11 Oct 2026. OpenAI published 722 mathematics manuscripts produced by an undisclosed internal model, drawn from roughly 4,000 research questions grouped into 372 result families, with each result averaging about 3 hours of ChatGPT Pro thinking compute. The release covers papers, proofs and partial reasoning traces, but the model itself stays closed. In products, agent orchestration scaled sharply: Anthropic opened Claude Code Projects to all Pro and Max users, where a coordinator splits goals across parallel cloud sessions sharing one project memory, and added dynamic workflows to Claude managed agents that can coordinate up to 1,000 agents in a single run, detecting 66 of 70 hidden bugs in a 116k-line codebase versus 14 to 27 for a single agent. On training, Xiaomi's MiMo-V2.6 report describes agents building, auditing and grading their own tasks, lifting MiMo-V2.6-Pro from 58.4 to 72.6 on DeepSWE after $2.6M of RL. Elsewhere, the FT reports Nvidia is in early talks over open-weight model company Reflection AI at a $25B valuation.
Building a scorer on Alibaba's Qwen3.5-9B signals how Chinese open-weight base models are now infrastructure even inside Microsoft's own stack. Teams running LLM routers, agents or verification pipelines can plug it in via Microsoft Foundry or OpenRouter for calibrated probability outputs rather than text generation. Compared with general chat models, it targets fixed-option scoring tasks specifically.
小米这篇论文有意思，训练循环基本让智能体自己跑，还解决了智能体钻空子刷分的问题，DeepSWE 涨到 72.6。
OpenAI 拿没公开的模型一口气写了 722 篇数学论文，平均每篇烧 3 小时 ChatGPT Pro 算力，模型却藏着不给用。
The result suggests the agent harness — not just the foundation model — is a decisive lever for robot performance, with a near-doubling of RoboDojo achieved purely through system design. Developers building embodied or multi-agent pipelines can borrow the multi-point EEF prediction and feedback-subagent patterns without retraining models. It also points to a shift in research effort from model scaling toward recursive harness optimisation.
The release matters less for pen spinning than for the data loop: the glove's 120 Hz fingertip tracking and 526-point palm pressure grid turn human demonstrations into joint angles, which is the bottleneck for imitation-learning pipelines. Labs that already own a WUJI Hand 2, a camera and an NVIDIA GPU can reproduce the full sim-to-real setup today, since code, motion clips, trained models and printable pen files ship under Apache 2.0. Robotics groups in Singapore and Shenzhen evaluating dexterous-hand platforms can compare it against the 1,000 Hz control rate claimed here.
The key lesson is that document-grounded RAG underperformed while learning from recorded lawyer decisions with recency weighting drove accuracy to 91% — a blueprint transferable to code review or customer support systems. Builders of legal tech and enterprise agents can copy the feedback-loop architecture, including balanced positive/negative few-shot selection. Uber's choice to ship inside Word rather than a standalone tool is a adoption pattern worth stealing.
Claude Code Projects 推出正式版了，Pro 和 Max 用户都能用，多个线程并行干活还共享记忆，冲突会走 merge 处理。
Anthropic 给 Claude 智能体加了动态工作流，一次能同时跑 1000 个子智能体，找 Bug 测试里单智能体只找到 14-27 个，它稳定找到 66 个，很直观。
The design lesson is that tiered model routing beats a single flagship model: classification, routing and field extraction do not need top-tier reasoning, and cheaper Chinese models handle them well enough to cut inference cost by double digits. Engineering leads building internal agents can copy the four principles, especially single-pipeline architecture and shipping features behind default-off flags. It is also a concrete case study of small teams automating sales ops rather than headcount.
Claude Code 现在能让 Claude 自己写调度程序，一次指挥最多 1000 个 Agent 分阶段干活，还能查进度、设预算，做批量材料处理的可以试试。
The shift here is that evaluation expertise, not raw compute or model size, becomes the core asset for companies selling training data to frontier labs. Practitioners working on RL pipelines, synthetic data or benchmark design should watch how Mercor turns this advice into concrete evals and data products. Lambert's explicit bet on open models for post-training gives open-source teams a possible edge over closed-lab-only approaches.
The takeaway is that agent-mediated purchasing will spawn a new optimization industry targeting model preferences rather than search rankings, and unbiased evals for product quality are inherently hard to build. Practitioners building agent workflows or growth strategies should start treating model training preferences and review data as ranking factors. Compared with traditional SEO, the incentive shifts from gaming keywords to gaming the models themselves.
The argument points to a measurable shift in retail economics: if agent-driven search removes the cost of comparison, pricing power moves from brands that buy distribution to brands that win on quality, and subscription margins built on forgetfulness come under pressure. Product, growth and subscription teams in Southeast Asia's e-commerce and SaaS markets can use this to stress-test retention assumptions before agent adoption scales.
Nvidia 可能拿下 Reflection AI，用授权加挖人的方式绕开反垄断审查，Beam 这模型用更少算力对标 GLM 5.2，值得看看细节。
纳德拉亲自写的长文，给开发者列了七条可观测性原则，核心是别信模型、留好刹车，做Agent的可以对照检查。
The 73.43% figure matters less than the gap: prior automated reviewers caught under 15% of core-claim errors, so MLR moves claim-level checking from near-useless to partially reliable, while still missing roughly one in four. Conference and journal programme committees weighing AI-assisted triage can use the Contradiction Benchmark as a baseline before trusting any reviewer agent on submissions.
Splitting ideation into three RL-trained roles rather than one monolithic agent appears to be what drives the novelty gains, a design choice others building AI-scientist pipelines can copy with a 27B open model. Builders of autonomous research or agent workflows should compare it against their Claude Code or Codex SDK stacks. The paper and chat-with-paper links are public, so the evaluation methodology can be inspected directly.
The finding that a missing think tag or short reply does not mean a model skipped reasoning undermines a common evaluation shortcut. Teams benchmarking reasoning models or building answer-only constraints should re-check how compliance is measured, since suppressing visible reasoning costs roughly 15 accuracy points. The result spans 5 models, making it a broad behavioral pattern rather than a single-vendor quirk.
The most useful part is the practical diagnosis layer: entropy and advantage tracking, clip ratios and KL penalties shown with real training logs rather than equations alone, which is what most GRPO implementations get wrong. Researchers and engineers training reasoning models with open-source models can apply the stabilisation tricks directly to their own RLVR runs. Code and checkpoints are provided, so it works as a follow-along exercise rather than passive viewing.
The framing clarifies why reasoning-style models changed developer workflows: test-time computation replaces next-token completion, which explains the leap in coding-agent reliability. Engineers evaluating agent frameworks or choosing between reasoning and standard models can use this lens to judge where test-time reasoning pays off. The original chart on X condenses the argument into a single visual.
Knowing which phase a given optimisation targets — faster GPUs for prefill, smaller weights or faster memory for decode — turns technique lists like quantisation and speculative decoding into a coherent decision framework. Practitioners running local inference or serving stacks can use pp128 and tg64 benchmarks separately instead of a single speed number. The guide is hosted on llama.app and centres on llama.cpp tooling.
The lecture breaks down always-on agent design in plain language, which most public material on proactive agents does not do. Practitioners building autonomous or background agents can borrow the harness and context-engineering patterns directly. Watching the harness and context sections (10:06 to 34:26) gives the most actionable material.
Cost-aware routing in the agent harness, rather than a single frontier model everywhere, is becoming a practical default for coding agents. Teams building multi-step agents can benchmark cheaper models per task type and keep premium models for escalation only. The DeepSeek V4.1 Flash versus Opus split shows a concrete tiering pattern worth copying.
