Skip to content

#推理

3 today

Oct 24Saturday

Hacker News front page

Show HN: Searching for optimal strategies in a puzzle game with chance (Drop7)

一项针对 Drop7 的自主策略研究显示,基于 n-tuple 查表的评估器在深度 3 搜索下、512 局未见过的公开开发局中平均得分 487,066,比同一搜索内手工评估函数(326,717)高出约 16 万分。该结果于 2026-09-05 首次在 screen 层级击败冻结的公平评估叶,次日以 SHA-256 校验的同一查表复现,但仍远低于百万分目标,尚未进入资格赛。

Google Blog · Gemini

See what 5 builders are making with Gemini 4 Argon

Google 展示了社区开发者使用 Gemini 4 Argon 构建的 5 个项目,涵盖三维模拟、实时数据可视化和多人浏览器游戏。案例包括水上飞机模拟、结合 NASA 野火数据与美国地质调查局地震数据的交互地球仪、喷气发动机模拟、由单张图像生成的交互式马车,以及多人夺旗游戏 Backyard Rascals。

AI HOT · Models

Moonshot AI releases Kimi K3.1 Pro with 1M token context, full weights due by November 4

Moonshot AI released Kimi K3.1 Pro, which supports a 1M token context, and says the full model weights will be released by November 4. A companion model, Kimi K3.1 Pro Highspeed, runs at 2-3x the speed. Both are available from today on Kimi.com, Kimi Work, Kimi Code and the Kimi API.

Why it matters: Moonshot disclosed vulnerability-task results before and after removing cybersecurity training data, giving a concrete comparison for risk control in open weights.

Hacker News front page

Kimi ships K3.1 Pro and a Highspeed variant on Kimi Code

Kimi made K3.1 Pro generally available on Kimi Code and released K3.1 Pro Highspeed, which has the same capabilities and outputs about 2–3× faster. K3.1 Pro has 3T parameters and a 1M context window, accepts image and video input, and runs only in thinking mode. Kimi says coding, 3D creation and financial research and modeling are notably improved.

Why it matters: The Highspeed variant trades about 2× the quota for roughly 2–3× faster output, while tool calls and script execution do not speed up.

Hacker News front page

Kimi launches K3.1 Pro and K3.1 Pro Highspeed with 1M-token context

Kimi released K3.1 Pro, a 3T-class model with a 1-million-token context window, and K3.1 Pro Highspeed, which runs 2–3× faster. After Kimi added cybersecurity risk controls to training data, K3.1 Pro's pass rate on an internal exploit test fell from 90.5% to 81.0%.

Why it matters: The drop from 90.5% to 81.0% on an internal exploit test shows what safety filtering in training data costs in offensive-security capability.

Latent Space

AI Is Getting Good at Math. Here’s Why.

AI 数学能力的进步,被 Ankit Maloo 归因于模型能快速推进候选思路,并结合并行智能体、验证器和失败反馈缩小搜索空间。他用洞见数量、洞见之间的工作量及搜索分支数解释问题难度,认为模型更适合完成概念突破较少、后续计算复杂的任务。文章强调人类判断与模型探索的互补,也主张公开失败路径,以帮助数学研究积累理解,而不只公布最终答案。

Oct 23Friday

Hacker News front page

The Rise and Fall of Jev

通用分类器 Jev 发布 33 天内经历爆红与迅速降温,其宣称"无法幻觉"实为输出符合 schema,并不代表逻辑正确,实际错误率甚至高于常规语言模型。作者指出 Jev 并非全新思路,类似模型 Laya 早其 1.5 年提出,且已有 Laya、Kev、Nimble、NanoJev 等性能更优的开源权重模型。

OpenAI News

OpenAI opens Decisions API for faster constrained-answer apps

OpenAI has released the Decisions API, built on GPT-6 Luna. For decisions with a fixed set of answers, it runs up to 10x faster than using GPT-6 Luna through the Responses API.

Why it matters: OpenAI moves constrained classification and routing decisions into a dedicated API, cutting the latency and cost that pile up across high-frequency small decisions in app workflows.

TechCrunch · AI

Microsoft’s Mustafa Suleyman says its AI outperforms doctors on hard cases as he chases ‘medical superintelligence’

Microsoft 在 Science 发表的研究显示,MAI-DxO 在复杂病例模拟中的诊断准确率为 83%,参与比较的全科及普通内科医生为 36%。该系统协调多个 AI 智能体问诊、安排检查并讨论诊断,但对照组未包括通常处理此类病例的专科医生或医生团队,其对基础模型的增益在急诊病例中也未显现。

TechCrunch · AI

OpenAI opens Decisions API to all customers, adding voice and vision

OpenAI opened its Decisions API to all customers on Thursday and added voice and vision support for the public version. The service is a reconfigured ChatGPT Luna that handles decision tasks with faster inference and lower cost. OpenAI says response time drops from one to three seconds to 100, 120 and 180 milliseconds. Language tutoring product Speak cut its action interval from 1.4 seconds to .13 seconds using the API.

Why it matters: OpenAI turned Luna into a low-latency decision service, and Speak cut its action interval and interaction wait after swapping out its existing LLM.

Oct 22Thursday

Hacker News front page

Achieving AI-style reasoning without an LLM

Minotaur 是一个不依赖语言模型和 GPU 的 AI 架构,基于 ELC 通过测量而非预测来推导答案,在单个普通 CPU 核心上每题约 0.25 秒完成演绎、归纳、溯因、诊断、信念修正等推理。在未见其代码的人设计的盲测中,它在提问会改变证人记忆的警察审讯场景,以及仅凭数据学习、诊断和预测系统等任务上,表现与当前用文字推理的语言模型相当。

Hacker News front page

WorldTUI: I trained a 360M model to pretend to be a terminal

WorldTUI 基于 Qwen2.5-Coder-0.5B 微调,用 360M 参数模型根据按键预测终端屏幕变化,背后没有真实 shell 或文件系统。训练使用 336k 段真实终端会话,并通过错误修复训练改善一致性,通过并行生成屏幕行和顺序验证加速推理。

Computing Life · Share · Yage

Why multi-agent AI costs more: what the extra budget buys

Multi-agent setups spend more tokens to gain specific engineering benefits, and whether a task gains depends on its parallelism, how much information it must cover, and its context state. Drawing on Anthropic, Cognition and several studies, the article sorts the gains into four kinds: shorter end-to-end time, gap-filling, more search diversity and context isolation. Team cost is usually several times that of a single agent.

Why it matters: It breaks multi-agent selection into task bottlenecks and payoff conditions, matching independent subtasks, information coverage and context pollution to different delegation and isolation choices.

Hacker News front page

LLM agents play Pokemon showdown

Oliver Grabner 构建了 Pokémon Showdown 2v2 多智能体环境,Qwen3.8-27B 队友自由交流时对脚本化最高伤害机器人的胜率为64%,禁言时为42%。

The Decoder

Deepmind founder Hassabis shortens AGI timeline from five to two to three years

Google DeepMind 创始人 Demis Hassabis 将 AGI 到来的预期从今年早些时候的3至5年缩短至2至3年。其定义是系统匹配人脑的完整认知范围,并提出“爱因斯坦测试”,即仅用1901年前的知识训练的系统能否重现爱因斯坦1905年的论文。他近期还提议设立由行业出资、政府控制的监督委员会,在模型发布前进行审查,并警告糟糕的监管比没有监管更糟。

Oct 21Wednesday

OpenAI News

OpenAI and partners use thousands of AI agents to mine biological data for new research leads

OpenAI, PerturbAI and Scripps Research ran thousands of AI agents driven by GPT-Rosalind to analyze experimental data from millions of mouse brain cells and propose research leads in 32 minutes.

Why it matters: With the model and resources held fixed, OpenAI found the distributed agents covered more literature, and blind expert reviewers preferred their reports in 75% of comparisons.

Latent Space

DeepSeek open-sources DeepInfer, discloses 155T tokens/day inference on a single cluster with caching

DeepSeek released DeepInfer, the inference engine behind V4.1 Flash, along with deepseek-recipe, the component that handles API requests. The engine supports distributed inference, speculative decoding and structured output. DeepSeek also published production figures showing 155T tokens per day on a single cluster with caching.

Why it matters: The production numbers and cache hit rates let teams estimate inference margins, since cache billing and hardware cost assumptions drive the math.

Hacker News front page

Alibaba hikes DeepSeek V4 Flash 0731 output prices by 230%

阿里将 DeepSeek V4 Flash 0731 的输出价格从每百万 tokens $0.40 上调至 $1.32,涨幅230%。Stork 自有价格目录记录显示,该模型输入价格也从每百万 tokens $0.20 升至 $0.44。

OpenAI News

Terminal X matches senior investor judgment 88.6% with GPT-6.1

Terminal X 的 Eval Agent 基于 GPT-6.1 Sol,在 500 条评分判断中与资深人类评审的一致率达 88.6%,比次优模型高 220 个基点。该 Agent 按 54 项标准评分,对偏离基金投资授权范围的回答检出率提升约 25%,单次查询成本降低约 30%。一家私募股权客户借助自动化的尽职调查与投资备忘录流程,实现年化 240 万美元节省。

Computing Life · Share · Yage

开放模型任务成本解析:为什么 Muse 比 DeepSeek 更省 token 却更贵?

开放模型的单任务成本同时受 token 消耗和单价影响,Muse Spark Open 输出量不足 DeepSeek 的一半,任务费用却高出三倍多。Artificial Analysis 的评测中,两者综合成绩均为 39 分,平均每任务费用分别为 0.84 美元和 0.27 美元,计算涵盖输入、缓存、推理思维链及最终答案的 token 费用。

Oct 20Tuesday

Hacker News front page

Maxwell: A new kind of model that sits between code and hardware

doubleAI 推出代码执行模型 Maxwell,根据程序运行时的硬件与数据优化代码,并对每项改进进行形式化等价证明。它在 CPU 和 GPU 上运行,收集足够运行时数据后,在程序持续运行期间替换为针对当前机器的优化代码。页面称数据处理流水线加速 38 ×、AI 推理内核加速 4.25 ×,LTX-2.5 在两块 B200 上生成每段 4K 视频的耗时从 250s 降至 134 s。

The Decoder

MiniMax M3.1 Flash scores 42 on Artificial Analysis index at under $0.22 per task

MiniMax released the reasoning model M3.1 Flash. Its Artificial Analysis Intelligence Index rose to 42 from 29 on the previous M3, while cost per task fell to under half, about $0.22.

Why it matters: M3.1 Flash costs less per task than DeepSeek and leads on hard Q&A and factual accuracy tests, though its overall intelligence index is slightly lower.

Hacker News front page

Inception releases Mercury 3, scaling its text diffusion model past 40x Mercury 1

Inception released Mercury 3, a diffusion-based large language model. The company says it beats GPT-6 Luna and Claude Haiku 5.5 on several benchmarks and runs 4–8x faster.

Why it matters: Mercury 3 scales past 40x the parameters of Mercury 1 while keeping similar speed, giving a like-for-like read on text diffusion at larger size.

Oct 19Monday

AI HOT · Models

MiniMax releases M3.1 Flash with nearly 3x faster inference and 70% lower usage cost

MiniMax released M3.1 Flash, which it says runs online inference nearly three times faster and cuts overall usage cost on agent tasks by 70% versus the previous generation. The model scores 73.7% on DeepSWE v1.1 and reaches 200+ TPS in everyday output. Pricing is $0.15 per 1M input tokens, $0.60 per 1M output tokens and $0.015 per 1M cached input tokens.

Why it matters: MiniMax claims a 70% cut in agent-task usage cost over the prior generation, with benchmark results on complex tasks, coding and computer use at leading-model levels.

Oct 18Sunday

Hacker News front page

Will it be a golden age or a dark age for mathematics (and science)?

Francesco Maggi 撰文指出,大语言模型能加速数学发现,但人类理解、科研独立性与人才培养将影响数学和科学走向黄金时代还是黑暗时代。文章区分 Lean 形式化证明的正确性与人类理解,强调数学进步还包括创造概念框架,并讨论署名、同行评审、研究保密与商业模型能力获取不均的问题。

Oct 17Saturday

DeepSeek · API updates

DeepSeek-V4.1-Flash-1017 update keeps API pricing and calling method unchanged

DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017, with the official web and app updated to match. API pricing stays the same and the model is still called through deepseek-flash.

Why it matters: DeepSeek improved the model while holding API price and calling method fixed, and cut task runtime and output length to lower task cost.