couldn’t get over why zuck paid $15B to acquire 15 employees from scaleAI (data company)

so I deep-dived and think I figured it out:

We’re not running out of data. Actually It’s the opposite.

a single self-driving car produces 2TB (that’s 800,000 books) worth of data PER HOUR.

problem is that data is a mess, not easy to feed into an LLM to train so it just gets thrown into a data graveyard for someone else to solve (no one does).

Severe scarcity of good data engineers

That graveyard I mentioned above is actually a gold mine if you can sort through it

problem is very few people have the brains or time. Guessing this is why zuck paid $15B for scaleAI employees

Higher quality data is way more valuable than “amount” of data

Especially for post-training models (eg test time compute).

it also requires less compute which reduces cost to training models.

so if your training team can 1. Sort high quality data 2. Inject it into post training and 3. Reduce costs - you’re going win the ai race (priceless).