Brett Winton

that datacenters are the likely most attractive honeypots for botnet infestations, further advantages orbital compute, which will be much easier to harden against infiltration (due to its modularity and relative infrastructural simplicity) (a botnet won't be able to compromise a satellite by tunneling through its air conditioning system.)

at its cybercab event Tesla disclosed it had completed 1 million driverless robotaxi miles 5 days later its run-rate seems to be 15x that. Active users fast approaching Waymo's (with the miles per active user discrepancy likely due to current restricted car supply.) https://t.co/vMd4WkmpFb

amidst talk of AI companies shifting token pricing to outcome pricing the @grok bot marketplace is very clever For now, hire a bot. In the future, an Agent as a Service platform. https://t.co/yYn9s11gTw https://t.co/vS9opyIL0j

on the importance of real world data I suspect that simulation data for tesla costs on the order of $50 per hour. And the reason they can do it so *inexpensively* is that roadways are relatively constrained environments with an extraordinarily limited menu of object types and interactions. If it weren't for the difficulty (or danger) in triggering certain road situations, conditions and corner cases, it may literally be cheaper to pay somebody to drive for data collection. Tesla is blessed with an enormous lake of real world data, including monumentally challenging corner cases, that it can use to seed its simulation training. Even still, they are on the cusp (but not quite ready) to massively scale robotaxi. This should help illustrate the enormity of the software challenge for general purpose humanoid robots. Contra robotaxis, with humanoids you can't collect high fidelity data with a sensor set that you know to be sufficient to solve the problem. With robotaxis, you know that a steering input and visual input is enough. With humans, you know that prioproceptive sense, sense of touch and vision are enough; with the former two it's not clear the fidelity or instrumentation that's required. Contra robotaxis, there is no fleet of deployable commercializable assets that can encounter the corner cases on the customer's dime, so even given relatively inexpensive simulation capability, seeding the simulation with the many more orders of magnitude of difficulties in humanoid space would prove quite daunting. Also contra robotaxis, the simulation space for a generalized humanoid is effectively unbounded. The bot itself has a wild number of degrees of freedom, the objects it encounters will as well, and the element interactions have to be much more complexly modeled (even leaving aside introducing other agents.) Net, its no surprise that in trying to crack the generalizable humanoid problem companies are seeking data; they need it! And, it is somewhat daunting, that facing a much more complex and challenging problem-set, humanoid developers are all starting at a severe data disadvantage relative to those attacking robotaxi.

Perhaps explanatory as to why the utility of diffusion models has lagged. Training on the same compute infrastructure you need 10 times as many image tokens as you do language tokens to max out performance. From this, fair to infer also that robotics also needs a higher mix of data to parameter count.

a vastly underrated characteristic seemingly common across all of @elonmusk enterprises: product capabilities continually improve even when they lie askance to the primary vector of improvement for the company; yes, this improvement helps the user experience, will marginally help robotaxi economics, creates a better product, but in no universe does it immediately or materially move the needle on model Y sales. it does mean that Tesla owners have a simply better product. Similarly, starlink is basically the only game in town, and yet its packaging is carefully considered, set up is near instantaneous and user-friendly; these are all things they didn't "need" to do. There is an attention to rate of improvement at the edges of the entire product portfolio that is a manifestation of corporate structure and culture, and given the breadth of products being worked, seems wildly ahistorical. Perhaps it is a function of the ruthless part and process elimination he subjects the organizations to. Very very very good design is the natural end-point of part and process elimination. That Tesla kept taking weight out of the Model X, even after its last and final refresh, is another example, and something no other auto company in history would consider. For the users of his company's products, the impact is subtle and remarkable and compounds. It also, over time, should compound into accelerating economic performance (of which I suspect we will see plenty of evidence over the next couple of years.) Though I think it's fair to say that much of Wall Street does not care for these sorts of niceties, the diffusion curve is long, but it bends towards compounding superiority.

we're actually experiencing the spaghetti-fication of a slice of the business and investing landscape in real time yet so much of the world--the vast vast majority--still sits outside of the tidal disruption radius https://t.co/4QWL5YdUmd https://t.co/j3esettcyH https://t.co/Q8E8CaT5vZ

@AstroGrower disagree. I think the market clears at $1 per mile (on a price basis), by coincidence I think that's a price that is high enough to keep high-cost providers alive/limping along while tesla should be able to clip quite high gross profits.

It's very fair to say that we're also wildly bullish on spaceXai *And* I think there is a profound public market misunderstanding (and therefore) overdiscounting of Tesla's prospective robotaxi cashflow. In some ways Tesla is cursed by having apparent (but not actually relevant) public market comps. Whereas SpaceXai clearly has no real public market (or private market) proxy. As a steward of Tesla shareholder capital I would prefer to defer a share swap until Tesla's implied robotaxi value more closely approximates Waymo's on a EV to forward rev basis. Waymo's most recent round prices it at ~160x 12 month forward annualized revenue (assuming an acceleration and doubling in revenue growth) If Tesla threw a year's worth of giga-Austin's Model Y production into the robotaxi channel then that could crudely yield ~$20b in annualized revenue. Pick your percentage of Model Y production (and whatever expectation you have for production coming off the cybercab line) that drives itself into robotaxi service, but it's fairly clear that it's not priced today.

agree when people purchase a vehicle they purchase for the tail event. when they rent a vehicle (as you are effectively doing with ride-hail), you rent for use. 93% of miles are serviceable with a 2 seater. As the market develops, more form factors will fill in market niches. One of the reasons why I think Tesla ultimately does license FSD is because it won't have the bandwidth to fulfill all of the different form factors that consumers would be willing to pay a premium for.