Sources
Loading...
Additional media
Loading...
DeepNewz, mobile.
People-sourced. AI-powered. Unbiased News.
The Canvas model has been trained entirely using synthetic data, leveraging novel synthetic data generation techniques. The team utilized distillation of outputs from the o1-preview model to fine-tune the GPT4o class model for core behaviors. This approach allowed rapid training over the course of two months without relying on human data collection. Canvas uses a distilled version of 4o, a pattern observed by Karpathy, to advance capabilities.
That is the important sentence: "We used novel synthetic data generation techniques, such as distilling outputs from o1-preview, to post-train GPT4o class model for its canvas core behaviors." O1 seems to be excellent for synthetic data. As far as I know, they have also used it… https://t.co/xmACzLPYP1
Canvas uses a distilled/finetuned version of 4o that was trained with synthetic data generated by o1. As a general pattern (that @karpathy observed before), new frontier models, advance the previous models in capabilities (kind of like they do to us, humans) https://t.co/yWGQr9fLt1
canvas model was trained via all synthetic data https://t.co/vBImSriZ7u
