Daft

Some people think we're running out of data. Sammy from Eventual says it was never about volume though.... Use something like Claude today and you get output in the top caliber of software engineers on certain tasks. Pretty good. But the labs are pushing further. And further. Here's what nobody sees. As you climb the stack of quality, sourcing that data becomes exponentially harder. The last mile isn't a little harder. It's a different problem entirely. Which is exactly the problem Eventual exists to solve.

Some people think passion means 996. Sammy of Eventual defines it completely differently. You only have 24 hours in a day. Passion doesn't mean grinding around the clock. Passion means it's always running in the background thread of your brain. You're always thinking about it. You're immersed in the problem. And here's the distinction. If you're solving problems that come up every day, where it's about grinding them out? You don't necessarily need passion for that. But when you're solving problems that have never been solved, when you're on the frontier, those problems have to live in your mind constantly. And that only happens if you're genuinely passionate about it.

Some engineers are paid to write code. At Eventual our engineers are paid to find the problem. What separates them? Number one: intellectual curiosity. Good engineers are systems thinkers. They love going deep into a problem. Rabbit holing. Never afraid to go lower and lower level until they've uncovered the real problem. Number two, and this is becoming everything: Agency. Because the coding part of the job matters less and less now. What matters is uncovering the right problem to solve. That's the core of engineering. First find the right problem. Then solve it in a deep, critical way.

A teleoperated robot episode doesn't start when the recording does - the operator is still setting up while the camera rolls. The robot's own joint positions tell you which frames matter without decoding a single pixel. With this method, trimming all of DROID - 500 hours of robot data - took 32 seconds on a laptop. https://lnkd.in/gthAb6zN

Zuck is spending tens of billions on GPUs. Here's what actually happens. Someone understands the algorithms. Or someone understands the data. They train a model. And they're using 10 percent of the hardware. The rest of that utilization? Money lit on fire. So the engineers getting the big bucks understand how to build the algorithms, how to run the training, how to sort the data. And how to make all of it efficient. That gap is the reason Eventual exists. Everyone can build the model. Almost nobody can make the machine earn its cost.

Some engineers chase what's hot. Sammy, the founder of Eventual, says the rarest skill is perspective. Engineers get caught up in what's hot right now. What's hot for the next three months. They're not taking the long view. He sees it constantly. People look at a startup's revenue growth curve and jump on the bandwagon. RAG got hot, and suddenly every engineer he knew needed a RAG startup. That collectivist mentality. Most people follow the pack. The real work is digging in and asking: What am I actually passionate about? What do I want to be part of? Where is this industry going? The three-month view is crowded. The ten-year view is empty.

Some engineers spent a decade building web apps. Physical AI, the problem Eventual is built around, makes that look small. Software engineering became synonymous with building a web app. A full stack application. In physical AI, you're not building a web app. The product is physical intelligence. Systems that move and change things in the real world. Web engineering runs on server-grade hardware. Cloud instances. Comfortable. Physical AI runs somewhere else. NVIDIA hardware. Embedded systems. A completely different set of constraints. And that is very fun engineering. Any software engineer should look at it and think: Hell yes. That's the environment I want to work in. Because those are challenges you'd never get to touch anywhere else.

Some engineers optimize inside the box. Sammy built Eventual to make the box bigger. In software, constraints make sense. But in physical AI, with GPUs, that frame breaks. Because if you make something more efficient, you don't just save money. You buy room for higher quality. Train more efficiently, train a bigger model. Improve inference, put a bigger model on the robot. So why optimize within the constraints you're given? Why not make the constraints better? That's where young engineers fall short. They look at the box they're handed and interpret it as a box. The move is to open it. To ask, given every dimension I have, what can I actually exploit here?