Just brought to my attention and it deserves a response because it seems really fishy. I don't have the time to 'correct' it, but I'm sure there will be people trotting it out so I'm just going to preemptively demolish it here. 🧵
The article looks at Tesla's safety report data and claims to adjust it for demographics and road use, suggesting that AP is no safer than non-AP. It's a scholarly article from a prolific author in the space, but it's badly done and appears biased. /2
Jan 14, 2023 • 9 tweets • 3 min read
On the topic of Tesla's Safety Report: I keep hearing assertions that these numbers are not representative of actual safety because AP is only used on roads with few accidents. This seems to be FUD. 🧵 /1 tesla.com/VehicleSafetyR…
In particular this notion that AP is only used on highways and that highways have 1/3 the accident rate of other roads gets mentioned a lot. There's no factual basis to that as far as I can tell. I can't find the original source of that, so let's do our own math: /2
Jan 11, 2023 • 8 tweets • 3 min read
@1LoafOfMeat Graph is their public data with averaging applied. I can't rule out them lying. I have put significant effort into trying to find a better way that they could release the data to inform users without making the confusion worse - and I couldn't come up with one. /1@1LoafOfMeat After pulling the NHTSA crash database and breaking it down many different ways I was unable to find a simple, clear, and accurate way to convey crash statistics. Every approach I came up with had the potential to be gamed by cherry picking thresholds and definitions. /2
May 23, 2022 • 15 tweets • 6 min read
@MerrillEarnest My take is that it's FUD. The core argument is that increasing features in an NN scales poorly, that HW3 is maxed out on current functionality, and that upgrading from HW3 is infeasible. All of these are wrong and the entire argument is a joke. /1@MerrillEarnest This is reminiscent of arguing Tesla is unprofitable because they lose money on every car and thus *must* approach bankruptcy as they expand. Every component of the argument is flawed. /2
Jan 18, 2022 • 12 tweets • 3 min read
People misunderstand the value of a large fleet gathering training data. It's not the raw size of the data you collect that matters, it's the size of the set of available data you have that you can selectively incorporate into your training dataset. /1
This is a critical distinction. The set of data you choose to train with has a huge impact on the results you get from the trained network. Companies that just hoover up everything have to go back through the collected data and carefully select the items to use for training. /2
Oct 26, 2021 • 6 tweets • 1 min read
Today Tesla release a whitepaper describing new numerical formats being used in Dojo. These novel numerical formats are created with the needs of neural network training in mind. A brief explanation follows: 1/6
Common numerical formats in current use have shortcomings that make their use in the training of neural networks more complex and less efficient that could be ideally achieved. This paper describes two new formats: CFP8 and CFP16. 2/6
