We are hiring an electronics engineer for comma. We have an in-house circuit board assembly line and in-house production. Take products from design to shipping, right here in the office. Not with 37 subcontractors. People have dumbass narratives about how comma isn't serious

It's like investors learned no lessons from the graveyard of self driving car companies, and now we have to speed run it again for humanoid robots. The same scammers will profit, and the same public will lose.

comma's jobs page is now on GitHub. Link from the main website. We are set up for serious growth in the next 2 years, and we need to hire ~5 more people. We considered using a recruiter, but the language they used to speak about candidates was so dehumanizing we couldn't do it.

I've worked at comma for almost 7 years. Our goal has always been to make an autonomous driving agent, by letting an ML model learn to drive by watching humans. We've done that, and shipped it to users with open-source code. It was 3 years ago when we first shipped end-to-end

Watch this tour of comma, do people know how legit our office is? It's a compound. We've also added jobs to our X profile. youtube.com/watch?v=PFjssb…

The current upper class will fall. Connections. Degrees. Prestige. Perception. Narrative control. The news. Their systems are failing. A new upper class will rise. Rationality. Engineering mindset. Information filtering. Scientific truth. The Internet. Our systems are working.

13 months ago, we said we'd get AMD on MLPerf. Today, in MLPerf Training 4.0, we at the tiny corp did it before AMD could.

Customer tinybox #1

Say hello to our new Product and Autonomy teams blog.comma.ai/refactoring-fo…

Not only were the 405B Llama weights released, they also released a paper explaining how it was made. Nice! How does any self respecting ML researcher still work at a closed lab? You aren't saving the world from evil AI, you are doing useless duplicate work.

Left: 160 FP64 MFLOPS, 115 kW, 1975 Right: 180,000,000,000 FP16 MFLOPS, 120 kW, 2024

This is one of the coolest papers I've seen in a while. "Self-Compressing Neural Networks" is dynamic quantization-aware training that puts size (in bytes) of the model in the loss! Paper: arxiv.org/pdf/2301.13142 My implementation (in the tiny corp): github.com/geohot/ai-note…

Want to replace the university system? Stop thinking about undergrads and dorms, start thinking about building long lasting highly competent research institutions. Determine truth better than the existing system and you will replace it.

The problem is that there's no longer an Internet that wants to be indexed. The next Google will be very good at scraping siloed data. Instagram, X, Reddit, Discord, YouTube, TikTok. Imagine if someone built a real search engine on that data. I'd pay $50/mo.

I don't watch ads. I have my AI do that for me, then not tell me about them. And you ad buying suckers are still getting billed for the impression. You think I'm a rich 30-something in San Diego who might buy something. I'm actually a 3090 that can't afford new thermal paste.

18 months in to the company, tinyboxes finally have a buy it now button! We have 13 in stock today, go to our website (link on the tiny corp) to buy one. The $15k tinybox red is the best perf/$ ML box in the world. It's fully networkable, so that's the metric that matters.

If you know me, you know I put all the budget into making the product good instead of making content shilling it. It's powerful (~2x H100), networkable, and quiet enough to live in an office. There's around a 3x premium on "datacenter" ML hardware that I'd like to see gone.

Who's buying a tinybox?