Joanna, if your reading this, welcome back .. really miss your blog !

In my ideal future, Joanna would write about things like :

  - how do we secure LLMs leading up to AGI, presumably they can bypass our best attempt at airgap firewalls
  - is consciousness different from thinking from intelligence ?
  - how do we ensure AI is distributed so all humans can benefit, and not used as a multiplier of extreme wealth inequality ?
  - whats missing in current AI LLMs ?

I'll take a crack at the last one : imo, the current generation of LLMs is missing :

  - a bias for truth / fact .. or a mode switch to make it bias truthiness
  - reasoning by chains of formal language [ formal logic ]
  - reasoning by chains of probabilistic inference [ bayesian logic ]
  - reasoning by deep simulation [ stochastic modelling ]
  - spatial reasoning [ 3D model of space, machinery, physics. 3D reconstruction. model of humans and animals with bones, muscles, mass ] 
  - mathematical modelling [ proposing formulae, checking fit ]
  - psychological model of humans, of human populations, their needs / motivations / rewards and psychoses

Essentially AGI would require a proper merging of RL style [ NN learning from a stochastic simulation of future states ] and current ChatGPT style LLMs.

An implication of the above is that future AGI will need to run on fast branching CPU _and_ massively parallel GPU with a fast data path between them - ie. balanced compute.

However, the best part will be discovering what the real Joanna writes about !