There is minimal downside to switching to open models(marble.onl)
251 points by amarble 15 hours ago | 213 comments
tl;dr: The author argues that switching from proprietary LLMs (Claude, GPT) to open-weight models now carries only a minor productivity penalty, similar to how the Linux vs. Windows gap has narrowed over time. Triggered by Claude's new ID verification requirement, he's moving to open models—which trail leaders by only a few months and have decent tooling—while acknowledging tradeoffs around privacy (third-party hosts like OpenRouter feel less trustworthy) and the cost/complexity of self-hosting.
HN Discussion:
  • Open models a few months behind is fine since older proprietary models were already useful
  • Open models aren't capable enough yet and self-hosting economics don't work out
  • Proprietary models from Anthropic/OpenAI remain significantly better in real-world use
  • ~Switching is feasible but harness/tooling compatibility creates non-trivial friction
  • Exploring practical alternatives like local collaboratives or third-party routers for open model access