My new PC arrives tomorrow. Once I source myself two RTX 3060's I'll be an AI owner, no longer dependant on cloud APIs.
Currently the bottleneck is Agents. If you want a large language model to actually do anything you need an Agent. Agents so far need a human in the loop to keep them sane. Until that problem is solved most human jobs are still safe.
GPT 4o incorporated multimodality directly in the neural network, while reducing inference costs to half.
I fully expect GPT 5 (or at the latest 6) to similarly have native inclusion of agentic capabilities either this year or next year, assuming it doesn't already, but is just kept from the public.
> It will probably strongly favour places like China and Russia though, where the economy is already strongly reliant on central control.
I think you may be literally right in the opposite sense to what I think you intended.
China (and maybe Russia) may be able to use central control to have an advantage when it comes to avoiding disasterous outcomes.
But when it comes to the rate of innovation, the US may have an advantage for the usual reasons. Less government intervention (due to lobbyism) combined with having several corporations actively competing with each other to be first/best usually leads to faster innovation. However, the downside may be the it also introduces a lot more risk.
That's a very weak form. The way I use "agentic" is that it is trained to optimize the success of an agent, not just predict the next token.
The obvious way to to that is for it to plan a set of actions and evalute each possible way to reach some goal (or avoid an anti-goal). Kind of what AlphaZeros is doing for games. Q* is rumored to be a generalization of this.
Currently the bottleneck is Agents. If you want a large language model to actually do anything you need an Agent. Agents so far need a human in the loop to keep them sane. Until that problem is solved most human jobs are still safe.