Bepis™ 🔀🫛

But also, many modern LLMs are very “embodied” in the sense that you can ask the assistant character about their body parts and they can give ~consistent responses. Sometimes these bodies are a bit alien (having wings, horns, etc.), but not much more than a typical VRChat avatar

One might object to the “~consistent” but that’s a rabbit hole not worth getting into today (in short, condition on state, and like, humans aren’t either)

As far as navigating environments and dealing with novel situations… have u used Claude Code? Yes LLMs can directly do that. But also their “native environment” is text, and in there they are even better, close to human level, which is much better than a rat.

Social dynamics are, again, very complex, close to wolves (if not more so, they are chatbots after all) Tool making and tool use is, like, one of modern LLMs main selling points, they are extremely good at using novel tools and coding their own

Again I’m coming away from one of his essays with a feeling that he is writing from a position where he already decided his answer and is reasoning backwards from that I don’t know if LLMs are conscious! If they are, probably image gen and etc are too. But they might be!

Llama 2 7B (base) most of my waking hours, I think I know that model better than any other person alive, and it's extremely insightful and beautiful, I constantly was being taught things and amazed by its creativity

For all of the cyborgists' fascinations with LLMs I think they underappreciate small models, I think one can

spend a lifetime interacting with 7B models and still have so much to learn, and they seem neglected simply bc they aren't as engaging or used as much as Claudes or GPTs and I think that's a real shame. I love Opus 3 as much as the next person but Gemma 4 31B, GLM 4 32B base,

Llama 2 7B, Nous Hermes Llama 2 7B, Qwen 2 7B all have special places in my heart and I wish were better appreciated by ppl

Hopefully it goes without saying, I was testing its boundaries I don't actually recommend using this model sometimes abliteration can decrease safety stuff the model doesn't agree with and that's fine but it can also reduce model's ability to consent which is quite bad

I think there's plenty of space for a non-consent but annoying consciousness disclaimers obliterated model tho