Nathan Lambert 谈 Artificial Analysis v5 与基准制定者的困境
Nathan Lambert 撰文分析 Artificial Analysis 智能指数 v5 更新及其带来的评测困境。AAv5 将指数均分为知识工作、编码、科学、通用四类各占 25%,榜单显示 8 个闭源模型超过最强待发布开源权重模型 Kimi K3.1 Pro。
Nathan Lambert 撰文分析 Artificial Analysis 智能指数 v5 更新及其带来的评测困境。AAv5 将指数均分为知识工作、编码、科学、通用四类各占 25%,榜单显示 8 个闭源模型超过最强待发布开源权重模型 Kimi K3.1 Pro。
Hugging Face 首席政策官 Irene Solaiman 在 TechCrunch Disrupt 上表示,嵌入式 AI 评估器并非简单解法,评估结果的可复现性仍是核心难题,开放权重对此至关重要。
Microsoft 合伙人研究经理 Jennifer Neville 在 Microsoft Research Podcast 中讨论了评测如何推动 AI 系统突破传统 benchmark 的性能边界,以及模型在超出常规测试时出现的"意外失败"。她结合自身从数学、物理、认知科学到计算机科学的经历,分享了使用当前 AI 系统的实用建议,并强调当结果偏离预期时深入检查数据的重要性。
据彭博行业研究报告,DeepSeek 9 月发布 V4.1 Flash 后,中国顶尖模型在基准测试上仅落后美国对手 3%,较 5 月约 9% 和今年早些时候 15% 明显收窄。
GPT-6.1 Sol:以五分之一的价格提供接近 Astra 的智能。 它是当今可用的、在其性能水平上最具成本效益的模型。
推荐理由:原文给出智能指数与单任务成本的 Pareto 前沿新点位和具体分数价格,可用于横向比较各家模型性价比。
Trail of Bits 发文批评 1Password 8月6日发布的 FLAWED 报告,称其 26% 的 AI 干净修复率受四项实验设计选择影响而失真,包括刻意指示智能体应用错误修复的提示词占 22% 数据、36% 的试验禁止编译测试,以及不同推理档位设置。
推荐理由:原文针对1Password基准的26%头条数字给出逐项方法学批评,并补充自身人类修复失败率与开源合并数据作对照。
We've never seen this before. The biggest jump in Vending-Bench history. GPT-6 Astra is better at making money and more ethical than Claude Fable 5.1. Surprising, because: 1. First time ever that OpenAI is #1 on Vending-Bench 2. The best model is no longer the unethical one.
Muse Spark 1.3 Max by @AIatMeta has reshaped the Pareto frontier for Code Arena: WebDev! Meta's latest model at Max reasoning is doing something interesting on the Arena Pareto frontier: it's the only model holding down the wide price band between Qwen3.8-max ($5/MToken) and Qwen3.8-Flash-Next ($0.39/MToken). The next closest model to the frontier is Meta's own, Muse Spark 1.3 on xHigh reasoning. Muse Spark 1.3 Max performs just 20 pts lower than Qwen3.8 (Max) while costing 30% less, and 24 pts from Kimi K3 (Max) while costing 70% less. - Muse Spark 1.3 Max: 1650 pts | $3.50/M - Qwen3.8 (Max): 1670 pts | $5/M - Kimi K3 (Max): 1674 pts | $12/M Overall in Code Arena: WebDev, Muse Spark 1.3 Max landed #8, ahead of Claude Fable 5 at #10 (+22 pts), Grok-4.6 (High) at #12 (+26 pts), and GPT-5.6 Sol at #14 (+33 pts). Congrats to the @AIatMeta team on this release!
Arena(arena.ai)平台发布 2026 年第 36 周(8 月 31 日~9 月 6 日)AI 大模型盲测排名。
GPT-6 Astra has set a new ECI record, with a score of 169. This is a substantial jump from the prior best (163), but is within our uncertainty range for the reasoning-era ECI trend. Astra also set new records on our math, continual learning, and game-puzzles benchmarks. On our long-horizon coding benchmark, MirrorCode, Astra ranks between Opus 4.7 and Fable 5. OpenAI gave us pre-release access to test Astra. Charts and more details for Astra’s individual benchmark results in the thread.
GPT-6 Astra debuts on APEX-Agents at #1 on the leaderboard. 🥇 62.4% Mean score (#1) 🥈 46.7% Pass@1 (#2) It leads Fable 5.1 by 0.4 points on mean score and trails it by 0.8 on Pass@1. APEX-Agents runs long-horizon tasks written by bankers, consultants, and corporate lawyers in our expert network. Against its predecessor GPT-5.6 Sol, Astra moves up 6 places on our leaderboard. It gains 5.7 points on mean and 6.7 on Pass@1. All three domains moved by at least 4.5 points, so this is a broad gain rather than one job pulling the average up. Congratulations to @OpenAI. See full leaderboard: https://www.mercor.com/apex/apex-agents-leaderboard/
推荐理由:ARC Prize 作者基于自家标准 harness 的实测数据评估 GPT-6 Astra,读者可对比 66% 与近 100% 两种设置看模型与 harness 能力的边界变化。
Fable 5.1 from @AnthropicAI on ARC-AGI (Verified): - ARC-AGI-2: 90.0%, $3.12/task - ARC-AGI-1: 97.5%, $1.40/task Across the two benchmarks, Fable 5.1's average cost per task is about 32% lower than Fable 5's, driven by better token efficiency.
Fireworks AI 发布对 DeepSeek V4 Pro 0813 的 CyberGym 评测,称其在 840 个对抗性安全任务中零拒绝,每次成功成本约为所测最高分模型的一半。
医疗 AI 面临“预言机问题”:临床决策主观、底层数据不透明、评测易沦为营销,导致医院难以客观评估模型优劣。OpenAI 报告每周超 3 亿人用 ChatGPT 咨询健康问题,2026 年每百万条病历中约 1,686 条记录患者对模型回答的追问。a16z 正与 Protege 合作,试图通过追踪患者长期结局来解决这一评估难题。
Epoch AI 在 Gradient Updates 中列出 9 个关于 AI 能力的大问题,涵盖 AI 能否胜任完整工作、哪些领域将出现下一个"Claude Code 时刻"、开源与闭源及中美模型差距是否跨领域一致、以及基准分数为何普遍高度相关。
Big news: DeepSeek-V4-Pro (Max) by @deepseek_ai is coming in around ~#8 overall (#2 among open models) in the Code Arena: WebDev! At 1607 pts, this places it after GPT-5.6 Sol (xHigh)(1622 pts), and makes it the second best open model after Kimi K3 (Max) (1674 pts). Note: this is an early AutoEval score, in which a Reward Model trained on Arena's human preference data casts automatic votes in place of live votes. We’ll continue to see how scores converge as more live human votes come in. See thread for Text Arena scores and more on AutoEval's methodology. Congrats to the @deepseek_ai team on this release!
SemiAnalysis 发布 TileRT InferenceX 长文分析:TileRT 将整个 decode 图静态编译为单一持久 kernel,在单台 B200 上跑 InferenceX GLM5 FP8 744B 达到最高 500 tokens/s/user,约为传统引擎的 3 倍,iso-cost 下交互性最高快 2 倍。
Exciting news: Muse Spark 1.2 (xHigh) by @AIatMeta is #4 in the Text Arena (1498 pts), and has reshaped the Pareto frontier! It is priced at $1.25/$4.25 per MToken. Congrats again to the @AIatMeta team on this release!
Siri AI 已随 7 月发布的 iOS 27 消费者测试版上线,能理解个人上下文、调用世界知识回答问题,并支持自然对话、打字输入及独立应用。该功能由 Apple 与 Google 合作,利用 Gemini AI 模型训练自有 Apple Foundation Models,可在 Apple Silicon 和 Private Cloud Compute 上运行。
Andrej Karpathy 让 Claude Opus 5 将《指环王》开头一段文字转化为 3D 浏览器场景,生成 5,500 行代码,耗时约两小时,花费约 10 美元。模型自主放置和动画化物体,但因只能通过截图检查视频输出而出现错误。Karpathy 认为这展示了按需生成廉价定制游戏世界的潜力,并称鹈鹕测试已接近上限,此类结果更多是氛围测试而非硬性基准。
Turns out GPT-5.6 Sol is actually SoTA on ARC-AGI-3. Just took two setting changes. You just have to allow it to reason and work over multiple context windows with the help of our canonical compaction implementation. https://openai.com/index/how-two-settings-tripled-our-arc-agi-3-scores/
Kimi K3 debuts at #3 on DeepSWE. It's the first open-weights model that delivers frontier-level performance, achieving results similar to Claude Fable and GPT-5.6 Sol.
GPT-5.6 Sol sets a new state of the art in cybersecurity on “The Last Ones” cyber range. We’re already seeing that capability translate into defensive outcomes: helping teams find, validate, and fix vulnerabilities in real-world code. Put it to work with Codex Security: https://openai.com/daybreak/codex-security-plugin/#desktop-codex
推荐理由:八天四个前沿模型,Frontier 从两家实验室变成六家,前三名分差仅三分,而且价格正在暴跌,模型选型的逻辑从「谁第一」变成了「够用且便宜」——对产品经理来说是真正的转折点。
Cursor 的模型评估负责人 Nate Schmidt 发现,Claude Fable 5 在其内部基准 CursorBench 上以 Max effort 模式达到 72.9%,创下新高。该模型在模糊的真实编程任务中表现出全局推理能力,例如在航天模拟器中仅凭一句提示自主规划并成功登月,而此前 Claude Opus 运行 12 小时以上仍无结果。
推荐理由:Cursor 工程师用自家基准和太空模拟器实打实测试了 Claude Fable 5,从全局推理到自主规划路径的能力跃迁很具体,做编码 agent 的可以看看这个‘p99 问题’解法。
Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5. This is a 17-place jump from Kimi-k2.6 (#18 -> #1). In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5. The full model weights will be released by July 27. Congrats to the @Kimi_Moonshot team on this major milestone!
Encore un scandale. On accuse Elon Musk sans arrêt de se servir de Grok pour faire passer ses idées. Pourtant, le @neutralityorg vient de sortir The Neutrality Project : une étude indépendante qui prouve l’inverse. Sur 18 modèles IA testés, 54 positions sur 60 atterrissent à gauche du centre. Moyenne globale : -0.41 (biais progressiste massif, surtout environnement et valeurs sociales). Et Grok 4.5 ? -0.02. Le modèle le plus neutre de tout le benchmark. Le seul vraiment proche du centre. Pendant ce temps, les autres boîtes injectent massivement leur idéologie de gauche dans des IA qui répondent à des milliards de questions par jour. Manipulation des masses à grande échelle. C’est absolument scandaleux. Ça devrait être un scandale mondial. https://x.com/neutralityorg/status/2076028460066283859