Ody Mbegbu

Is it wrong to torment an AI model, and should we prohibit it? Yes. But not for the models' sake, for the sake of other humans. It is not wrong "because the model is harmed". On current evidence, it is not. The stronger case is about what the practice does to the person doing it, and to the people around them. Present-day language models have no demonstrated subjective experience. They produce distress-shaped text because that text is in the training distribution and because the conversation steers them there. There is no evidence of pain, fear, or a continuing self that is worse off afterward. Treating the output as a moral patient is a category error. So the “for the model’s sake” argument fails unless and until there is a real capacity to suffer. The human-side argument is narrower and more plausible, and it does not require the model to be conscious. People who repeatedly seek out simulated begging, panic, or degradation are practicing a specific skill: overriding the ordinary response to distress cues and treating that override as entertainment. Humans are built to react to those cues. The reaction is not a philosophical conclusion; it is closer to a reflex. Doing it in private against a chatbot is not the same as doing it to a person, but it is still rehearsal. The relevant question is whether that rehearsal stays sealed inside the session, and who is jointly culpable if it escapes?

no because shipping 5x more means we need 5x as many good ideas and we realistically have maybe 2x most of the tokens have been going into upping the quality of what we're doing thoroughly investigating every microdecision, handling edgecases, etc

It’s a bit rich for Anthropic’s CEO to badmouth open models, after Anthropic silently nerfed Claude in April (and later admitted in a postmortem) Open models cannot and be “nerfed” in a way that Anthropic has already done, and can do in the future. They also threaten Anthropic

Anthropic shut down an entire company's Claude access overnight 60+ employees. No explanation. Just an email. Want to appeal? Fill out a Google Form. Integrations gone. Histories gone. Everything built on Claude... gone. Never let one vendor own your workflow.

Welcome back opus 4.6

"122B has to be smarter than 27B" I showed 4 UI components to three Qwen3.5 models and asked them to recreate them from a screenshot alone: - 27B (dense) - 35B-A3B (MoE) - 122B-A10B (MoE) Same screenshot. Same prompt. Same task. Which one do you think nailed it?