He’s building images for a specific SBC, not trying to merge any code in the Haiku code base, so I fail to see your concerns.
2 Likes
Finally, someone gets it. Anyway, I find it extremely amusing that this is filed under “Proprietary” when all of my scaffolding is MIT. The bias is amazing.
FYI, regarding code leakage - that happened in the early days, and well before modern model training techniques and QA eliminated it. SOTA models these days come with massive IP protection guardrails that sometimes even prevent them from quoting reference works like books and documentation directly, so I suggest you get acquainted with the current state of affairs and how both Anthropic and OpenAI have been steadily improving on that.
As someone who also looks at on-device inference, I have also seen small “local” models go down that same path (partly to comply with concerns, partly because they simply cannot hold much more than generic patterns to recall due to their size).
But some open weight models (like Deepseek) are… interesting in that regard. I don’t use those for actual work.
4 Likes