said by HTM, AGI knows what it did to the world (self-awareness), also signals from sensors.
google research
gwern wrote a fiction. he thinks agi starts from automl-zero which is similar to lazero and metalazero by name and perspective.
by design lazero can be deeply aligned, inspecting and studying user’s actions. it also has its own exploration space. however, these expectations can never be fully satisfied at the same time. if you want more power, you have to let go.
lucidrains repositories
this one got lots of state-of-the-art implementations for close-sourced papers and also repos for AGI. stunning.
AGI related
JEPA-pytorch (WIP) yann lecun’s version how agi will be built
PaLM scaling language model with pathways
side projects
make a video text to video generation
nuwa text to video generation
opencog
moses (supervised) for evolutionary program synthesis
repos on github
aixijs general reinforcement learning in browser repo
brain simulator 2 on windows platform
materials and links
DQfD: Learning from Demonstrations for Real World Reinforcement Learning (paper)
jiaxiaogang’s god-knows-what theory and training logs
awesome deep reinforcement learning (deep-rl)
awesome agicocosci exhausitive list of papers and repos for cognitive science and AGI