Latest
100 位 Apple 生态专家聊 AI 怎么用的报告公开了，Xcode 这块相对独立，跟生态外的体感可能不太一样，做 iOS 开发的可以对照看看。
Confidence thresholds fitted on your own agent's traces, not generic accuracy, are what decide whether a small router model helps or hurts. Practitioners building tool-using agents can cut costs by training a small decision model on trajectories their own agent actually reaches, mistakes included. Compared with routing every tool call blindly, the thresholded approach preserved success within 3 points.
LobsterAI 按月公布自家办公 Agent 的真实使用数据，9 月任务规模、用途分布、缓存成本差 7 倍的原因都在里面，做产品或运营的可以拿来当参考。
The takeaway is that cost optimisation for security agents should start with token accounting, since generic compression tricks usually break the reasoning that finds bugs. Teams building vulnerability-hunting agents can instrument their own runs and try AVRI-style code-reuse to cut spend. The measurements on Codex and OpenCode give a concrete baseline to compare against when evaluating cheaper agent pipelines.
QA-only evaluation cannot tell whether failures come from memory writing, retrieval or answer generation, so teams may ship agent memory systems that look fine in testing yet mis-store facts. Builders should add direct checks of stored facts after each update, using FEB-30-style scripted histories where ground truth is known. The 20.1% correct-action rate is a concrete red flag for anyone reusing Mem0 or Hindsight architectures without stage-level instrumentation.
Cohere 的新论文，实测多智能体写代码怎么省钱：便宜的模型只做规划，强的模型当 lead，成本能砍 70% 掉不了几分。
字节发布的 Agent 使用报告，重庆人凌晨还在跑任务占 19.1%，单次任务从 6.3 万涨到 20.2 万 Token，Seed 模型占了图像视频生成一半以上。
网易 LobsterAI 拿真实用户数据写的报告：微信是最大入口，七成人只用一个模型，跟 X 上的玩法差别挺大，做 Agent 产品的可以看看。
LobsterAI 这份报告挺有意思：假期里大家拿 Agent 干最多的居然是编程，使用高峰还挪到了晚上 10 点，还扒出了大家常用的省钱模型组合。
Anthropic 做了 3 年可解释性研究的人聊 Claude 内部怎么运作，发现模型会提前规划好几个 token，还挺颠覆直觉的。
Verified-by-construction environments sidestep the usual reward-hacking problem in RL for open-ended analysis tasks, a reusable recipe for teams training agents beyond fixed-answer benchmarks. The diagnostic finding that frontier models compute correctly but conclude wrongly matters for anyone building data-analysis agents who should audit the reasoning step, not just the arithmetic. The paper is available on arXiv (2610.21656) with a Chat with Paper interface at academy.dair.ai.
Anthropic 研究员聊了三年扒开 Claude 内部看运行机制的事，居然发现模型会提前规划好几个 token，听完挺刷新认知的。
The result suggests frozen LLMs can accumulate and reuse knowledge over long interactions without retraining, pointing toward cheaper continually learning systems. Developers building RAG pipelines or agent memory should compare this compression approach with Mem0 and MemoRAG before defaulting to full-context retrieval, since it wins most head-to-head benchmarks at a fraction of the token cost.
OpenAI 一次发布 300 多个 AI 解出的数学结果，把密码学家吓到了——如果 AI 真能啃动数论难题，我们的支付和聊天加密可能受威胁。
dair-ai 每周论文榜来了，这次 7 篇里有三篇 Harness 系列，还有智能体科学家论文，搞研究的可以翻翻。
Diffusion language models are mostly benchmarked on GPUs, so this shows a practical path to running one locally on consumer NPUs where power budgets matter. The measured 8.9x energy gain over CPU INT4 is the kind of number edge-deployment engineers can use when sizing on-device inference. Anyone working on AMD Ryzen AI hardware or quantized LLM deployment should watch for the full poster at PyTorchCon North America 2026.
Anthropic 自己员工的说法：六成人说 AI 已经把合适任务的活儿干得只用原来四分之一时间，还预测五年后七成五工时交给 AI。
The non-obvious move here is replacing score-based feedback with same-task A/B behavioural comparisons, which sidesteps the missing-verifier problem that blocks most self-improvement systems. Engineers building agent scaffolds can study the Propose–Probe–Compose loop and reuse its evidence-gated lesson library design in their own stacks. As an open-source release from Google Research, it is directly comparable to other RSI and agent-tuning frameworks that depend on task-specific verifiers.
Anthropic 自家员工说 2030 年 AI 能干掉自己 75% 的活，比美国大众的 28% 乐观太多，这个认知差距本身就很说明问题。
The best single-game performer was not the most efficient learner, so one-shot benchmark scores can mislead when choosing an agent for long-horizon work. Teams building self-improving systems should track held-out learning curves against compute cost and verify that saved tools and notes are actually reused. The frozen-weights design makes the results directly relevant to anyone running agents under a fixed budget.
Anthropic拿自己公司开刀做研究：Agent工时已超人类6倍，初级岗从27%砍到9%，招人标准也变了，数据很具体。
新加坡国立大学的 Simon Chesterman 写了一篇关于 AI 智能体责任归属的文章，讨论出事了到底算谁的，观点挺扎心。
The key insight is training the decision model on trajectories generated from the agent's own mistakes, which lets a 4B model outperform the larger Jev on out-of-distribution tasks. Builders running agent harnesses with high token bills can apply the fitted-threshold cascade to shave costs without touching task outcomes. The paper is hosted on DAIR.AI's open academy site, so the full write-up is freely available.
The non-obvious trick is labelling training data with the agent's own post-mistake states, which makes the small decision model generalise beyond familiar trajectories. Builders running agentic pipelines can replicate this distillation-and-cascade pattern with any open 4B model to slash inference cost without changing outcomes. Compared with fixed-threshold routing, fitting the confidence threshold on the agent's own trajectories is what preserves task results.
The reward-design audit is the most transferable part: dense progress rewards let failed runs harvest half the reward, so teams training RL agents should measure how much reward failure earns before trusting their curves. RL and agent-training engineers can apply the strict-success-versus-partial-credit split to their own reward pipelines. Compared with generic terminal-agent RL reports, this one pairs benchmark gains with a concrete failure-mode analysis.
斯坦福两位教授和 François Chollet 一起聊 AI 安全，把飞机防撞系统的验证方法搬到 LLM 上，还有用 Lean 证明代码正确的 CSLib 演示，硬核但好懂。
The methodology detail matters as much as the gains: fixed short SFT runs replaced proxy metrics for evaluating midtraining data, a practice teams can copy when tuning their own pipelines. The full recipe and ablations are public, so labs and engineers training coding agents can reproduce or adapt the stage directly. NVIDIA's Nemotron models also give an alternative to Qwen- and DeepSeek-based coding stacks for agentic workloads.
RL environment counting has become a reporting gap: different labs may describe the same training run with wildly different sandbox numbers, making cross-lab comparisons unreliable. Practitioners tracking frontier RL scaling should read both the DeepSeek paper and the Beam blogpost before citing either figure. The expectation that V4.1F-1017 outperforms Beam sets up a concrete head-to-head comparison to watch.
Verified-evaluation loops turn LLMs from code generators into search engines for scientific design, a pattern transferable beyond quantum computing to any domain with a cheap automatic checker. Researchers in circuit synthesis, evolutionary search and automated discovery will find the divsteps result a concrete demonstration. The tasks span chemistry, optimisation and number theory, showing breadth across domains rather than a single cherry-picked benchmark.
Microsoft 在 Nature 发的天气预报研究，把可预测范围推到 2-6 周，思路是给模型配验证系统，搞 ML 的可以看看。
The takeaway is that attention architecture choice has become a strategic fork: American spinouts standardise on SWA while Chinese labs treat attention design as proprietary IP. Engineers evaluating open-weight models should look past parameter counts and check which attention variant a model uses, since it shapes cost and context behaviour. The comparison with Mistral's stack gives a concrete reference point for judging the trade-offs.
Benchmarks that pause the world reward slow, careful reasoning and hide latency penalties, so leaderboard rankings may invert under realtime conditions. Developers building latency-sensitive agents can copy the TORSI pattern: keep a fast model in the live loop and let a stronger model write and test its instructions offline between rounds. The result applies to any realtime control task, not just games.
Ethan Mollick 分享的一篇论文，用 Gemini 4 智能体组队打 Minecraft，看 AI 怎么自发分工，跟人类协作思路还挺不一样。
The interesting point is that polysemy in single neurons may be tractable: a neuron can hold two distinct meanings that each form clean, readable peaks rather than an indecipherable superposition. Interpretability researchers and anyone studying mechanistic interpretability of open-weight models like Qwen can use this framing to defend multi-concept neuron analyses against claims of overfitting. Qwen's open weights make these findings reproducible for practitioners anywhere, including ASEAN research groups working with Chinese open-source models.
LlamaIndex 挑了篇长文档检索论文 VL-DocIR，用 LlamaParse Agentic Plus 解析，连跨列表头和 35 个数字都没丢，做文献处理的可以试试。
OpenAI 放出 300 多道 AI 解出的数学题结果，密码学家坐不住了，说再突破可能动摇支付和加密货币的加密根基。