Skip to content

#推理

3 today

Today · Nov 17Tuesday · 3 items

The Decoder

OpenAI's GPT-6 Astra tackles the Standard Model's toughest math problem on a lattice

OpenAI 的 GPT-6 Astra 协助完成一项证明,可能为粒子物理中延续 50 年的标准模型格点化难题提供解法。物理学家 Nathaniel Craig 提出通过比较逐步细化的格点修正重叠费米子相位,模型将其转化为可行构造并主导详细证明,成果已作为预印本发表。三位格点场论专家审阅并提出问题;Craig 表示,这项工作更像起点,距离实现现实物理模拟仍有大量工作。

The Decoder

Alibaba's Qwen4-Max takes on Opus 5.5, GPT-6 and Kimi K3.1 Pro at a lower price

Alibaba released Qwen4-Max, its flagship model, built on a hybrid attention architecture with native image handling and a one-million-token context window, aimed at industrial, scientific and financial tasks. Artificial Analysis measured its intelligence index at 43, nine points above Qwen3.8-Max, with cost per task falling from $10.02 to $3.27. Some evaluations still trail rival models.

Why it matters: Qwen4-Max gains nine points on the Artificial Analysis index over its predecessor and cuts cost per task to about a third.

Hacker News front page

Nvidia faced chip competition and China limits before the Huang controversy

Nvidia was already under pressure from chip rivals, China market restrictions and challenges to its CUDA ecosystem before the Huang episode. The article says Nvidia holds 80 percent of AI accelerator sales, while AMD's share rose from 2 percent in 2022 to 7–10 percent in 2026. Google, Amazon and Microsoft are also building their own chips.

Why it matters: Specialized inference chips and cloud in-house silicon compete with Nvidia, and AI-assisted code migration lowers the cost of switching hardware, weakening its software lock-in.

Yesterday · Nov 16Monday

Hacker News front page

Qwen releases Qwen4-Max with native vision and a one-million-token context

Qwen released Qwen4-Max, built on the Qwen4 architecture with native vision and a one-million-token context window. It is available now on QwenCloud through Token Plan and API access.

Why it matters: Qwen4-Max raised task performance on the official Work360 benchmark and cut average cost per task from $4.43 to $1.67, a 62% drop.

OpenAI News

A step toward defining the Standard Model on a lattice

Nathaniel Craig 在 GPT‑6 Astra 协助下提出标准模型的候选非微扰格点定义,并在 10 月预印本中给出构造。方法通过比较逐步细化格点上的费米子态确定相位规则,Astra 协助设计细化方案、论证收敛并主导详细证明的开发,外部专家参与审阅与提问。该构造适用于场强受限的有限格点间距,距离实际模拟仍有大量工作,有限间距下的反射正性及连续时空定义问题仍未解决。

Nov 15Sunday

Hacker News front page

Will we put data centers in space?

Forethought 研究估计,轨道 AI 数据中心在 2030 年前不太可能占据显著算力份额,但若 Starship 开发按计划推进,可能在 3–5 年内具备成本竞争力。模型显示,发射成本降至约 $100/kg 并适度降低设备质量后,轨道数据中心总成本可接近地面方案,散热预计并非主要瓶颈。无法维修的芯片故障会增加配套硬件成本,星间带宽限制则可能使早期轨道算力主要用于推理。

The Decoder

Anthropic says outdated prompting habits slow Claude Opus 5.5

Anthropic advises Claude Opus 5.5 users to drop phrases like "think hard" and start at the default Medium reasoning effort. In internal tests, removing such wording made responses start faster with no measurable quality drop, and Opus 5.5 at Medium matched Opus 5 at High. Other tips: state the goal, output format and when to ask follow-ups in one go, attach charts or files directly, and name unwanted design styles.

Why it matters: Anthropic's guidance boils down to stating a clear endpoint, cutting prompts that push the model to think harder, and raising reasoning effort only when needed.

Nov 14Saturday

Hacker News front page

Google shows Antigravity /teamwork results in scientific research and engineering optimization

Google and academic partners used Gemini 4 and Antigravity /teamwork to produce results in math, formal verification, molecular structure discovery and systems engineering. The setup pairs parallel exploration with independent checking and isolates context per task.

Why it matters: Google's /teamwork combines parallel exploration with independent verification and isolates context by task, a collaboration pattern that carries over to long-running research and engineering work.

Latent Space

[AINews] The No AI Slop Manifesto

《No AI Slop 宣言》在 AIE CODE 大会闭幕演讲中由 Dex 发布,已获多位演讲者联署并登上 Hacker News 首页。同期 llama.cpp 与 ggml 发布 1.0,移除公共头文件中 35 个废弃函数;SGLang 也在 SGLang Summit 上宣布 1.0。

Hacker News front page

Unsloth Qwen3.8-27B Quantizations – Dynamic v3.1 Update

Unsloth 发布采用 Dynamic V3.1 预览算法的 Qwen3.8-27B GGUF,称多数基准在相同体积下准确率提升 5-10%。此次更新覆盖多数 2bit 至 6bit 量化版本,3bit 因提升不显著而保留旧版;智能体基准使用 256K 上下文,其余基准使用 128K。

Hacker News front page

After 600 experiments, here's what I discovered about JEPA world models

Haddy Alchaer 对 LeWM 进行了 600 次实验,找到四种无需额外训练的规划方法,将长时程任务成功率从 22% 提升至 85%。方法包括每隔 25 步设置子目标、为记忆图跨演示跳转增加代价、用变换后的演示补齐记忆,以及调整目标评分;在跨演示取起点与目标的更难测试中,最佳规划器成功率为 56%。

Dwarkesh Patel podcast

The avalanche that hit math is coming for AI development

Dwarkesh Patel 预测 AI 将迅速逼近顶尖人类研究员水平,并以 2028 作为实现 AI 研发自动化的中位年份判断。他以数学能力进展及预训练、后训练、推理效率的可验证指标支撑这一预测,同时给出 25% 的概率认为该节点会推迟至 2032 之后。他对持续自我加速的软件奇点与受算力限制的短期加速各赋予 50% 概率,后者预计将把 3–6 年的 AI 进展压缩到 1 年内。

Hacker News front page

AI unlocks Gödel's notebooks, said to contain hundreds of proved theorems

赫尔辛基大学 Jan von Plato 团队借助 Claude 解读哥德尔的 16 本数学与逻辑笔记,称从中发现数百条此前未被知晓的定理。团队向 Claude 提供笔记转录、研究著作及相关文献,数天内获得针对 1,009 页材料的 650 页解释,并需逐页核查。

Nov 13Friday

TechCrunch · AI

Israeli VCs see Nvidia leadership crisis as opening for rivals

Jensen Huang's resignation has worried Israel's tech sector, where VCs say a leadership crisis at Nvidia could open room for AI chip and inference software competitors. Nvidia employs 6,000 people in Israel, and locals are watching whether a successor keeps the company's commitments there. Investors said slower decision-making could leave space for rivals, but if uncertainty slows AI spending the market shrinks too, and competitors still need better economics to win business.

Why it matters: Israeli investors say the Nvidia leadership crisis could open a window for chip and inference software rivals, while also shrinking the market if AI spending slows.

The Verge · AI

Google's new Gemini 4 Hydrogen model is better at breaking down complex questions

Google says Gemini 4 Hydrogen can break down and answer complex questions item by item, including multi-part questions, hard-to-read typos and highly specific requests. Google has made it the default model in AI Mode, so decomposed answers and multi-subtopic retrieval now feed straight into search.

Why it matters: Google made Gemini 4 Hydrogen the default in AI Mode, so its handling of complex, multi-part questions lands directly in the search experience.

The Decoder

Google releases Gemini 4 Hydrogen, scoring above Argon less than four weeks after its launch

Google released Gemini 4 Hydrogen for everyday tasks. At its default high thinking level it scores 51 on the Artificial Analysis Intelligence Index, above Gemini 4 Argon's 50, and reaches 54 at the xhigh level.

Why it matters: A new Google model matches rivals' intelligence on an independent benchmark at a lower price, with the gap showing up in agentic tasks and in cost at higher thinking levels.

Ars Technica · AI

Google launches Gemini 4 Hydrogen, replacing Gemini 3.8 Flash across its products

Google released Gemini 4 Hydrogen with immediate access and swapped it in for Gemini 3.8 Flash across its products. The model has four thinking levels: low, medium, high and a new xhigh. Google says high mode runs 2.5 times faster than Flash 3.8, cuts fabricated statements by more than 5 times, and scores over 3 times the old model on Terminal-Bench.

Why it matters: Google opened Hydrogen to everyone and replaced Flash outright, a lower bar than its earlier staged releases limited to pricier subscriptions.

Hacker News front page

Google launches Gemini 4 Hydrogen and Orchestrator in Antigravity

Google launched Gemini 4 Hydrogen in Google Antigravity today and added an Orchestrator option that switches between Hydrogen and Gemini 4 Carbon within a single task.

Why it matters: Orchestrator swaps planning and execution models mid-task, nearing Carbon's task success rate while using less quota and finishing faster.

TechCrunch · AI

Google Search adds Gemini 4 Hydrogen to handle complex queries

Google Search is rolling out Gemini 4 Hydrogen to all search users worldwide to improve how it understands and answers complex questions. Google says the model handles multi-part questions, specific requirements and hard-to-read typos better, and gives more useful links, images, structured answers, comparison tables and custom tools in AI Mode.

Why it matters: Google is putting Gemini 4 Hydrogen into Search so multi-part questions and specific requirements get sharper answers, with images and comparison tables added.

Google Blog · Gemini

Google releases Gemini 4 Hydrogen for everyday coding and agent workflows

Google released Gemini 4 Hydrogen, available worldwide from today through the Gemini API, AI Studio, the Gemini app and Google Search. It focuses on fast coding and agent execution.

Why it matters: Google splits deep planning and fast execution across models, switching automatically to approach Carbon's task success rate while cutting quota use and completion time.

Google Blog · Gemini

Google Search gets Gemini 4 Hydrogen upgrade, rolling out worldwide

Google Search now runs on Gemini 4 Hydrogen and is available to all users worldwide, with better question understanding, retrieval and visual answers. The upgrade improves instruction following and tool use, handling multi-part questions, hard-to-read typos and specific requests, and runs parallel searches on related subtopics to surface a wider set of links. Answers include more images, structured content, comparison tables and custom tools, and supported AI Mode languages get finer language understanding.

Why it matters: Google is upgrading search for every user, widening link coverage by searching related subtopics in parallel so people find sites they may have missed.

The Decoder

OpenAI’s data suggests that frontier models are hiding their thoughts because developers optimize for capability and compute

OpenAI 安全团队的新分析指出,预训练能力、架构选择、强化学习和教师模型数据都会影响模型隐藏推理及操控 CoT 的能力。研究覆盖 GPT-5 至 GPT-6 Astra,发现更强的基础模型能在不输出推理链时完成更多推理步骤,较小模型通过更长推理链可匹配大模型的能力且更易监控,但会增加部署计算成本。

Nov 12Thursday

Hacker News front page

Talking Zen 6 with Mike Clark at AMD In-Depth

AMD 原 Zen 架构师 Mike Clark 在访谈中详解 Zen 6 与 Zen 6c 的设计差异:Zen 6c 已从过去的密度版本升级为顶级性能核,每核配备 4MB L3,32 核共 128MB L3,并拥有该核心可达到的最高 IPC。

The Decoder

Google launches Gemini Decision 1, handling 1,000 decision questions per call

Google DeepMind released Gemini Decision 1 and the Gemini Decisions API, which can process up to 1,000 separate closed-ended questions in a single call.

Why it matters: Google turned closed-ended decisions into an input-billed endpoint, and batch processing with no output token charges lowers the cost of these calls.

Nov 11Wednesday

Hacker News front page

An Open Field Guide to Inference Engineering

Modelplane 的推理工程指南面向平台工程师,系统讲解模型与 GPU 的配合、生产部署及职业学习路径。指南涵盖权重与 KV cache 显存预算、prefill 与 decode 瓶颈、引擎优化、网关鉴权、可观测性和扩缩容,并强调用真实流量开展并发评测。

TechCrunch · AI

PSI is trying to make a business out of AI-driven scientific discovery

Physical Superintelligence(PSI)正以数据中心优化获取收入,并探索将 AI 驱动的理论物理突破转化为实际技术。其研究人员 David Turturean 使用提前获准访问的 Anthropic Fable 5.2,求解了提出已有18年的 Aaronson-Ambainis 猜想,结果经 Lean 自动形式化并交由相关研究者审阅。

The Decoder

Google opens Gemini 4 Carbon, a fast model aimed at complex coding and research

Google has opened Gemini 4 Carbon after two weeks of testing by safety experts. It sits alongside Gemini 4 Argon and targets complex multi-step coding, debugging and open-ended scientific research. Carbon carries the same per-token price as Argon, so deeper reasoning and longer verification loops raise usage.

Why it matters: Carbon and Argon share a token price, so deeper reasoning and longer verification loops can push the real total cost of complex tasks higher.

Google DeepMind

Google DeepMind releases Gemini 4 Carbon for complex coding, cybersecurity and scientific reasoning

Google DeepMind has released Gemini 4 Carbon to developers, enterprises, scientists and consumers. It targets complex knowledge work, long-horizon agent workflows, cybersecurity and scientific reasoning. Google prices Carbon at the same entry rate as Argon, but Carbon spends more tokens on deeper exploration and verification.

Why it matters: Google positions Carbon for complex multi-step tasks, charging the same entry price as Argon while spending more tokens to explore and verify.

Google Blog · Gemini

Google releases Gemini 4 Carbon for complex agent tasks and scientific reasoning

Google has released Gemini 4 Carbon to developers, enterprises, scientists and consumers, targeting complex knowledge work, agentic coding, cybersecurity and scientific reasoning. Carbon ranks first on CursorBench 4.0, WANDR and AutomationBench, and Google says it costs more than 2x less per task than the next-best external model.

Why it matters: Google sets Carbon at the same entry price as Argon, with Carbon spending more tokens on reasoning and verification for complex multi-step tasks.

Nov 10Tuesday

Google DeepMind

Gemini Robotics 2.5 boosts general dexterity in real-world robot tasks

Google DeepMind released Gemini Robotics 2.5, a vision-language-action model it says improves how fast robots learn skills, along with dexterity and safety. The model is being opened gradually to robotics partners.

Why it matters: DeepMind is testing Gemini Robotics 2.5 on sorting car parts, moving robot reliability checks onto a real production line.

Google DeepMind

Google DeepMind unveils AlphaProof Code, formally verifying Signal's core crypto code

Google DeepMind published AlphaProof Code, a multi-LLM agent framework that formally verified key security properties of Signal's core cryptographic code in four weeks. The work moves formal verification from the protocol design to the actual implementation.

Why it matters: DeepMind pushes Signal's formal verification from protocol design down to real code, covering the design-to-implementation gap where bugs tend to hide.

Latent Space

Anthropic opens IPO roadshow, Claude Fable 5.2 tops LLM leaderboard

Anthropic started its IPO roadshow on Monday and released Claude Fable 5.2 and Claude Mythos 5.2. Fable 5.2 ranks first on the Artificial Analysis Intelligence Index v5 with 65 points.

Why it matters: Fable 5.2 cuts the price per token by 20%, yet GDPx cost per task rose from $67.16 to $88.15.

Simon Willison

llm-anthropic 0.31 adds support for Fable 5.2 and Mythos 5.2

Simon Willison released llm-anthropic 0.31, adding Anthropic's Fable 5.2 and Mythos 5.2, both launched today, so the new models can be called from the CLI.

Why it matters: Willison tested Fable 5.2's five reasoning levels; at max it ran out of tokens without producing an image, while the two earlier versions succeeded at the same level.

TechCrunch · AI

Anthropic releases Mythos 5.2 and Fable 5.2, raising concerns about recursive self-improvement

Anthropic released Mythos 5.2 and Fable 5.2 on Monday, citing gains in coding, reasoning, vision and communication. Fable's input price drops from $10 to $8 per million tokens, responses are 30% faster at every effort level, and Enterprise Frontier Safeguards support zero data retention (ZDR).

Why it matters: Fable's input price falls from $10 to $8 per million tokens, making Fable 5.2 cheaper than main rival GPT-6 Astra.