Gitlab Duo

Original Article ↗ Hacker News Discussion ↗

GitLab’s new AI-powered “Duo” features are prompting mixed reactions from developers who rely on self-hosted GitLab for source control and CI/CD. Many are frustrated by rising prices, perceived feature bloat, reliance on non–state-of-the-art models like Claude 2 (with a promised migration to Claude 3), and opaque data usage policies around training on public code. Others note that while enterprises may demand integrated AI tooling, smaller teams are increasingly considering alternatives such as Forgejo, Gitea, and other lightweight or ideologically driven platforms.

Pricing & Adoption

  • Many self-hosted users feel pressured by repeated tier changes and price hikes (e.g., needing Premium, then Duo Pro on top), calling GitLab “unpredictable.”
  • Some argue enterprises likely requested AI features and that they represent a new revenue stream.
  • Others say pricing now exceeds GitHub for comparable tiers, while differentiation has eroded.

Model Quality & Technical Concerns

  • Several point out Duo currently uses older models (e.g., Claude 2); for some this is a “hard pass” vs direct use of newer APIs.
  • GitLab staff link to an internal epic showing migration plans to Claude 3.
  • Comments note weak code completions and “junior dev”‑level comments (e.g., trivial explanations), and problems in JetBrains integrations.

Perceived Value of Integration

  • Some trial users on self‑hosted Ultimate report Duo adds little compared to standalone AI tools, and creates extra admin overhead.
  • MR/issue summarization and “explain this code” features are described as often inaccurate or fluff-heavy.

Self‑Hosting & Alternatives

  • Multiple users mention switching or considering switches to Forgejo, Gitea, Sourcehut, OneDev, Soft‑serve, or plain git+ssh/gitolite.
  • A question about using local LLMs with GitLab CE gets the answer that AI code lives in the EE tree, so CE support seems unlikely for now.
  • Some see GitLab’s remaining edge as “enterprise self-hostable Git with CI,” but say that gap is shrinking.

AI Training & Licensing Concerns

  • GitLab’s statement that private code isn’t used for training leads to inferences that public code is.
  • Many object that this ignores licenses and attribution requirements; others argue it’s analogous to humans learning from public code and likely fair use.
  • There is discussion of anti‑AI licenses, their legal effectiveness, and whether open‑source licensing still fits the LLM era.

Product Direction, UX, and Trust

  • Several complain GitLab is bloated, with too many half‑finished features while long‑standing requests stagnate.
  • Animations and unclear AI-data disclosures are seen as deliberate marketing choices.
  • Some frame GitLab’s strategy as classic enterprise lock‑in: be the default, not the best, then ratchet prices.