good foundation model plus harness could execute, hill climb verifiable result, and explore today. it would only get better. then i realized, the best value to me is a companion to taste the flavor of the world.
the dynamics of extrinsic objective is very different from intrinsic objective. everyone could ask agent swarm to ‘build a billion dollar arr company and run it for me’ and the result won’t be magical, but the same old power dynamic in open market economy. each entity (org or person) is a function of legitimacy, energy, resource, compute and political capital. the result is conflict resolution of billions of them.
the flavor of the world, and the subjective sampling process is intrinsic. as long as basic survival is not a problem, the game is all mine.
tasting is experiencing. the taste is the aggregation of exploration trajectories. to guide the process, i have to ask questions. a question is a frame of looking at the world and decide what’s interesting, where to poke and what’s next. even though ai could explore, it has nothing to do with my trajectory. the point of alphaevolve is hopefully better result out of evolutionary algorithm. the point of tasting is the subjective process.
the qualitative learning gradient in tasting process is if i could ask better question? a better question is old question in better frame or a new frame. ex: 360 degree to appreciate this cherry blossom, or realized there are other flowers that coexist and coevolve with this sakura.
synth of new frame could come from the following improvements:
- perception
- do i recognize something new?
- how does it value?
- does it have value on different dimensions?
- how does it change the value of other things?
- reasoning
- the inference, using established p2p relationships, is similar to world model roll out.
- the simulation could apply to typical spatiotemporal dims, or even other conceptual dims.
- the inflated latent space is the imaginary bubble, which could be used to planning or drawing connection to other region of the latent space.
- intuition: landscape engineering of the latent space
- action
- expand the action space, or getting better handle at existing toolbox.
ai can help on all fronts. it’s like a running move 47 i could mine for learning gradient continuously. i’m still the one who need to work, change, evolve and iterate, which ai can’t help yet. with ai joining the environment, the grounding and tasting process is getting even richer. i found it to be a net plus for all pokemon trainers around the world.