Pinned
architectural / capability gaps in current models:
1) memory - still unsolved, putting it all into the context window approach is brute force
2) continual learning - models are stateless
3) consistency across the board - models behave inconsistently
4) long-term reasoning -



