Ed Jones

Sharing my own summary of the great thoughts on Tesla $TSLA FSD Beta V12.3 by Brad Gerstner @altcap and Bill Gurley @bgurley on the BG2 Podcast: First end-to-end variant and unanimously praised for its human-like performance, and seeing reports of hours of autonomous driving without human intervention or disengagements, seems to be a game-changer versus previous FSD versions. Summary: Tesla replaced C++ code deterministic model with neural net end-to-end model driven by imitation learning. It is video in and control out, faster and more accurate. This is done via uploading videos from their best drivers, the videos are the input, and the output is the steering wheel, brake, and gas pedal. In this model, it’s pixels in, there is not identifying of objects, the driver’s behaviour is the label. Previously, it was coding every object, circumstance like a piece of software, deterministically identifying objects, and if X happens, then Y. What happens is that it ends up being a patchwork big nasty rat’s nest of code that builds upon itself. What is fascinating is that the Tesla team decided to chuck the whole old model and put in a new things is just crazy. This is eleven versions of what came before, which were just slightly better patchwork models. BG2 learnt that the rate of improvement was an order of magnitude of 5-10X better per month as a model versus prior systems. This can happen because of two things: (1) being able to upload the videos from the best drivers from over 10GB per night over the fast WIFI infrastructure, and (2) the footprint of the sheer number of cars Tesla has on the roads, 5 million. It is the quantity and the quality of the data. The car processes the data on the edge, ~99% of the data that it collect never makes it back to Tesla. If 80% of the driving is the highway and there’s nothing interesting that happens, you can just throw out all the data. What it is looking for are adverse events, the outlier moments, the abrupt movements, if the gas goes fast, hard brakes, steering wheel jerks, and all that is super relevant and it has higher model weights. ➡️ Excited to see the continued development of FSD and looking forward to @elonmusk sharing more color on it in the months and quarters to come. youtube.com/watch?v=bD1rrBBc…

@FedexSTi I think it is understanding Dlocal’s value proposition which is emerging countries payments at scale. Because of capital controls, and countless differentiation, it doesn’t make sense for every company to go into a country to sell up a payments team for pay-in or pay-outs, just to do business there for countless local payment methods. Think Netflix and Spotify for example. If they got a provider like Dlocal who can do it at scale and cheaply, it makes sense to just use them, rather than building a payments team in each country for local payments just to do start to do business in every country they enter. Revenues are thus highly recurring, and is merchant dependent which benefits as a lot of the merchants themselves are growing rapidly. Merchants are larger merchants as well, which tend to have lower churn versus SMB type merchants. Emerging payments infrastructure is fragmented and difficult to build and costs little, but once it is figured out, and built into their API, the infrastructure moat becomes stronger and stronger with network effects when they become the dominant player, and it makes little sense for someone else to come in. Add on FX, which pushes the takes rates higher. This explains why the take rates for emerging market payments is much higher than say moving USD or EUR in developed markets where most can do it and set it up quite easily. Hope this helps to provide some color. Happy to have a chat and share my thesis separately as well.