openai

OpenAI — AI lab; vendor of the GPT model family and the Codex coding agent. Initially cited at one remove via Tw93’s three essays; 2026-04-27-harness-engineering-codex-agent-first is the wiki’s first directly-OpenAI-authored source (Ryan Lopopolo, Apr 2026 — the Codex “no manual code” case study). 2026-04-27-the-second-half-of-ai is the second: Shunyu Yao (OpenAI; ReAct, CALM) on AI’s “first half vs. second half” — the wiki’s first thesis-level source from inside OpenAI, framing the field’s bottleneck as evaluation rather than methods.

What this wiki currently knows

Models / products referenced

• GPT-3 — the 175B baseline beaten by 1.3B InstructGPT on human-preference evaluation; cited as the article’s opening proof that post-training can outweigh pretraining scale.
• InstructGPT — Ouyang et al. 2022; the historical reference for “post-training reorders capability ordering”.
• o1 — the public reference for reasoning-models; opened the second scaling axis (inference compute).
• Codex — coding agent / engineering team. The “no manual code” experiment in 2026-04-27-harness-engineering-codex-agent-first is the primary source for the throughput claim that prior wiki sources cited at one remove: 3 engineers (now 7), ~5 months, ~1,500 PRs, ~1M LoC, ~10× hand-coded baseline. See codex for the full case study.
• ChatGPT Plugins / function calling — Weng‘s 2023 agent survey (2026-06-04-llm-powered-autonomous-agents) cites both as the productized form of tool-use: tools supplied by third parties (Plugins) or self-defined (function calls). The 2023 baseline the wiki’s later tool-design discipline (agent-computer-interface) builds on.

Techniques / programs attributed

• RFT (Reinforcement Fine-Tuning) — the productized post-training interface where the user supplies task definition, grader, and reward signal. Named alongside RLHF and DPO as the three paths today’s post-training systems route between (2026-04-27-llm-training-principles-paths-practices).
• Deliberative Alignment — OpenAI’s recursive-safety-reasoning approach: the model evaluates safety policy during inference rather than via reflexes baked into weights. Contrasted with Anthropic’s Constitutional AI / RLAIF.

Public posts in this wiki

• Harness engineering: leveraging Codex in an agent-first world (2026-04-27-harness-engineering-codex-agent-first, Ryan Lopopolo, Apr 2026) — the wiki’s first OpenAI-authored source. Primary citation for the Codex throughput numbers, the AGENTS.md-as-table-of-contents pattern, the ralph-wiggum-loop, the custom-linter regime, and the entropy-and-garbage-collection cleanup-loop pattern.
• The Second Half (2026-04-27-the-second-half-of-ai, Shunyu Yao, Apr 2025) — thesis essay based on talks at Stanford 224N and Columbia. The “recipe” (language pretraining + scale + reasoning-as-action) now generalizes broadly enough that benchmark-hillclimbing has been industrialized; the field’s bottleneck shifts to evaluation design — the utility problem. Drives the page’s coverage of Yao’s RL-three-components frame and OpenAI’s pre-2022 environment-first plan (gym → World of Bits → Universe → Dota → robotic Rubik’s Cube) → priors-first realization (WebGPT / ChatGPT) → recipe-era (o-series / deep research / computer-using agent).
• Building a safe, effective sandbox to enable Codex on Windows (2026-07-12-codex-windows-sandbox-engineering, David Wiesen, May 2026) — the wiki’s third OpenAI-authored source, and the first about the shipped external codex product. Design history of the Windows sandbox: rejected native primitives → unelevated prototype (enforced writes, advisory network) → shipped elevated design (dedicated sandbox users + firewall rules, four binaries). Primary source for agent-sandboxing.
• Designing agents to resist prompt injection (cited via prompt-injection; not yet ingested).
• The deliberative-alignment explainer (cited via Tw93; not yet ingested).
• Introducing Aardvark (cited via 2026-04-27-harness-engineering-codex-agent-first as another agent participating in the same codebase; not yet ingested).

Third-party investigations of OpenAI APIs

• Investigating how Codex context compaction works (2026-04-27-codex-context-compaction-investigation, Kangwook Lee, Mar 2026) — reverse-engineers OpenAI’s encrypted compact() API used by Codex models via a 35-line prompt-injection probe. Recovers the compactor LLM’s system prompt, the compaction prompt, and the handoff prompt — all near-identical to the open-source non-codex versions in the Codex repo. Confirms server-side Fernet encryption (AES-128-CBC + HMAC-SHA256) on the blob round-trip. The wiki treats this as third-party evidence of current behavior, not as published spec — see codex for the absorbed details and context-engineering for the verbatim prompts.

Stance

The wiki’s view of OpenAI shifted with the 2026-04-27-harness-engineering-codex-agent-first ingest: previously OpenAI was a source of programmatic patterns cited at one remove via Tw93 (Codex’s harness-engineering checklist, Deliberative Alignment, the o1 reasoning-model paradigm). The Codex post resolves the previously-flagged “wants primary verification” caveat on the 1500-PR throughput claim (harness) — it is the primary source. 2026-04-27-the-second-half-of-ai adds a second OpenAI-authored register: Yao’s essay isn’t a technical post but a thesis on where the field is — useful as the wiki’s load-bearing reference for “the recipe works; evaluation is now the bottleneck”, and as the primary source that retroactively explains why the LLM-agent recipe generalizes (priors + reasoning-as-action — see react). Other claims sourced through Tw93 (e.g. the OpenAI prompt-injection guidance) still want their own primary ingests when they land in raw/.