OpenAI
The corpus represents OpenAI through three first-party engineering views: the Codex agent loop, low-latency full-duplex voice, and ChatGPT-scale PostgreSQL operations. Together they emphasize control-plane design, state management, and infrastructure choices around model-facing products.
Recurring themes
- Codex is described as a loop over instructions, tool calls, permissions, and history, with prompt caching and compaction as key context-window controls. [src]
- GPT-Live separates low-latency audio streaming from slower reasoning, tools, persistence, and delegation; stateful inference and warm handoffs serve long conversations. [src]
- The database account describes read replication, selective write sharding, pooling, workload isolation, and failover as scaling tools for a read-heavy service. [src]