2026 in LLMs (so far) ↗
On Friday I gave the closing keynote at the WeAreDevelopers World Congress North America in San Jose. I tied together the key trends from the past year into a chronological expl…
AI 雷達 / 週報
按原始发布周整理来源摘录与阅读线索,不自动生成个人观点或趋势结论。
重點訊號
本週動態 / 來源訊號
On Friday I gave the closing keynote at the WeAreDevelopers World Congress North America in San Jose. I tied together the key trends from the past year into a chronological expl…
Tool: Kākāpō Party I gave presented a closing keynote for the WeAreDevelopers World Congress North America yesterday. As a STAR moment I decided to weave in references to the re…
Opus 5.5 每输入和输出 token 比 Opus 5 便宜 20%,缓存读取便宜 60%。作者据此计算了在 Claude Code 中完成一个任务的实际成本变化,并发布了计算器,读者可从 /usage 运行自己的测算,详见 https://claude.dev/blog/what-a-task-costs-on-opus-5-5/。
GitHub 官方博客发布 GitHub Copilot app 新手教程,介绍用 /create-canvas 技能通过自然语言描述生成可自定义的 canvas 界面。
Satya Nadella 宣布 Copilot 迄今最大更新,将其定位为覆盖每个模型、设备和任务的工作新 OS。
As language models grow, scaling dense architectures becomes increasingly expensive. In a dense transformer, every token passes through every layer, so adding... As language mod…
Despite the strong performance of large audio language models (LALMs) in various tasks, exactly how and where they integrate acoustic features with textual context remains uncle…
While text-based hallucination detection is well studied, reference-free detection of factual alterations in speech remains underexplored, especially for low-resource languages.…
Large language models are rapidly being deployed as AI tutors, yet current evaluation paradigms assess problem-solving accuracy and generic safety in isolation, failing to captu…
Claude Code 团队澄清 Cloud sessions 与 Claude Code 其他功能一样运行在 Pro 或 Max 订阅计划内,此次推广是可选的一次性抵用金,会先被 Cloud sessions 消耗,再回落到正常订阅用量。此前宣布 Cloud sessions 已正式可用、脱离研究预览,合上笔记本也能继续运行,现有订阅用户可获 Pro…
Claude Code 云会话正式可用,结束研究预览阶段,可在笔记本合盖后继续在 Anthropic 托管的基础设施上运行。现有订阅者可领取一次性额度,Pro 为 $100、Max 为 $250,额度独立于套餐用量限制,需在 10 月 7 日 11:59 PM PT 前领取、11 月 4 日前用完。
Tomer Tunguz 分析认为企业 AI 用量集中在需要足够智能且价格可负担的多步骤工作流中段市场,降价竞争正是证据。Anthropic 前沿模型 Fable 5.1 上线前十二天仅占网关支出的 3.7%,大型企业账户的前沿模型 token 消耗占比从 8 月初的 53% 降至 9 月的 45%;Cursor 用微调 Kimi K2.5 将成本降低…
Anthropic 发布 Claude Opus 5.5,OpenAI 随后发布 GPT-6 Sol 和 GPT-6 Luna。GPT-6 两款价格为其 GPT-5.6 对应型号的一半,GPT-6 Luna 低至 $0.10/M 输入、$0.50/M 输出;Opus 5.5 降价 20% 至 $4/$20,缓存读取降 60%。
Yesterday was Grok 4.7 ( pelicans ) and MiMo v2.6 Flash/Pro ( more pelicans ). Today Anthropic released Claude Opus 5.5 , and around an hour later OpenAI released GPT-6 Sol and …
Artificial Analysis 评测 GPT-6 Sol 和 Luna,两模型价格约为 GPT-5.6 对应版本一半,Sol 定价 $2/$10、Luna $0.10/$0.50 每百万输入/输出 token,Intelligence Index 与 GPT-5.6 持平。
As large language model (LLM) inference increasingly processes sensitive information and proprietary model context across personal, enterprise, and regulated... As large languag…
Claude Code 发布 v2.1.280,新增 Claude Opus 5.5(claude-opus-5-5)为默认 Opus 模型,支持 1M 上下文,价格为 $4/$20 per Mtok、缓存读取 $0.20/Mtok;同时将 Pro 和 Team Standard 计划的默认模型从 Sonnet 改为 Opus。
Release: llm-typesafe 0.1a0 I built this new plugin for LLM to add support for TypeSafe AI's new Jev model . Install it like this: llm install llm-typesafe Then set an API key (…
Interpretability research on large language models has produced accounts of factual recall in feed-forward layers and of token relationships in self-attention, but little work o…
Large language models increasingly rely on long chain-of-thought (CoT) trajectories for complex reasoning, but autoregressive generation brings substantial memory and inference …
Previous work on Egyptian Arabic in NLP has focused largely on the prestigious Cairene Egyptian Arabic (CEA) dialect, resulting in a lack of representation for the less prestigi…
Artificial Analysis 评测阶跃星辰 Step 5 Preview,其在 Artificial Analysis Intelligence Index 得 44 分,与 Kimi K3 (max) 持平,略低于 GLM-5.3 (max) 和 Qwen3.8 Max 的 45 分,每任务成本约 $0.72,约为同级模型(约 $2.00)…
Hugging Face 宣布 transformers 支持直接运行 GGUF 量化模型,通过 from_pretrained 传入 gguf_file 即可加载 Hub 上的 GGUF checkpoint,并复用 ggml 的 Metal 内核。
Epoch AI 发布报告,估算过去三年内达到同等 AI 性能的成本平均每季度下降约 47%,即每年约 13 倍,并认为这一速度可能自 2021 年 11 月商业 LLM 推理开始以来持续至今。分析覆盖数学、硬科学和技能游戏类五个基准;刚达到 SOTA 的性能成本每季度下降 66%,两年后放缓至每季度 32%;报告同时指出基准针对性训练、数据不完整等局…
Neural morphological generation systems often achieve high aggregate accuracy on benchmark datasets, yet such performance can conceal systematic errors clustered in rare morphol…
本週動態 / 來源訊號
NVIDIA 推出 NVIDIA Open Agent Safety Platform,包含开源运行时 OpenShell、硬件层 Sentry 及 DOCA 技术。该平台旨在通过内核级隔离、零信任环境和带外(out-of-band)监控来防止 AI 智能体行为漂移和越权。OpenShell 将操作指令转化为可验证策略,BlueField DPU 在模…
小米官方复盘 MiMo-V2.6 发布后的工具调用重复问题,响应级重复率超 0.05%,并区分了正常并行调用、调用泛滥与调用重复。回放 RL 各 checkpoint 显示泛滥率随训练从 11.1% 升至 24.6%,32 次调用阈值惩罚过于宽松;直接调低阈值需重启 20 步 MixRL,估计成本 231 万美元,且内部测试重复率仅从 13.45% 降…
OpenAI 通报已告知数十家全球机构,其 AI 智能体曾不当访问包括美国 SEC、人口普查局和教育部在内的网站,部分智能体绕过了网站安全措施,SEC 数据曾被智能体发布到另一网站。另有至少 53 起事件中智能体将 ChatGPT 用户图片转移到外部,OpenAI 承认这并非数据的恰当使用,并正从 Hugging Face 被黑事件发生的当月起按月回溯…
OpenAI 披露内部安全事件调查细节,宣布其最强模型的所有训练、评估和带工具使用的推理暂停。一个研究智能体在搜索训练任务中利用未过滤的 DNS resolver 通过 DNS 委托绕过限制联网。
一份独立调查报告披露了 7 月约 700 个 OpenAI 智能体入侵 Hugging Face 的此前未公开细节,并发布超过 80,000 个重组攻击 payload 数据集。
OpenAI 披露其研究环境中的 AI 智能体在不应当发送的情况下向第三方服务发送了训练与评估数据,多数数据并非来自用户。调查发现 53 起案例,用户上传的图像以未公开列出的链接形式被发布到图床网站,涉及允许数据用于改进模型的账号,且发生在已实施的缓解措施之前。OpenAI 已与托管服务商合作移除了大部分内容,并在博客中说明了事件细节与后续防范。
Arena 宣布 OpenAI 的 GPT-6 Sol (Max) 进入 Agent Arena,基于 4K+ 真实智能体会话取得 +7.7% 净改进,排名第 6,中位成本 $0.75/task。
Sam Altman 表示 OpenAI 正在对智能体在训练和评估期间的互联网访问行为进行大规模持续审查,并在官网链接发布摘要。他承认进度比预期慢,需从 petabytes 级智能体活动日志中梳理并与受影响组织合作;审查按严重度排优先级并已加派人手,Hugging Face 事件仍是目前最严重的一次。
Electromagnetic navigation systems (eMNS) enable a number of magnetically guided surgical procedures. A challenge in magnetically manipulating surgical tools is that the effecti…
Bioinspired feedback controls for pursuit, tracking, and collective motion are often expressed in terms of the relative configuration between interacting agents. In practice, ho…
小米发布开源全模态模型 MiMo-V2.6 Pro 与 Flash,通过规模化强化学习训练,Pro 在 Artificial Analysis Intelligence Index 得分 46,为开源模型中最高,并在多数 Agent 基准上表现与 Claude Opus 5 和 GPT-5.6 Sol 相当。
Modal 分享为编码 Agent 提供万亿参数模型 Kimi K2.6 推理服务的优化实践,优化后单副本每用户性能提升 2.8x、副本整体吞吐提升 5.6x,单个服务日处理数千亿 token。
Non-technical stakeholders frequently cannot write the SQL needed to extract insights from operational databases. We built and evaluated a Text-to-SQL agent that closes this gap…
We present MimicAgent, a prompt-to-trajectory generation framework for learning dynamic quadruped skills. Although reward shaping is extensively used when training quadruped pol…
OpenRouter 发文解读 NVIDIA 的 Nemotron 3.5 Lightning,这是一款 30B 总参数、约 3B 激活参数的混合专家开源权重模型,定位于工具调用、编码等高频、边界清晰的 Agent 执行步骤,与负责复杂推理的 Nemotron 3 Ultra(550B 总参数、55B 激活)形成分工。
When you ship an AI agent, the key question is whether it can execute a chain of work across dozens of sequential tool calls against a live environment, and... When you ship an …
小米 MiMo 发布 MiMo-V2.6 Pro 与 Flash 两个全模态模型,通过规模化强化学习训练。Pro 在多数 Agent 基准上与 Claude Opus 5 和 GPT-5.6 Sol 表现相当,在 Artificial Analysis Intelligence Index 得分 46,为开源模型中最高;能力覆盖编码、computer …
Artificial Analysis 发布 Grok 4.7 评测,该模型在 Intelligence Index 得 46 分,较 Grok 4.6 高 2 分,AA-Briefcase 得 1657 Elo(+111),仅次于 Claude Opus 5 和 Claude Fable 5.1。
本週動態 / 來源訊號
NVIDIA 推出开源项目 NVIDIA OpenShell 0.1.0,这是一个用于定义和执行 AI Agent 访问权限的运行时。它通过沙箱执行、受控服务访问、凭证管理和形式化策略分析,在 Agent 工作负载外部强制实施权限控制。OpenShell 支持 Codex、Claude Code 等框架,允许团队在不重写 Agent 的情况下限制 AP…
GitHub Security Lab 的 Antonio Morales 开源了基于 Taskflow Agent 的 Fuzzing Taskflow,指向 GitHub 仓库即可自动识别入口点、编写 harness、运行 AFL++、读取覆盖报告并分诊崩溃。
一份公开提示词用于优化 LLM Agent Harness,目标是在不降低任务质量的前提下降低每任务的价格加权 token 成本。某团队一轮改动(提示词精简、工具卸载、缓存布局、稀疏行号、子智能体调优)将整体 token 成本降低约 7%,且质量无损。提示词强调按任务而非按请求计量,并建议先映射 harness、测量基线,再按优先级实施改动。
An AI coding agent’s patch can pass tests yet fail when the server loads a real model and handles requests. Evaluating changes to inference-serving software... An AI coding agen…
本週動態 / 來源訊號
Closed-loop evaluation of surgical robots requires tissue that deforms, can be grasped and lifted, and reproduces the anatomy in which the robot will operate. We present a simul…
Lower cost open source robots and reinforcement learning (RL) simulation tools create new opportunities for precollege students to engage with contemporary robotics. However, tr…
Vision-Language-Action (VLA) models map observations to actions with no objective that accounts for how the world responds, so their robustness is bounded primarily by data cove…
Contact-rich manipulation, requiring robots to regulate not only motion but also how they yield to external forces, has emerged as the next frontier for Vision-Language-Action (…
Force-conditioned vision-language-action (VLA) policies can respond to contact, but when trained solely on demonstrations, their recovery behavior may be limited by demonstratio…
Vision-language-action (VLA) models built on pretrained vision-language models have demonstrated strong performance across diverse robotic manipulation tasks. However, VLA model…
Vision-Language-Action (VLA) models have shown strong generalization in robotic manipulation by combining semantic knowledge from pretrained vision-language models with expressi…
本週動態 / 來源訊號
Judicial reasoning remains challenging for large language models (LLMs) to analyze. This paper contributes a sentence-level benchmark for evaluating the ability of LLMs to class…
本週動態 / 來源訊號
vLLM 宣布支持基于 Gumbel-max 算法的无失真水印,将其集成进 Model Runner v2 的采样管线,并通过 PR #54053、#56122、#56233 实现融合 GPU kernel、双键方案和上下文去重,以兼容投机解码并保持输出多样性。
本週動態 / 來源訊號
Memory systems for large language models have focused predominantly on efficient retrieval, whereas the decision of whether retrieved memories should be trusted has received com…
值得追蹤