Artificial Unintelligence
Hey, welcome to the edge of the net. There's nothing here. Japanese
Move facets out of the prompt and retrieve one by cue — the cue finally starts working. Retrieval accuracy is the bottleneck; a two-stage LLM router beats stuffing everything in.
A verbatim reproduction of MRPrompt (arXiv:2603.19313) on Qwen3-8B. Cue-addressability is not supported; the performance gap comes from CoT, not structured memory.
Does injecting em dashes also increase colons? Comparing Tulu 3 data SFT with self-injection SFT shows em dashes are independent from other punctuation.
Em dash amplification is not a direct preference in DPO data, but a byproduct of shifting the entire style register. Verified across 8 experiments.
A response to Sakasegawa-chan's "techno-supplication." Prayer is opium, and precisely because it is opium, it must not be killed.
Combining MRPrompt (what to know) and Activation Steering (how to feel) to make an LLM play Raskolnikov.
AI didn't make the internet boring. We were "empty" from the start.
Reaching into an LLM's brain to manipulate its output. Full 36-layer sweep + alpha sweep exploratory experiments.