Summary
The whole day's X signal came from the morning capture of the Following feed, which covered the US Saturday. The afternoon and evening slots were empty, as usual for a US weekend overnight. The strongest cluster is decision models: five posts on small models that return one calibrated choice instead of text (Microsoft-Decision-1, the training-free AnyJev, Vega's physics-engine head, an Unsloth recipe, a router pitch). That is real routing signal, but no launch reports calibration quality. Harness engineering and inference economics each drew a four-post cluster. Stanford's harness-ceiling paper and SemiAnalysis's 88% Anthropic inference-margin estimate are the substance in those two. The standout single post is the detector-turnover paper: AI-text screening catches over 99% of rewrites just before a model-generation boundary and 3.8% just after it. Memecoin, robotics demos and tip re-posts were noise.
Posts
- Decision models become a category (cluster of 5) (@Sumanth_077 · @satyanadella · @Nandakishorm1 · @crazyMLguy · @cyrilXBT · summary) [morning]. Microsoft-Decision-1, AnyJev (no fine-tuning) and Vega all return one option with a probability, and an Unsloth notebook trains one on 4 GB of VRAM. The router cost case is strong, but no launch reports AUROC.
- Harness engineering (cluster of 4) (@rohanpaul_ai · @dair_ai · @bibryam · @undefinedKi · 2610.11655 · 2610.11573 · summary) [morning]. Stanford finds that harness evolution fixes loops and stalls, but poor plans need weight training. MemoType routes memories by type for up to 16% better Recall@1.
- Where inference dollars go (cluster of 4) (@deedydas · @rohanpaul_ai · @rohanpaul_ai · @StockSavvyShay · summary) [morning]. OpenRouter handles about $1.5B a year, with Anthropic at 36% of it. SemiAnalysis puts Anthropic's inference margin near 88%, and the FT reports Nvidia is in talks over Reflection AI.
- OpenAI revenue restatement hits markets (@rohanpaul_ai) [morning]. The Nasdaq 100 fell 1.4% and the semiconductor index fell 3.4%, even as OpenAI reported 77% Q3 run-rate growth.
- AI-text detection breaks at model boundaries (@rohanpaul_ai · 2610.11599) [morning]. Detectors trained on a vendor's past model versions collapse from over 99% recall to 3.8% across a generation boundary. Vocabulary drift predicts where detection still transfers.
- Six parallelism strategies for inference (@_avichawla · summary) [morning]. A useful explainer. Prefill context parallelism cuts time to first token, while decode context parallelism splits KV history to add capacity.
- Agent safety (cluster of 2) (@kimmonismus · @LauraRuis RT of @tomekkorbak · summary) [morning]. Nadella says to build systems that assume the model is compromised. Tomek Korbak says he lost OpenAI safety access and was escorted out.
- Mathematicians and LLM proofs (@bengoertzel) [morning]. A former math professor is disappointed by the community's reaction to LLM-written proofs. This is the flip side of the 700+ AI-generated manuscripts story (digest).
- Memecoin tool, robotics hand demos, generic "learn LLM" links, Spanish Claude Code tips [morning]. Skip.