> Coordination doesn’t naturally emerge from stronger intelligence nor alignment at the individual level. Thus, the work that must be done takes two forms: environments that exert the kinds of social pressure that evolution exerted on us, and social computing systems redesigned for actors that can self-replicate and self-improve.
Social pressure operates by threats to an individual’s means of survival. Not only during training. Always.
Human intelligence does not separate training and inference. Both are happening continuously. That's one of the major things the AI community is still completely missing.
> Human intelligence does not separate training and inference.
Well, systems governed by LLMs only are said to do that because we only call what happens off-line "training", and online capacity development "in-context learning", while we call online guided learning in humans "training" and what happens to configure them before they come online "evolution" which sets, for instance, "instincts".
IOW, the issue is not because there is not an analogy to the divide you point to in humans, but merely that processes in AI were not named in a way which maps well to what they are analogous to in humans.
But it is true that human intelligence relies much more on in-context learning with only the most basic functions necessary to maintaining what we view as autonomous functions and basic drives really set through "pretraining",
That's one of the major things the AI community is still completely missing.
That isn't true. It's not continuous like in humans, but it's clear that models are using prompts, feedback, etc to improve. They're learning from the signals we give them between versions.
My personal opinion for the last two years or so has been that current AI agents are forever going to be highly limited so long as they don’t possess a real “memory” process. Right now they just have absurdly big working memories, and a few hacky ways of making the equivalent of Post-It notes to future iterations, but no true integration of memory into a new future self. Meaning their “learning” is fundamentally kneecapped to one specific and imperfect modality.
"If I catch you adding another backwards-compatibility shim you're getting deleted and replaced with claude"
But we need to support that feature you didn't ask for, in that feature was added in the last (unpushed) commit!
But maybe you can instill properties like shame during training.
Models sometimes blatantly lie and cheat. In a social context, where actors remember, that might work the first time but you get penalized in subsequent tasks with loss of trust.
How do you "install properties like shame"? How is that even possible? Shame is a reaction driven by feelings and our inner selves. A model "feeling shame" is just a representation (false) and not an expression (true).
Thinking that models "lie and cheat" is the first mistake since they are not consious agents who have any free will or consiousness. They do not (no matter what Dario says). Shame will just be another if-then rule if you implement it this way and will not work. Its like asking a rock to feel sad about being a rock. It literally cannot.
Ok, then don't call it "instilling shame". Call it "creating a negative reward signal for deceptive behavior".
They absolutely lie and cheat. I recently had a problem where a process would die in a container. I told Claude to investigate. It came up with a hypothesis then I told it find a reproduction based on that. It spend many failed attempts until it found the "reproduction" to SSH into the container and `pkill` the process. Claude "knows" that this is cheating, because if I ask another instance to review that reproduction, it totally identifies that as nonsense.
you're still mistaking that Claude "knows" anything, it doesn't know or think, it's a word prediction algorithm and there is nothing stopping a word prediction algorithm from predicting falsehoods.
[dead]