Scaling Git at OpenAI
Explosive growth in AI-generated code is pushing OpenAI’s Git infrastructure to its limits, prompting a custom system (“Pando”) to handle petabyte-scale monorepos and tens of thousands of commits per day. Commenters debate whether this reflects genuine productivity gains or just “AI slop” — vast amounts of redundant, low-quality, or constantly rewritten code that humans can no longer fully understand or review. The thread raises broader questions about maintainability, security, and the future role of human developers as large labs redesign source control for agent-driven software development.
Perceived code and commit explosion
- Many are stunned by the sheer commit/LOC volume and question how it can be an asset rather than a liability.
- Others note that, relative to company size, the volume may be “mild,” and that traditional LOC metrics are becoming less meaningful than repository size and churn patterns.
AI-generated “slop” vs productivity
- Multiple commenters describe AI-agent output as voluminous “slop”: overbuilt abstractions, duplicated logic, long-winded comments, and endless “fix” commits.
- They report small tasks ballooning into huge token and commit counts, with agents repeatedly patching their own mistakes.
- Supporters argue this is acceptable if compute is cheap: code and tests can be regenerated or refactored repeatedly until benchmarks are met.
Human oversight, trust, and responsibility
- One side argues LLMs can reliably write and review most code, with humans only needed for out-of-distribution or high-stakes cases.
- Critics emphasize incentives (people get fired for bad code, models don’t), regulatory and customer-data responsibilities, and the need for someone to actually understand how data is processed.
- There’s concern about managers eventually having to answer “I don’t know, the model did it,” and about loss of human debuggability.
What is all this code doing?
- Speculation includes: RL environments, training/serving infrastructure, internal apps, massive experimentation, and capturing every agent edit for training.
- Others suspect duplication, ephemeral scripts, and “caching” generated code that could instead be regenerated on demand.
- Some worry the benefit to users is unclear given relatively modest visible product changes and sometimes clunky client apps.
Git, monorepos, and alternative VCS models
- Several note that Git already struggled at large monorepo scales (Meta, Google); OpenAI is hitting similar or worse limits with agent-driven workflows.
- Debate over whether this is fundamentally a monorepo problem vs a Git problem vs a new “agent-scale” use case (high commit rate, snapshot/log-like usage).
- Some argue big players should adopt entirely new, centralized VCS topologies rather than stretching Git; others see Git evolving into more of a protocol layer.
Broader implications
- Some see this as impressive proof that LLMs can build complex systems quickly.
- Others view it as a warning: exploding codebases, larger attack surfaces, and a development culture that optimizes for throughput over clarity and maintainability.