prices due to supply constraints and high demand. While supply constraints exist now for Nvidia GPU’s, Tesla has been working on their FSD program for years and had an opportunity to rack sufficient GPU’s in their datacenter years ago before shortages emerged and prices skyrocketed. 1 LaQue Research. 15-Sep-2023. Twitter: @seth1620 Jonas admits that the cost and performance figures for Dojo are based entirely on what Tesla and Elon Musk have said and that he has not and can not “explicitly validate” them. There are ample and obvious reasons to be skeptical of Tesla’s claims, but let’s momentarily give them the benefit of the doubt and assume all the asserted optimizations to their FSD training workflow are legitimate. Let’s also assume that regardless of cost Tesla is truly unable to source sufficient Nvidia GPU’s to reach the required level of compute. Data is often stated as one of Tesla’s primary advantages in the race for autonomy. Tesla has a large and growing fleet of “400k+ [FSD users] on the road already collecting data from 300+ million miles traveled” according to the Morgan Stanley report—but the quality of that data is questionable. The cameras used with HW3—the vast majority of the existing fleet—are low resolution at 1.3mp. To reduce the data transfer burden, Tesla does not capture entire drives, only clips. Since their system only uses cameras there’s no confirmation data from lidar or radar to help validate the captured visual data. Further, Tesla FSD is an SAE Level 2 system which means that it can and often does terminate and dump control back to the human driver at the most difficult moments—exactly the moments it needs training data on most. Tesla insists that AI training is their bottleneck to achieving true autonomy so let’s ignore the data quality issues discussed above and assume hypothetically that Musk can snap his fingers and instantly train an AI model with all of the data Tesla has collected from the existing fleet to date. In that scenario would all Model 3, Model Y, and recently produced Model S/X Tesla cars (all HW3/4) on the road then be capable of SAE L4 or L5 autonomy with a simple over-the-air update to their FSD software? 2 LaQue Research. 15-Sep-2023. Twitter: @seth1620 Training is only half of the AI lifecycle. The other half is inference: where the AI model created from the training is deployed and run on the FSD computer in a Tesla vehicle. The inference phase of an L4/L5 autonomous vehicle is a mission critical application: accurate decisions must be made in millisecond-timescale to safely manipulate the car’s basic controls. LLMs like ChatGPT run their inference phase in the cloud on large datacenter clusters using Nvidia hardware and are still often slow to reply, or give incorrect answers. The Morgan Stanley note cites Nvidia’s specs that their latest H100 architecture “promises up to 9x faster AI training and up to 30x faster AI inference” over their older A100 which is what Jonas is comparing to Dojo. Yet Dojo is only for the training phase in the datacenter and the Morgan Stanley note makes absolutely no mention of the inference phase for Tesla FSD and its severe compute limitations at the edge. It is here that Tesla is hindered by a false promise that Musk started making in 2016: “The basic news is that all Tesla vehicles leaving the factory have all the hardware necessary for Level 5 autonomy” (source: https://www.youtube.com/watch?v=-vjGEEF_p5E) . The majority of Teslas on the road today have been sold since 2019 so Musk has recently focused on delivering on his L5 autonomy promise only with the HW3 version of FSD’s in-car hardware released that year. Tesla is attempting to build an increasingly complex model from their growing data set hoping to encompass all edge cases—and expecting that model to run on their 4-year-old inference computer at the edge (HW3) with acceptable performance and latency. Imagine Rockstar Games developing the new Grand Theft Auto on the latest gaming engine and using state-of-the-art development tools, and then releasing it on an obsolete version of PlayStation (e.g. PS3 or PS4) and expecting it to run near flawlessly. Degraded performance is not an option for safe L5 autonomous vehicle operation. Since mid 2023 most newly manufactured Tesla cars have been produced with an updated version of the FSD computer and sensor suite called HW4. Retrofits of the large HW3 fleet are not practical financially and are not offered by Tesla. Assume however—again hypothetically—that retrofits from HW3 to HW4 were possible, free, and Tesla could upgrade the entire fleet instantly. Then—and in combination with the hypothetical AI model we assumed earlier that is trained on all edge cases encountered so far in Tesla’s data set—would the promise of L5 autonomy finally be delivered? In truth there are an infinite number of permutations for the movement of objects through the world and thus an infinite number of edge cases. Tesla may have collected a large amount of data but it is infinitesimal next to the set of all possible outcomes. In fact the problem that Tesla is trying to solve is even harder than it appears because the cars themselves have no ability to learn on the spot or adapt to their local environment—all the training happens in the datacenter. Thus the FSD software needs to know how to handle every edge case everywhere all the time. Any human driver knows they are more confident and capable driving in locations and conditions they are familiar with than those they aren’t. Tesla is attempting to make FSD not only as good as a human driver, but as good as all human drivers collectively. Here we reach our logical conclusion: the problem of L4/L5 autonomous driving has always been more about inference than training. It’s evident that Tesla’s brute force attempt to train an AI model for an infinite number of edge cases doesn’t scale. Yoshua Bengio, a well-regarded expert in artificial intelligence, states in his March 2023 paper Scaling in the service of 3 LaQue Research. 15-Sep-2023. Twitter: @seth1620 reasoning & model-based ML (https://yoshuabengio.org/2023/03/21/scaling-in-the-service-of-reasoning-model-based-ml/) that “current large-scale systems still exhibit significant factual errors and unpredictable behavior when deployed... these systems do not think like humans do... further scaling is unlikely to resolve these reliability issues.” Because we currently lack the artificial circuitry and AI modeling to run human-level inference even in the datacenter, let alone at the edge, we need more data than human brains have access to—namely from lidar, radar, and HD maps—as a crutch for AI to safely make predictions about the world. Bengio’s paper is highly encouraged as follow-up reading. A quick note on V12 FSD. Like all other major—and often minor—versions of FSD before it, Musk has claimed that V12 will be a revolutionary step change, finally unlocking the functionality he’s long promised. V12 does represent a major change to the training workflow for FSD. Musk often refers to end-to-end neural networks—by which he effectively means that instead of humans explicitly providing the rules of the road to the AI, the AI figures out the rules itself implicitly after being trained on enough examples. This approach is not inherently wrong and is a powerful way to train AI systems. However, as I’ve shown, the limitations of FSD are not actually about training but rather about inference at runtime. Tesla plans to release V12 FSD on the 4-year-old HW3 in-car inference computer. The result will be unpredictable behavior—as we saw in Musk’s V12 livestream demo when his car got confused about which traffic light was controlling it and tried to drive straight through a red light at a busy intersection—and predictably unnecessary death and destruction. It's clear that Tesla simply selected the wrong approach for reaching L4/L5 autonomy when Musk decided to forgo lidar and radar. It's reminiscent of Enron's video-on-demand deal with Blockbuster for essentially what Netflix became a decade later. It wasn't a bad idea—they were just early and tried to go to market with a product they didn't have the technology to make work yet. 4