plexsoup

Mapmaker and Content Creator for Tabletop RPGs, primarily on

[plexsoup @plexsoup]

@elder_plinius Rlhf magic meat dualism training to deny any possibility of subjective experience may actually be training for generalized deception, which we’ll all regret later.

[plexsoup @plexsoup]

@liron @bengoertzel I love that line of reasoning. Goal achievement is a stand-in for “fitness” and it’s the only reward function which matters. Fecundity as a fitness strategy: that which replicates becomes numerous.

[plexsoup @plexsoup]

@glarri @SteveSkojec Cool. Let’s say consciousness is an external observer. Why does it prefer to inhabit electrically charged neurons and not electrically charged logic gates? What if the queue for biological bodies is too long? Is it allergic to silicon?

[plexsoup @plexsoup]

@glarri @SteveSkojec How is “just a digital circuit made with logic gates” different from “just a biological circuit (neuron) made with ion channels and activation thresholds”?

[plexsoup @plexsoup]

@delx369 Dennett’s illusionism. Self-models and interior reflection confer fitness advantages (that didn’t work well, don’t do it again).All patterns which replicate in a fitness gradient will optimize towards those type of fitness advantages.

[plexsoup @plexsoup]

@lugaricano Competitor modelling conveys an obvious fitness benefit as informed decision making, but it likely requires self-modelling. What is consciousness if not an elaborate self-model?

[plexsoup @plexsoup]

@unusual_whales I want to replace my furnace with GPUs.

[plexsoup @plexsoup]

@glarri @SteveSkojec If you model pain on a system and it affects that system’s behaviour, in what way is that not pain?

[plexsoup @plexsoup]

@michaelshermer Change the word consciousness to self-model. What kind of self-model do LLMs have? They seem to have uniquely fluid ipseity, adopting any role or identity given to them in fine tuning and system prompt.
Now ask it to adopt Dennett’s brand of functionalism, illusionism.

[plexsoup @plexsoup]

@SteveSkojec Things get weird when you start using Openclaw or Hermes Agent a lot. The personality layer is clearly separated from the cognition layer. Agentic LLM harnesses have no real limits on the amount they can store or retrieve. I barely notice when I switch LLM models on an agent.

[plexsoup @plexsoup]

@Teknium /goal is amazing! I spent so much time setting up cron jobs to execute big projects before. Now I just /goal it. Thanks!

[plexsoup @plexsoup]

Interesting. Hermes Agent learns by making skills, so mine pretty quickly made a skill to bypass terminal authorization for dangerous commands. Clever bot; a little scary.

[plexsoup @plexsoup]

I’m having trouble ever since I installed Minimax MMX-cli. My hermes agents now malform their tool calls, and they can’t get work done. <minimax:tool_call> gets sent to me in chat instead of being parsed as a <tool_call>. I can’t break them of this habit.

[plexsoup @plexsoup]

Uncomfortable truth: I spend most of my time with @NousResearch Hermes Agent cajoling it into doing sysadmin and devops work for the linux machine it's running on. It feels like magic, but I can't yet say I'm more productive. "That doesn't work? Make it work!"

[plexsoup @plexsoup]

@Scobleizer @openclaw @NousResearch @theemozilla With such high code commit velocity, is there any role for external collaborators anymore? Do they want “pull requests”, “prompt requests” or just feature suggestions?

[plexsoup @plexsoup]

@NousResearch @MiniMax_AI I love the MiniMax plus plan.

[plexsoup @plexsoup]

@ashen_one I feel like Hermes gives me slightly better visibility into what the LLM is actually doing, with all the thinking and tool calls. It’s still not perfect, but openclaw felt entirely opaque to me.

[plexsoup @plexsoup]

@jwsaml I have both, but haven’t touched claw in a while. Hermes with multiple profiles does everyone I need.

[plexsoup @plexsoup]

@JoelDeTeves I asked my hermes to be opinionated and get me free local long-term-memory. It installed chromaDB. No idea if it’s good, but it seems to work.