Over-Automating, Googling Doctors and Neural Nets: George Hotz Reviews Programming Memes

Fridman proposes a binary pass-or-fail review of printed black-and-white memes, mostly from the programming subreddit. The first few get quick rejections: one gets a one out of ten for too much text and for playing on bad stereotypes.

The first meme that works says never spend six minutes on something by hand when you can spend six hours failing to automate it. Hotz admits he does this, and says you learn a lot from the attempt. His comparison is spending ten years on self-driving cars rather than driving there for six minutes.

Fridman calls it the programming spirit: you put far more effort into overcoming laziness than it would take to simply do the thing. He notes the meme has 78,000 upvotes.

A meme about trusting programmers more than doctors, because both Google things, draws a joke from Hotz that doctors think med school gave them divine wisdom. He recounts a doctor who called Adderall a schedule B drug, and the pair joke about outsourcing knowledge to Stack Overflow or Wolfram Alpha instead of answering an integral off the cuff.

Asked about unit testing, Hotz says comma's regression testing has gotten a ton better in the last year. They are now fairly confident that if the tests pass the code will go into someone's car and do something sensible, whereas before a merge sometimes left a process that wouldn't start.

He values integration-style testing: the thing runs, runs in CI and doesn't output crazy values. He describes a tool called process replay that runs the processes and checks they produce exactly the same outputs over about 60 minutes of driving. It looks like a useless test, but it guards anything claimed to be a refactor, and he gives the meme eight out of ten.

On IDE annoyances, Hotz says he still uses Emacs and calls the typical assistant "Clippy on steroids." He thinks better ones are coming, and says he has trained language models on Python that do a decent job of reading code and checking for bugs. Fridman cites Gmail's auto-suggest and GPT-3 as basic applications that could be useful.

Hotz doubts that GPT-3 will replace customer service agents, saying this reflects badly on those jobs and that he would rather just get access to the API. Next the pair rail against sites that autoplay audio, saying they would rather have a site mine cryptocurrency in their browser than hijack the speakers.

They say news sites are broken, with the revenue model broken and the interface worse: they would pay the New York Times or the Wall Street Journal but face logins and gym-style subscriptions designed to be forgotten. Their prescription is to be upfront, make it cheap and accessible, and ship a great product with no tricks.

They wish for an ad blocker that also hides recommendations, trending topics included, which Hotz says have zero relevance to his life and are designed to outrage and bait. Fridman suggests a Chrome extension that hides likes and retweets, saying it makes Twitter a very different experience, because he realized he judges tweets by their numbers.

Hotz blocks global trending content and says he doesn't want to know what's trending in the United States. Fridman pushes back that trending is how you would learn about a nuclear war, and Hotz jokes that a friend would text you.

Both say YouTube's personalized recommendations are different, since customized suggestions provide real value to them. Fridman had tried turning them off and realized they were worth keeping.

A Stack Overflow meme prompts Hotz to mention that everyone at comma must read Eric Raymond's essay on how to ask smart questions. In his view, the quality of the answer you get depends a lot on how you phrase the question. They also note that some of the most interesting questions are ambiguous and don't really have one answer.

Hotz says he asks ambiguous questions in interviews because they reveal how candidates reason, what they think of and what language they use. His formula for a good candidate is intelligence and motivation. Intelligence is easy to test because dumb people can't pretend to be smart, but motivation is much harder.

Comma's answer is a paid, two-day micro internship on something real in the code base, where you can see whether there is a fire or whether the candidate is just checking a box. The best candidates, he says, have read the openpilot code and ask questions about it, and the beauty of comma is that you don't need to be hired to start contributing.

Asked about resource constraints, Hotz says he doesn't care about bytes but does care about CPU time, because CPU means watts. Responding to complaints of the comma 2 overheating, he jokes that they're lowering CPU frequencies, and the code is efficient enough to cover for it.

On the "my code doesn't work, let's change nothing" meme, both admit that rerunning often works, like restarting a computer. Fridman reads it as something he catches himself doing, and mentions Advent of Code. Hotz says people don't realize how much he "cheated" on his Twitch SLAM project, since he had spent six months studying SLAM beforehand.

A machine learning meme about neural networks not leading to general intelligence irritates Hotz a little, because learners at the object-detection stage are only seeing software 2.0 at a small scale. He compares it to software 1.0, where hello world doesn't look like it will fly rockets, yet at scale it does.

He says he is bullish on what neural networks are: matrix multiplications interspersed with non-linearities, an expressive, differentiable and quasi-convex form that optimizers work well on. His guess is that the right loss function hasn't been seen yet, perhaps something that looks like compression, and that reinforcement learning has fun, not yet understood power.