跳到正文
#推理
10月14日周三
  1. Google DeepMind：Blog85

    Google DeepMind 发布 Gemini 4 Argon，全球上线 API、Gemini 应用和 Antigravity

    Google DeepMind 发布 Gemini 4 Argon 并全球开放，定价为每百万输入 token $2、输出 $10（缓存输入 95% 折扣），支持 1M 统一输入输出 token 上限，并在 Vals Index 排名第一。

    同一件事，精选收录的是《Google 发布 Gemini 4 Argon：面向复杂工程与科研任务的前沿模型》

10月13日周二
  1. OpenRouter42

    Scaledown 的约束解码模型家族（Summarize、Extract、Classify-1）已在 OpenRouter 上线，每个模型专为单一任务打造，支持长文本/图片/PDF 摘要、自然语言定义字段提取 JSON、自定义标签分类。发布期 8 折优惠：输入 $0.04/M tokens，输出免费。

    引用Neal Patel@neal_k_patel

    You already gave the LLM the answer. So why is it writing it out from scratch? Introducing a new AI model architecture: Constrained Decoding Models (CDMs) LLMs were built to write. Most production workloads just need an answer: fast, consistent, and affordable. CDMs are purpose-built for 1 task each. They read your input once and skip the token-by-token generative loop that makes LLMs slow and expensive. Not just a decision model, but a family of 3 CDMs live today: • Classify: define your labels and get back a calibrated probability + reason for each. Used for routing, tagging, and intent detection. • Extract: describe the entities in natural language and get back well-formed JSON with confidence scores + citations. Used for web scraping, document processing, and PII redaction. • Summarize: pass in long text, images, or PDFs and get back a consistent summary under 1,000 tokens. Used for context compaction, metadata generation, and transcript digests. Built for production: • 1M tokens of multimodal (text + image) input, 31x larger than Jev • 77.2% on DecisionBench vs Jev's 72% • 200x faster, 467x cheaper than frontier LLM APIs • $0.05/M input tokens. Output is free (yes, even for summarize) Live on your favorite platforms today: @OpenRouter @vercel_dev @LiteLLM @LangChain @TrustedRouter @itsArthurAI @useRouterPlus Get 50M input tokens for free (for the first 50 hours) at link below 👀

  2. Rohan Paul54

    ScaleDown 发布 3 个面向单一任务的 Constrained Decoding Models（Classify、Extract、Summarize），不做逐 token 生成，只对输入做一次 prefill 并从原文中直接选出答案。

    引用Neal Patel@neal_k_patel

    You already gave the LLM the answer. So why is it writing it out from scratch? Introducing a new AI model architecture: Constrained Decoding Models (CDMs) LLMs were built to write. Most production workloads just need an answer: fast, consistent, and affordable. CDMs are purpose-built for 1 task each. They read your input once and skip the token-by-token generative loop that makes LLMs slow and expensive. Not just a decision model, but a family of 3 CDMs live today: • Classify: define your labels and get back a calibrated probability + reason for each. Used for routing, tagging, and intent detection. • Extract: describe the entities in natural language and get back well-formed JSON with confidence scores + citations. Used for web scraping, document processing, and PII redaction. • Summarize: pass in long text, images, or PDFs and get back a consistent summary under 1,000 tokens. Used for context compaction, metadata generation, and transcript digests. Built for production: • 1M tokens of multimodal (text + image) input, 31x larger than Jev • 77.2% on DecisionBench vs Jev's 72% • 200x faster, 467x cheaper than frontier LLM APIs • $0.05/M input tokens. Output is free (yes, even for summarize) Live on your favorite platforms today: @OpenRouter @vercel_dev @LiteLLM @LangChain @TrustedRouter @itsArthurAI @useRouterPlus Get 50M input tokens for free (for the first 50 hours) at link below 👀

  5. Rohan Paul45

    SubQ 推出 SubQ Snap，号称首个面向企业真实工作的长上下文决策模型，支持 384K 多模态上下文，是 Jev 的 12 倍。其延迟随上下文线性增长：100K tokens 660ms、384K 2.5 秒，而标准二次注意力在该区间成本约涨 15 倍。

    引用Alexander Whedon@alex_whedon

    Introducing SubQ Snap, the first long-context decision model for real enterprise work. Snap is the only decision model or LLM where compute per token stays constant as context grows. • 384K tokens of multimodal context, 12x larger than Jev. • 61.7 on Decision Index v0.2.1 vs Jev’s 57.9, outperforming on 13 of 16 enterprise benchmarks. •Processes 100K tokens and returns a decision in under one second. •$0.05/MM input tokens. Output is free. Available today to anyone. The first 10,000 sign-ups get 100M free input tokens. We’ll likely hit capacity fast and will add users as we add compute. Compute offers welcome 🙂. Latency will be variable as we tune the platform to this new model and category of users. Model speeds are 80ms for minimal context, 250ms for 32k, 660ms for 100k, and 2.5 seconds for 384k. Visit: https://subq.ai/

  6. Amazon Science61

    AWS 提出 priming 方法，可将预训练 Transformer 改造成混合状态空间架构

    AWS 团队发布论文《Priming: Hybrid state space models from pre-trained Transformers》，提出 priming 方法，复用预训练 Transformer 权重构建混合架构，将预训练成本降低逾 99%，全流程消耗不足源模型 0.5% 的 token 预算。

    推荐理由：AWS 团队提出 priming 方法，把预训练 Transformer 改造成混合架构的预训练成本降低逾 99%，并开源了 8B 和 32B 模型与代码。

10月12日周一
  3. 公众号：卡尔的AI沃茨58

    大AI理论：分三阶段调推理档位榨干模型能力

    作者提出“大AI理论”，用三阶段对话把模型能力拉到极致。第一步拉满推理并选计划模式，用语音、图片、视频不过滤倒出想法整理成计划；第二步切中档推理快速干活，前提是需求描述足够详细，作者引Terminal Bench 3.0说明详细需求加中档推理质量高于模糊需求加高档推理；第三步把推理调回高档让AI审查中档成果，称能揪出约20%的问题，审查耗时约为从头生成的三分之一。

10月10日周六
10月9日周五
  2. IT之家·人工智能73

    陶哲轩批评 OpenAI 发布 719 篇 AI 数学证明：人类来不及理解和消化

    陶哲轩于 10 月 7 日在 mathstodon 发帖，批评 OpenAI 10 月公布 719 份 AI 数学解答（涵盖 372 个结果族）的做法，称这会把开放问题变成可批量消耗的资源。他不反对 AI 辅助研究，但反对把快速攻克著名难题当产品展示，指出约 42% 证明未经形式化处理、仅 10 份附推理链，未达到 AGMAI 建议的披露标准，并担心出现大量无人能消化的证明。

  3. Arena.ai71

    Arena 公布 Anthropic Claude Haiku 5.5 (High) 的真实评测结果，在 Code Arena: WebDev 以 1587 分首秀排名第 30，略在 Pareto 前沿之外。

    引用Claude@claudeai

    隆重推出 Claude Haiku 5.5：我们迄今为止发布的最便宜、最快、最强大的小型模型。 平均而言，它的运行成本比 Claude Haiku 4.5 低约 75%。

    推荐理由：榜单方给出 Haiku 5.5 在 Code Arena 的实际得分与各档 Claude 模型的成本分差，便于横向权衡性价比。

10月8日周四
  2. Ant Ling43

    感谢我们的 day0 合作伙伴 @novita_labs 持续、可靠的发布支持 🥰 @novita_labs：🚀 来自 @AntLingAGI 的 Ling-3.1-Flash 现已通过 Novita 在 @OpenRouter 上线。 作为 Day-0 合作伙伴发布 ⚡ 🎁 10 月 13 日上午 9:00（PT）前免费。 560B 总参数 · 25B 激活参数。 专为混合推理和工具调用工作流打造。

    引用Novita AI@novita_labs

    🚀 Ling-3.1-Flash from @AntLingAGI is now available via Novita on @OpenRouter. Launching as a Day-0 partner ⚡ 🎁 Free until October 13 at 9:00 AM PT. 560B total parameters · 25B activated parameters. Built for hybrid reasoning and tool-using workflows.

  5. François Chollet45

    Chollet 指出，80 年代心理学研究曾假设学编程能像"思维体操"一样提升通用推理与规划能力，但基础研究及后续元分析证明这一迁移并不发生：学编程者算法思维与编码能力变强，收益却不迁移到通用认知任务。数学训练同理，只提升数学演绎与结构映射，不提高无关领域的基线学习速率。他认为通用智能是大脑的固有属性，而非可训练的技能。

    引用François Chollet@fchollet

    What if the jagged frontier is mainly math + code (which you can push arbitrarily far with RLVR), and everything else starts to plateau because it is still bottlenecked by human generated data? Model performance in non-verifiable areas has kept improving steadily, albeit much slower than for math and code. But is that steady improvement a side effect of a higher G (itself driven by RLVR), or only a function of the amount of new human data getting injected into training (which is still continually happening on a massive scale)? A lot of things depend on the answer to this question

  6. 公众号：数字生命卡兹克81

    OpenAI 在 GitHub 发布 openai/math 仓库，722 份数学手稿引发数学界震荡

    OpenAI 于10月7日在 GitHub 发布 openai/math 仓库，收录其未公开内部模型产出的722份数学手稿，归为372个成果组，覆盖数论、代数几何、偏微分方程等方向，每个结果平均花费约3小时 ChatGPT Pro 级别思考算力。

    推荐理由：作者把 OpenAI 的数学成果放进从四色定理到 Lean 的机器证明史里讲，读者能看懂这次争议为何激烈。

  7. Noam Brown47

    Noam Brown 指出 LLM 已跨过重要门槛，在部分研究问题上超越顶尖人类专家，这解释了近期数学成果为何显得突然爆发。他认为模型能力仍有明显短板，但随着持续进步，其他领域的突破将接踵而至，开启科学发现新时代。

    引用Mark Chen@markchen90

    The Navier-Stokes moment was always much more about the figure below than about the Navier-Stokes problem itself. The run represents a decade of mathematical progress in a week. Can't wait to point these tools at life sciences, building our next models, and alignment!

10月7日周三
  2. Greg Brockman55

    OpenAI 发布一系列由内部前沿模型产出的新数学结果，仓库地址 https://github.com/openai/math。发布前咨询了高等研究院数学与人工智能独立顾问小组，并参考其建议和公开推荐来决定发布方式。Greg Brockman 转发此事，称其朝加速科学发现和改善每个人生活质量的方向迈进。

    引用OpenAI@OpenAI

    We’re releasing a broad range of new mathematical results produced by an internal frontier model. We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results. https://github.com/openai/math

  3. Rohan Paul65

    OpenAI 在 GitHub 公开发布内部前沿模型产出的数学论文与证明材料，共 722 篇手稿，整理为 372 个结果族。测试时模型被喂入约 4K 个研究问题，每项结果平均使用相当于 3 小时 ChatGPT Pro 思考的算力；部分结果尚无 Lean 形式化，可能存在问题，仓库将持续更新。

    引用OpenAI@OpenAI

    We’re releasing a broad range of new mathematical results produced by an internal frontier model. We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results. https://github.com/openai/math

  4. Aravind Srinivas40

    另外，Perplexity Decision API 的价格减半了，现在每百万 input tokens 只要 2 美分。尽情享用！

    引用Perplexity Developers@perplexitydevs

    pplx-decider-v1.1-27b, our updated open weights multimodal decision model, is now available. It scores the highest on the new @huggingface Decision Index 0.3 benchmark. pplx-decider-v1.1-27b costs half as much as v1, at $0.02 per million input tokens. https://huggingface.co/spaces/multimodalart/jev-decision-index

10月6日周二
  1. ARC Prize67

    ARC Prize 公布 DeepSeek V4.1 Flash 在 ARC-AGI (Verified) 的成绩：ARC-AGI-2 得分 72.9%，每任务成本 $0.13；ARC-AGI-1 得分 94.5%，每任务成本 $0.07。相比 V4 Flash 的最佳成绩，ARC-AGI-1 高 5.5 分，ARC-AGI-2 高 11.5 分，但每任务成本高约 250%。

    推荐理由：ARC Prize 官方实测数据同时给出分数和每任务成本，读者可以据此比较 DeepSeek V4.1 Flash 与上一代的性价比取舍。

  4. The Decoder：AI News62

    Reflection 发布开放权重模型 Beam，以更低算力对标 DeepSeek 与 Qwen

    Reflection 宣布其首个免费开放权重模型 Beam，定位是以最小算力实现强性能，对标 DeepSeek 和 Qwen 的开放权重模型。Beam 为 501B 总参数的 MoE 模型，每 token 激活 23B；据 Reflection 称在推理任务上以少三到四倍的算力匹配 GLM 5.2，编码和智能体基准接近 Qwen3.8-Max，但原始性能仍不及 Kimi K3。