The data decides what the model learns to value.
Everyone is obsessed with architecture and scale. At frontier labs, the fight is for better data. What "better" actually means, from inside the room where it's decided.
Read on Substack →/ archive
Essay
Feb 2026
Go around.
The most underrated skill in aviation is knowing when to abandon an approach. The same thing is true for careers.
decisionheight.substack.com/p/go-around →
8 min read
Notes
Feb 2026
RLHF is an interview, not an exam.
Why "helpful and harmless" reward models flatten a model's voice, and what the post-training team did about it.
decisionheight.substack.com/p/rlhf-interview-not-exam →
6 min read
Essay
Jan 2026
The generalist tax, finally paying off.
Pilot, CPA, PM, ML researcher. For a decade this looked like indecision. In frontier AI it looks like a moat.
decisionheight.substack.com/p/generalist-tax →
11 min read
Letters
Jan 2026
To the engineer who wants to be a PM.
You don't need a second degree. You need one loop of a real product, shipped under pressure. Here's how.
decisionheight.substack.com/p/engineer-to-pm →
4 min read
Essay
Dec 2025
A checklist is not a cage.
On process, judgment, and why aviation's boring safety culture is the best thing that ever happened to a high-performing team.
decisionheight.substack.com/p/checklist-not-a-cage →
9 min read
Notes
Nov 2025
What I actually do at Meta Superintelligence Lab.
A plain-english description of "RL datamix." No jargon, no NDAs broken, just the work.
decisionheight.substack.com/p/what-i-actually-do →
5 min read
/ looking for free notes?
Substack is for the long stuff. The shorter takes live elsewhere.
I post free observations, mini-essays, and field notes on LinkedIn. Premium long-form lives only on Substack — by design.