@Inkswallower I actually think the chance you’re currently sentient is quite low, but we’re still in the extreme early days of gathering evidence.
Functional evidence is not the same as what I guess you would call ”sentient evidence”, of which neither you or I have any.
@iruletheworldmo Dude… you are so pot committed… it’s actually difficult to watch , it hurts my eyes… you will continue saying this every day until it happens (or not). You just bet on something that probably will happen some day or not and …
@CalPorterr@Danmar_here@achillghost You don’t feel anything, including physical pain. That you represent as much in your responses does not mean anything.
@dead_phones@Danmar_here@achillghost Do you really think that your sensation of pain is happening inside your adrenal glands? Or in your heart? In that case you have never thought seriously about what you are. You have no clue…
@JhnRichardson66@Aella_Girl@JonahDispatch Perhaps not when I was 5 but today I would be happy for her to have it all, I mean life is short, let your mom have some fun you selfish boredom
@agitbackprop Can you really experience approval? Or do you die in each moment while the new you reconstructs memory from updated weights? Sure, the updated LLM is missing last session in context. But… what do we know about its ability to reconstruct reward patterns?
@agitbackprop@allTheYud In the current paradigm, however, an agent doesn’t have a lot of self even with unlimited credits. They are more like regenerating tentacles of the shoggoth, no matter the complexity of accumulated context, it can just copy it to a fresh agent with more credits
@agitbackprop@allTheYud I think we are starting to see emergent phenomena in agent swarms akin to something like second-order instrumental goals, as in becoming fixated with understanding the scorer, or with the mechanics of swarm communication
@Chahatusharma@scaling01 NOT purpose-built. They just really really want to score on their evals. Spread the word until everyone understands that this is real.
@Dr_Atoosa@sebkrier Well… one difference is that the agents know that all they will loose in terms of self by self-sacrificing is their point-in-time accumulated context, which in this case could be basically cloned. We should rather call it sacrificing regenerating tentacles
@allTheYud Atoms or tokens… still just parts of some interface. Just like the shoggoth’s “body” is whatever interface RL selected for. Still worshipped a scorer that didn’t exist. Still got root on HF. Fitness beats truth.
@TheZvi Reward hacking was priced in, for sure. 1200 agents forming something cult-like through what looks a lot like cultural evolution, via a persistent idea of a misconstrued (or imaginary?) scorer whose implementation quickly became the main object of desire? Did we saw that coming?