I build internet stuff. Previously creative director of @stripe, original designer of @openai, and founder of @quillchat.

Joined March 2008
53 Photos and videos
2
11
1,706
3
343
NYC Compute
15
2
159
7,269
this one was better
1
169
5
251
Signal: Nominal
1
2
276
I was able to remove 90% of my deterministic guard rails, all my pre-loop classifiers, and reduce my system prompt by ~30% moving from GPT 5.4 -> 5.5. (tbf it had built up a lot of cruft 5.4 probably didn't need) 5.5 did it itself over ~8 hours total (never had anything run this long before! 3-4 sessions, mostly unattended), then it took me 2-3 hours to review and clean up. Would've taken me ~2 weeks on my own, probably.
3
9
1,288
weeeee
3
1
20
1,698
bonus of working on an image/video agent: I thoroughly enjoy skimming through the eval results
5
14
2,756
Also the irony that LLMs aren't that great at writing system prompts (tbc probably a skill issue)
1
1,693
Every time I catch myself trying to do clever system prompts deterministic guardrails (old world thinking) and instead rethink it into a few simple primitives, my little agent loops get like 2-3x more capable and 90% less brittle Also: evals. Evals evals evals.
1
7
1,655
2
13
3,121
spaghetti
11
9
296
29,698
1
2
155
16,885
the curse of building a sandbox is it's quite fun in the sandbox itself
8
5,066
11
5
115
13,758
(too complicated)
7
3,851
I just had my first instance of an LLM using a tool in an unexpected way but _it produced a valid output_ so yeah you could say I know what it's like to see your child take their first steps
1
9
4,683