Summary
No curated reposts landed in the morning slot, so the signal comes from the morning capture of the Following feed, which covers the US Saturday. The strongest thread is a cluster of five posts on decision models (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 to train your own, and a router pitch. A second cluster of four posts is about harness engineering: the Stanford ceiling paper, DAIR's memory-routing pick, the "12 harness patterns from Claude Code" article and a 17-author agent survey. A third cluster of four is about inference money: the OpenRouter dollar conversion, SemiAnalysis' Anthropic margin estimate, the FT report on Nvidia and Reflection AI, and the market reaction to OpenAI's revenue restatement. Standouts outside the clusters: a detector-turnover paper showing AI-text screening collapses at model-generation boundaries, Tomek Korbak's account of losing OpenAI safety access, and Nadella's "assume the model is compromised" post.
Posts
Decision models become a category (cluster of 5) (@Sumanth_077, @satyanadella, @Nandakishorm1, @crazyMLguy, @cyrilXBT). Microsoft-Decision-1 is Qwen3.5-9B post-trained to return one option plus a calibrated probability, priced at $0.042 per million input tokens with free output. AnyJev turns any open causal LLM into a decision model with no fine-tuning by reading the next-token scores over the allowed answers, and corrects option-order bias without labels. Vega (800M and 4B) maps the question to an energy landscape with one valley per answer and lets a small physics engine settle a ball; the valley is the answer, its depth the probability. An Unsloth notebook trains your own on 4 GB of VRAM. The router pitch ("one Jev call in front of your agent and every task finds its tier") is the cost case, but no launch reports AUROC. See the routing summary.
Harness engineering (cluster of 4) (@rohanpaul_ai, @dair_ai, @bibryam, @undefinedKi; papers 2610.11655, 2610.11573). Stanford's "Harness Evolution Hits a Ceiling" labels failed agent runs by their first failure signal: loops and stalls are fixed by evolving the harness, but plans that are delivered and simply poor are fixed only by training weights (a LoRA on evolved runs cut 9B content failures from a quarter to one in twenty). DAIR's pick MemoType proves one retrieval strategy has a precision ceiling on mixed memory stores and routes each memory and query by type, up to 16% better Recall@1. The X article "12 Agentic Harness Patterns From Claude Code" and a 17-author Huawei survey (model plus a six-part harness) circulated alongside. See the harness ceiling summary.
Where inference dollars go (cluster of 4) (@deedydas, @rohanpaul_ai, @rohanpaul_ai, @StockSavvyShay). OpenRouter's public token stats converted to dollars give about $1.5B a year, roughly 1% of inference spend, with Anthropic at 36% of it this week and open-weight models at 26%. SemiAnalysis estimates Anthropic's inference margin near 88%, with subscriptions at about 50%, so subscribers get about 4x more compute per dollar than API buyers. The FT reports Nvidia is in early talks to acquire or further fund Reflection AI ($25B valuation, about $7.3B of compute committed). See the compute economics summary.
OpenAI's revenue restatement moves markets (@rohanpaul_ai). The Nasdaq 100 fell 1.4%, its worst day in seven weeks, and the Philadelphia Semiconductor Index fell 3.4%, even as OpenAI told investors run-rate revenue grew 77% in Q3.
AI-text detection breaks at model boundaries (@rohanpaul_ai, arXiv 2610.11599). The paper paired 4,000 pre-ChatGPT PNAS abstracts with rewrites by 23 model versions from three vendors. Detectors trained only on a vendor's past versions, set to falsely flag 1% of human text, catch over 99% of rewrites just before the sharpest generation boundary and 3.8% just after it. Vocabulary drift between versions predicts where detection transfers.
Six parallelism strategies for inference (@_avichawla). Data, tensor, pipeline, context, expert and hybrid parallelism, with the useful distinction that prefill context parallelism cuts time to first token while decode context parallelism splits stored KV history for capacity. See serving shorts.
Agent safety, two posts (@kimmonismus, @LauraRuis RT of @tomekkorbak). Nadella says companies should build AI systems assuming the model is compromised, since one AI watching another is still one black box checking another. OpenAI safety researcher Tomek Korbak writes that he was told leadership no longer trusts him and a security guard escorted him out. See the misalignment reports summary.
Mathematicians and LLMs (@bengoertzel). A former math professor is disappointed by the math community's reaction to LLM-written proofs, the flip side of The Decoder's report on mathematicians' shock at OpenAI's 700+ AI-generated manuscripts.
Skip. A memecoin browser tool, robotics hand demos, generic "learn LLM" links and Spanish-language Claude Code tip re-posts.