2 ms·
> Such a model could not be updated nor learn. It might have some limited ability for updates if the hardware had the model code but the weights were in memory
by dragonwriter 3y ago
> Such a model could not be updated nor learn.
It might have some limited ability for updates if the hardware had the model code but the weights were in memory that was updatable.
It might do in-context learning even without that.
- intrasight 3y ago> weights in memory That's basically how our neurons work. New neuron growth and connection isn't much of a factor in learning. Rather it's the synaptic restructuring (equivalent to AI model weights) that change relatively quickly. So we need to figure how how to "grow" mechanical brains. I envision this being done with a new generation of FPGAs tailored to this task.