Alexander Embiricos
Product Lead, Codex
About Alexander Embiricos
Alexander Embiricos is the product lead for OpenAI Codex, the coding agent that has grown 20x since August 2024 and powers OpenAI's own development - including building the Sora Android app in 18 days.
Career Highlights
- OpenAI (2023-present): Product Lead for Codex
- Codex growth: 20x since August, most frequently served models in API
- Sora Android: Built in 18 days, reached #1 in app store in 28 days
- Codex on-call: Running for OpenAI's own training launches
Notable Positions
On Codex as Teammate
The vision beyond autocomplete:
"A really smart intern that refuses to read Slack. It won't check DataDog unless you ask, doesn't attend planning, doesn't know what happened in standup. That's what we're building: a teammate, not a tool."
On Compression for Long Runs
Technical innovation enabling 24-hour agents:
"Compression enables 24-hour agent runs. When models approach context limits, they prepare a compressed context, restart in a new window, and continue. This required coordinated work across model, API, and harness layers."
On Quality Threshold
Practical framing for agent utility:
"Would I have written this prompt? Maybe 50/50. That's good enough. The agent's job isn't to be perfect - it's to maintain consistency and run 24/7."
Key Quotes
- "A really smart intern that refuses to read Slack."
- "Sora Android app in 18 days."
- "50/50 prompts are good enough."
Related Reading
- Application Over Training - The product focus Embiricos embodies
- Andrej Karpathy - Codex power user cited in discussion
Video Appearances
Codex as teammate vision
A really smart intern that refuses to read Slack. When you give it the right instructions, it can do amazing things, but it won't check DataDog unless you ask. That's what we're building: a teammate, not a tool.
at 00:08:00
Sora app speed
Sora Android app was built in 18 days. A completely new application, then 10 days later it was live. This is Codex accelerating OpenAI's own internals.
at 00:15:00
Compression for long runs
Compression enables 24-hour agent runs. When models approach context limits, they prepare a compressed context, restart in a new window, and continue. This required coordinated work across model, API, and harness layers.
at 00:22:00