📌 Continual Innovation Culture

• “We built a culture of repeatable innovation, and everybody else chases what we do.”
• “Most founders are just doing the same thing. But it takes a lot of effort to build a culture that is good at continual innovation.”
• “That culture is the value. It’s not any one model, any one product.”
• “Most companies don’t actually make the effort to build a company that innovates because it’s easier to just copy. you don’t get fired for copying your competitor.”
• “Our research team has to be great, but we also have to build a great product team, a great trust team, a great community team to go build these giant clusters.”

⸻

📌 Leadership & Micro-Leadership

• “There’s a very limited number of things you get to do as a leader at scale.”
  • “You can set the vision, you can set the culture, you can choose the people, and you can fix crises.”
  • …“You can evangelize to the rest of the world a little more; you can set the roadmap and you can pick people.”
• “Although people don’t like micromanagement, they actually do like micro-leadership.”
  • “If you tie it to values and principles, you can jump in with a lot of small course-corrections.”

⸻

📌 Devices & Design with John Ive

• “AI is such a crazy affordance that you can have a completely different shape of what it means to be a computer.”
• “Basically there have been two computing revolutions: keyboard-mouse-screen and the smartphone. The next one is coming.”
• “It’s very interesting to watch Jony… over the week people just sketch out the final form factor, and no one understands how that part works.”
• “The models will be good enough; what matters now is how you re-imagine everything else: the hardware, the OS, when it should bother you or not.”
• “Design is obsessive over every small detail, but also deeply about understanding the user and how the device fits their life.”

⸻

📌 Future of Models & Hardware

• “I believe in the not-super-distant future we will have a reasonable model that runs on a single GPU.”
  • “I actually suspect that model will take a surprisingly small amount of compute to train, but the inference workload will be dense.”