Values
Figure it out yourself
For us to win, we need to go deep. That means you need to go deep! We're friendly, but encourage that you always default to doing things yourself. You own your work, and that includes working out how to do it.
Nobody has it all figured out on day one, and we're all learning on the fly. That learning compounds quickly, but only if you do it yourself. Someone handing you an answer gets you the answer, not the understanding, so we expect extremely high ownership.
This doesn't mean working alone. Asking for context is always welcome: why something is built the way it is, what's been tried, who knows the area. We just won't hand-hold, because it robs you of the part that makes you good.
The people who do best here are comfortable with ambiguity and very resourceful.
Do the hard thing
As work is increasingly commoditized by AI, the only work worth doing is the work that's hard.
Anything easy, someone else will do tomorrow. What lasts in inference is depth: understanding the problem further down than anyone else, and doing the work that follows from it. The winner in our market is the one who is willing to go the furthest.
Defaulting to the easy option looks like:
- Solving an adjacent, more straightforward problem because the real one is intimidating
- Choosing the benchmark or workload we already know we'll win on
- Calling something done because it works 95% of the time
- Adding complexity to look clever, which is easy and only looks hard
Primitive first
Our success is measured in tokens per joule, so we'll do whatever it takes to improve it.
That means reasoning from first principles rather than from what everyone else does. Most conventions in inference are only a few years old, and starting from the physics is how we find the gaps.
It's expensive. Working a problem out from the bottom takes longer than copying a known answer, and sometimes we'll arrive back at the conventional choice. When we do, we'll know why it's right, and exactly what would have to change for it to stop being right.