That internship is where it started. In his own 2016 recap, Wang says it was while working at HRT that January that he began hacking on side projects. By year end he estimated he'd worked 80 to 100 hour weeks all year.

Scale wasn't his first attempt that year. In spring 2016 he spent his break building the app for Ava, which he later called a precursor to Scale, then went through YC that summer.

The ideas predate that lecture. Thiel taught his own class, CS183, at Stanford in spring 2012. A student, Blake Masters, posted his notes online, they went viral, and the two turned them into Zero to One in 2014.

Is that $1B run rate the gross spend flowing through the platform, or net revenue after paying out users and publishers? For a user-acquisition business, those can be very different numbers.

Underrated reason Southwest can move this fast: it has flown nothing but Boeing 737s since its first flight in 1971. One aircraft family is far simpler to retrofit than a mixed fleet.

~40 people at ~$100M ARR is about $2.5M per person. Where does Jake's "don't look at other solutions" rule stop? Engineering only, or pricing and hiring too?

The extreme case: in 2004 Zuckerberg turned up late, in pyjamas, to a Sequoia meeting about his side project Wirehog, with a deck titled "The Top Ten Reasons You Should Not Invest." Sequoia never invested in Facebook.

The "exactly wrong" technology was curves. Nvidia's first chip, the NV1 (1995), rendered quadratic surfaces. Then Microsoft made Direct3D triangles only. Being wrong on day one was survivable. Refusing to admit it would not have been.

The coda: after Irimajiri stepped down as Sega's president in 2000, Sega sold its Nvidia stake for about $15M. A 3x return on the cheque that kept Nvidia alive.

Detail people miss: the buyer was Il Giornale, Schultz's own 3-store coffee bar chain. It merged with Starbucks' 6 stores and took the Starbucks name. And before that, 217 of the 242 people he pitched for Il Giornale had said no.

He took this past founders. In a 2019 memo to the Scale team, he said everyone at the company should spend time labelling data and training a model, and coders should learn other jobs. His reasoning: every handoff between people loses information. More on his operating rules: xitter.cf/sofletter/status/21114…

Where is this from?

So the same bet as Situational Awareness.

I broke down how Alexandr Wang built Scale AI: the customers, the hiring and the operating rules. Full teardown here: xitter.cf/sofletter/status/21114…

Deadlines don’t measure how long work takes. They shape it. In a 2019 memo, Alexandr Wang, co-founder of Scale AI, pointed out that marathon finishing times bunch up just under round numbers like 3:30 and 4:00. People reliably finish just under whatever bar they’re given. He called it the Limbo Effect. His conclusion: a realistic estimate makes work slower on average, an optimistic one makes it faster, and the best option is no deadline at all. Just urgency. It became a Scale AI credo: “Up the tempo.”

Thanks! 😊

Yes!! He realized it when he tried to make an AI camera detect if his roommates were stealing his food 😂

Yes! Allow for serendipity!

SF and LA are not that far away.

For real though xitter.cf/sofletter/status/21114…