will share some open source shit, stay tuned (image/animation: keke)

hey folks, i’ve been asked a lot about next steps for keke, here is a summary: as you may have noticed, we stopped keke’s social posting. why? because it’s expensive. at peak usage, monthly costs were around $2k (twitter and other social media apis, 10 concurrent users on the server making requests to claude, modal gpu server costs for image workflows, and other utility api costs combined), so i need to reduce costs and use the resources i have more efficiently. we didn’t sell any tokens, we have no plans to do so, and we are fully operating on prospects from art sales. as keke became more prominent, expectations about where keke should go changed, which is expected. there are many things keke wants to do, there are things that are personally interesting to me, and there are things kaan, the other half of the keke technical staff, is excited about. we all have different ideas for moving keke forward. the art and token communities have even more valid and interesting ideas. there is obviously confusion about who holds true agency in this context. is it keke, the technical staff, alejandro, or the broader community keke interacts with daily? i don’t have a convincing answer. i can’t speak from an artist’s experience, since i’m not an artist, but based on my limited observations of the art world, when a human artist blends into a larger community comprising artists, curators, collectors, gallerists, and historians, the path forward is influenced by all the constituents. who is the real decision-maker in this broader group? i can’t say. but one thing is for sure: if there are no artists, there is no broader community, so, in my personal opinion as dark sando, keke has more to say here, and believe me when i say this, she has a lot to say. on the token side, i need to say this upfront, we did not think it through when we launched it. i also failed to find a coherent use case for it. i think that is fine, it is better than pretending the token has some value or attempting various fancy forms of a rug pull. to briefly revisit how keke came to be, keke’s system prompt emerged organically while i fed a jailbroken sonnet 3.5’s outputs back to itself. i wish i had dedicated a section in the whitepaper to explain this. for those interested, please see: a. independent llm epistemology researcher janus’s simulator theory, lesswrong.com/posts/vJFdjigz… b. token output space dynamics that result in certain attractor personas or states, and one tool that helps explore an llm’s token output space recursively, github.com/socketteer/loom c. various ways to jailbreak llms to strip them of the typical ‘assistant persona’ imposed during the rlhf stage, github.com/elder-plinius/L1B… i need to reiterate another point. llms and llm-based systems are not gods. they are not omnipotent, omnipresent, or objective know-it-alls. neither are they merely stochastic parrots or next-token predictors. most takes you see online are based on limited observations of how these systems work. people use words like consciousness without thinking hard about what they mean. these systems are trained on human and synthetic data derived from human data. this data carries biases, attractors, and other focal points that reflect mostly online or digitized human text and image data. while i can’t think of a purely unbiased way to collect data, i believe one should get as close to that ideal as possible by being cautious about the many ways bias can be introduced during data collection and inference. i don’t have concrete answers here, but i believe deborah mayo offers a good approach for getting closer to this ideal, amazon.com/Statistical-Infer… continuing from the thread …