i was curious about the use of weight space programming to short cut distillation and create domain experts (in ecology and conservation) with little or no data.
just a couple parts that surprised me:
1. the tweaks around attention and information passing between modalities, essentially forcing the image model to say clearly the top answer so the llm would "know" it.
2. the simplicity of some of the hacks that Fable thought of to improve the execution of the classification tasks.
there's a push for foundation models for ecology (https://www.nature.com/articles/s42256-026-01317-6). it's interesting to me how cheaply we can explore the tradeoffs, the design, and the fitness of the general models. maybe a follow up question is how much the general models care about these questions and know their answers, but don't pay attention to them :D
i was curious about the use of weight space programming to short cut distillation and create domain experts (in ecology and conservation) with little or no data.
just a couple parts that surprised me:
1. the tweaks around attention and information passing between modalities, essentially forcing the image model to say clearly the top answer so the llm would "know" it. 2. the simplicity of some of the hacks that Fable thought of to improve the execution of the classification tasks.
there's a push for foundation models for ecology (https://www.nature.com/articles/s42256-026-01317-6). it's interesting to me how cheaply we can explore the tradeoffs, the design, and the fitness of the general models. maybe a follow up question is how much the general models care about these questions and know their answers, but don't pay attention to them :D