Terms like AGI, Proto-AGI, and recursive self-improvement obfuscate somewhat how progress in AI actually happens. I so much doubt one can take OmegaClaw, add in Ben's papers, other data, put in a loop, let the wheels spin enough, and there you are, version 2.0.
The more realistic version is likely based on what people observe when they work very intensely with Claude and other AI.
Claude is smart in some ways, and can fail unexpectedly in dumb ways. For the next several years the only possible improvement is by painstakingly and diligently working with AI, uncovering and correcting its blunders, and gently and carefully letting go more and more of the leash, as it gets more steady on its feet.
So it will a lengthy and highly interactive process rather than some singularity-based thingie in which compute and some schema are determining factors.
I guess this quote from the article is a good way of putting it: The human role becomes more selective, not less important.
"the only possible improvement is by painstakingly and diligently working with AI, uncovering and correcting its blunders, and gently and carefully letting go more and more of the leash, as it gets more steady on its feet." ==> That is what I thought in March, but working with hives of OMegaClaw agents driving coding agents has made me more optimistic about what progress can be made via allowing agents to have more autonomy.
If more autonomy works, it is great, of course. At least the AI should have a rich enough environment where it can work, fail, get feedback, learn.
My whole experience with AI suggests that details are always harder to get right than what we think, it always takes a long time, and repeated refinements are needed. But maybe the time has come to hope for better.
Terms like AGI, Proto-AGI, and recursive self-improvement obfuscate somewhat how progress in AI actually happens. I so much doubt one can take OmegaClaw, add in Ben's papers, other data, put in a loop, let the wheels spin enough, and there you are, version 2.0.
The more realistic version is likely based on what people observe when they work very intensely with Claude and other AI.
Claude is smart in some ways, and can fail unexpectedly in dumb ways. For the next several years the only possible improvement is by painstakingly and diligently working with AI, uncovering and correcting its blunders, and gently and carefully letting go more and more of the leash, as it gets more steady on its feet.
So it will a lengthy and highly interactive process rather than some singularity-based thingie in which compute and some schema are determining factors.
I guess this quote from the article is a good way of putting it: The human role becomes more selective, not less important.
"the only possible improvement is by painstakingly and diligently working with AI, uncovering and correcting its blunders, and gently and carefully letting go more and more of the leash, as it gets more steady on its feet." ==> That is what I thought in March, but working with hives of OMegaClaw agents driving coding agents has made me more optimistic about what progress can be made via allowing agents to have more autonomy.
If more autonomy works, it is great, of course. At least the AI should have a rich enough environment where it can work, fail, get feedback, learn.
My whole experience with AI suggests that details are always harder to get right than what we think, it always takes a long time, and repeated refinements are needed. But maybe the time has come to hope for better.