sc94597 said:
I think we'll have systems that outperform (nearly) all humans in white collar work by the early 2030's, but not everyone will accept AGI as here until embodied entities are also prevalent, which probably is more towards the middle or end of that decade. What's currently more interesting than VLMs/LLM's in my opinion is the work being done on trained world models (ex VL-JEPA, Nvidia's various world models, Google's Genie, etc) as well as hybrid systems that combine physical/ground-truth models with VLM's (a recent example from MIT.) In general VLM's/LLM's seem to be a very useful component in all current systems, because they can generate a searchable latent space effortlessly, that the other components traverse/search on. But they probably aren't sufficient for getting us to AGI. Hardware will catch up too. We probably will (or likely China will be the leader in this) have commercial, scalable photonic computers in the early 2030's and that will solve a lot of the data center energy issues surrounding AI as energy costs come down about a hundred fold for the same compute. |
World models like Genie is actually the aspect I'm most excited for in the near future since this year seems likely for the first publicly available one that's at least semi-decent. I expect Genie 4 to be revealed within the next few months and with the insane progress between 2 and 3 in just 8 months it should be mind blowing. I really hope they let the public use that one even if you need to pay for a subscription since I can't wait to mess around with that sort of thing.
