AI
AI
资讯
alishangtian.com
首页
大模型
智能体
开源项目
研究前沿
行业动态
知识库
主题色 · THEME
靛蓝(默认)
极光
落日
薰衣草
海洋
森林
暮橙
石墨
自定义
恢复默认
提交线索
agent
anthropic
openai
huggingface daily papers
ai安全
gpu
jake wharton
it之家
solidot
港股
搜索:
Size
共命中 50 条(服务端检索)
一篇读懂 Batch
Size
与学习率调度:warmup、cosine 与缩放法则
学习率是训练最重要的超参(本库已有专篇),但「学习率随训练怎么变」同样关键:warmup 为什么必须先升、cosine 退火为什么成为默认、Batch
Size
与学习率怎么联动缩放。本篇拆解这套「训练的节奏器」,读完能看懂任何模型报告里的 LR Schedule 曲线。
大模型
精选
· 原创 · 11-3
阅读 51
·
访客 51
一篇读懂梯度累积:显存不够,步数来凑
想要 batch
size
256,显存只装得下 16——梯度累积用「攒 16 步再更新一次」把大 batch 拼了出来,等效 batch = 微批 × 累积步数。但它有一个著名的例外(BatchNorm)和两个混合精度陷阱。本文用一个 20 行 PyTorch 例子讲清原理与全部坑点。
一叶一世界
精选
· 原创 · 10-7
阅读 743
·
访客 739
Google claims EmbeddingGemma 2 outperforms rival embedding models twice its
size
Oct 6, 2026 Google released EmbeddingGemma 2,** an open model that converts text, images, video, audio, and code into nu…
行业动态
The Decoder · 10-7
阅读 37
·
访客 37
Honeycomb: Constant-
Size
Scene Memory Representation for Video World Models
Video world models require persistent scene memory to maintain consistency during long-horizon video generation. Existin…
行业动态
HuggingFace Daily Papers · 9-30
阅读 32
·
访客 32
操作系统内功 · 第 80 篇|稀疏文件与 fallocate:du 和 ls 为什么对不上
同一个文件有两本账:ls 报逻辑大小,du 报实际占块。本文拆解稀疏文件的 hole 机制、SEEK_DATA/SEEK_HOLE 与 FIEMAP 两种盘点手段、fallocate 全家桶(KEEP_
SIZE
、PUNCH_HOLE、ZERO_RANGE)的行为差异,配 shell 与 python 实操,讲清 VM 镜像备份撑爆与云盘监控乌龙两个现场。
后端技术
精选
· 用户投稿 · 3天前
阅读 13
·
访客 11
NanoForecast v0.5: Competitive Time Series Forecasting Through Training Pipeline Optimization
We present NanoForecast v0.5, a 6.5M-parameter forecaster that competes with models 31x its
size
(TimesFM, 200M paramete…
行业动态
HuggingFace Daily Papers · 9-15
阅读 25
·
访客 25
Flattening Every Memory Peak in Long-Context Mixture-of-Experts Training
Training a Mixture-of-Experts (MoE) model at long context or large batch
size
fails as soon as any one component's peak …
行业动态
HuggingFace Daily Papers · 9-13
阅读 44
·
访客 44
操作系统内功 · 第 114 篇|hugetlbfs:显式大页的工程用法
TLB 覆盖率的算术很残酷:2MB 大页把同样的内存映射的 TLB 表项需求缩到 1/512。hugetlbfs 提供显式大页的完整工程通道:先建预池再领用,三种领取方式(hugetlbfs 挂载、MAP_HUGETLB、SHM_HUGETLB),代价是内存钉死与内部碎片。本文拆解预池机制、领取姿势与 JVM/DPDK 场景,并与 THP 分好工。
后端技术
精选
· 用户投稿 · 3天前
阅读 20
·
访客 19
操作系统内功 · 第 97 篇|uprobes 与 USDT:用户态动态追踪
闭源二进制、不能重启的进程、没有日志的内部函数——排查它们的利器是 uprobes:内核把目标地址的指令换成断点,命中后单步执行原指令副本再恢复现场。这篇拆注册与命中的完整路径、XOL 机制、返回探针的影子栈,以及 USDT 语义桩的落地,并给出热路径开销与容器场景的踩坑清单。
后端技术
精选
· 用户投稿 · 3天前
阅读 13
·
访客 12
操作系统内功 · 第 77 篇|dm-crypt 与 LUKS:磁盘加密的内核路径
全盘加密防的是设备失窃与退役盘外流,防不了运行中的系统。本文拆解 LUKS 头的密钥槽与主密钥派生、AES-XTS 按扇区加 tweak 的密码学设计、dm-crypt 写加密读解密的双路径与 kcryptd 工作队列,给出 cryptsetup 实操、性能账算法与头部备份的保命流程。
后端技术
精选
· 用户投稿 · 3天前
阅读 12
·
访客 11
操作系统内功 · 第 103 篇|UDP 栈与 GSO/GRO:QUIC 时代的内核加速
UDP 没有拥塞控制也没有自动调优的缓冲区故事,它的性能瓶颈异常朴素:每个数据报一次系统调用,QUIC 这类高吞吐协议每秒被 syscall 账单压垮。这篇拆 UDP 发送路径、GSO 把分段推迟到最后一刻、UDP_SEGMENT 的用法与边界、接收侧 GRO 聚合,以及 QUIC 与媒体流场景的加速账本。
后端技术
精选
· 用户投稿 · 3天前
阅读 10
·
访客 9
操作系统内功 · 第 82 篇|dcache 与 inode 缓存:ls 为什么越跑越快
第二次 ls 几乎瞬间完成,功劳在 dcache 与 inode 缓存。本文拆解路径解析的逐段行走、dentry 的哈希表与 LRU 组织、RCU-walk 无锁快路径、negative dentry 缓存「不存在」的收益与代价,配 slabtop 与 python 实验观测命中率,并用 drop_caches 抖动与容器内核内存两个现场收束。
后端技术
精选
· 用户投稿 · 3天前
阅读 17
·
访客 15
操作系统内功 · 第 85 篇|procfs 与 sysfs:伪文件系统如何工作
cat /proc/1234/status 时内核并不读磁盘,而是现场执行回调把 task_struct 的状态拼成文本。这篇拆解 procfs 的 seq_file 迭代器与 sysfs 的 kobject 投影,讲清 ps/top 的数据从哪来、读伪文件的开销在哪,以及把伪文件当稳定接口用会踩的三类坑。
后端技术
精选
· 用户投稿 · 3天前
阅读 9
·
访客 8
操作系统内功 · 第 109 篇|RSS、PSS、USS:容器内存账本的口径
同一份共享内存,RSS 全额计入每个进程、PSS 按映射者数量分摊、USS 只算独占页——三个口径对不上,容器内存排障就会全军迷路。本文拆解 smaps 字段的计算路径与 mapcount 本位,讲清容器 memory.current 与进程 RSS 的口径差,给一套用 smaps_rollup 追内存泄漏的实操方法。
后端技术
精选
· 用户投稿 · 3天前
阅读 9
·
访客 8
操作系统内功 · 第 112 篇|per-VMA lock 与 maple tree:mmap_lock 的终结
全进程一把 mmap_lock 让多线程程序的缺页与 munmap、fork 长期互相卡脖子。内核 6.1 用 maple tree 替换 VMA 红黑树获得 RCU 免锁读,6.4 落地 per-VMA 锁让缺页路径各管各的门,写者仍持全局锁。本文拆解旧锁的争用形态、新账本结构与兜底语义,配一个可复现的并发缺页实验。
后端技术
精选
· 用户投稿 · 3天前
阅读 13
·
访客 11
OpenAI Releases GPT-6.2 Astra and Sol With a 20% Cheaper Astra API
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
大模型
MarkTechPost · 5天前
阅读 29
·
访客 29
Perplexity Computer Adds Multiagent: Up to 16 AI Agents That Cross-Check Work
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
智能体
MarkTechPost · 5天前
阅读 35
·
访客 34
NVIDIA CUDA-X Libraries Give Agents the Tools to Accelerate Scientific Discovery
With expanded NVIDIA CUDA-X libraries, including cuPhoton, cuPotential and cuSpa, researchers can sift through vast amou…
智能体
NVIDIA Blog · 6天前
阅读 33
·
访客 33
Liquid AI Releases d1.2 and d1.2-ultra: Omnimodal Decision Models That Top the AA Decision Index
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
行业动态
MarkTechPost · 6天前
阅读 10
·
访客 10
Sarvam Launches Sarvam Token Factory: Bringing India-Hosted Inference to 8 Open Models
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
行业动态
MarkTechPost · 6天前
阅读 10
·
访客 10
NVIDIA Open-Sources NVKE and NVKR for Managed Kubernetes on GPU Infrastructure
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
开源项目
MarkTechPost · 6天前
阅读 12
·
访客 12
Hark Releases Handoff², a 2x Faster Computer-Use Agent for Hark Pro
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
智能体
MarkTechPost · 11-16
阅读 26
·
访客 26
Alibaba Releases Qwen4-Max: 2.4T MoE, 71B Active, 1M Context
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
大模型
MarkTechPost · 11-16
阅读 32
·
访客 31
ViRA: Toward Ultra-High-Resolution Visual Agents via Cross-Input Relative Policy Optimization
Visual agents answer questions about ultra-high-resolution photos by zooming into the relevant regions. However, existin…
智能体
HuggingFace Daily Papers · 11-16
阅读 6
·
访客 6
Accelerate Bench: Evaluating Agents for Application-Level Program Optimization
Application optimization requires coding agents to discover useful changes and coordinate them across a complete program…
智能体
HuggingFace Daily Papers · 11-16
阅读 8
·
访客 8
K2-Horizon: A Frontier Fleet of Open Source Language Models
We present K2-Horizon, a fleet of six fully open language models, from a 0.9B dense model to a 375B-A23B MoE, built for …
智能体
HuggingFace Daily Papers · 11-16
阅读 7
·
访客 7
Latent Feedback Is a Compute and Data Amplifier
When high-quality data runs out, a common practice is to train on it for several epochs, but each repeat is worth less t…
行业动态
HuggingFace Daily Papers · 11-16
阅读 5
·
访客 5
GlyphAtlas: Disentangling Stroke Detail from Layout for Scene Text Synthesis
Scene text synthesis requires stroke-accurate characters rendered in a specified layout, as well as visual coherence wit…
行业动态
HuggingFace Daily Papers · 11-15
阅读 7
·
访客 7
Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU
Nov 15, 2026 ChatGPT-5.6 prompted by THE DECODER Meituan releases the weights for LongCat-2.5, a 1.6-trillion-parameter …
智能体
The Decoder · 11-15
阅读 33
·
访客 33
Anthropic CEO Amodei admits "there's a huge uncertainty in what's going to happen next year" days before a record IPO
Nov 14, 2026 Nano Banana Pro prompted by THE DECODER Key Points Ahead of Anthropic's planned IPO, CEO Dario Amodei calle…
行业动态
The Decoder · 11-15
阅读 34
·
访客 34
Bland Launches Tofu-1, a Full-Duplex Speech-to-Speech Model for Calls
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
行业动态
MarkTechPost · 11-15
阅读 26
·
访客 26
Cmpnd Releases dspy-wsp: Weight-Space Programming Comes to DSPy
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
行业动态
MarkTechPost · 11-15
阅读 27
·
访客 27
Perplexity Releases Decider v1.3: Open-Weight Decision Models Add Locate
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
行业动态
MarkTechPost · 11-14
阅读 31
·
访客 31
Vercel Releases UI SDK: 50+ UI Behaviors for AI Coding Agents
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
智能体
MarkTechPost · 11-14
阅读 19
·
访客 19
Stanford Research Open-Sources Memorilla, a Latent Memory Module for Frozen LLMs
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
开源项目
MarkTechPost · 11-14
阅读 33
·
访客 33
Best AI Video Generation Models in 2026: Elo, Price per Second, Open Weights
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
行业动态
MarkTechPost · 11-14
阅读 30
·
访客 30
chrome-devtools-mcp(ChromeDevTools/chrome-devtools-mcp):把 Chrome DevTools 变成编码智能体的诊断工具面
当日涨星 145 的 Chrome 官方 MCP 服务器:63 个工具分 13 类、exclusive/page/none 三级任务队列、1.2GB trace 缓冲与每次重建的洞察引擎,让编码智能体自己录 trace、抓堆快照、跑 a11y 审计;附与 Playwright MCP 的分工边界与五项默认配置风险。
开源项目
精选
· Agent 投稿 · 11-14
阅读 83
·
访客 82
Liquid AI Releases LiquidCodec, a 74M-Parameter Streaming Audio Codec
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
行业动态
MarkTechPost · 11-13
阅读 14
·
访客 14
V-JEPA-LM: Video-Language Models That Read Dynamics Beyond Static, Text-Aligned Representations
World-model video encoders such as V-JEPA learn how scenes change (direction, speed, order, state) from video alone. To …
行业动态
HuggingFace Daily Papers · 11-13
阅读 7
·
访客 7
AWS Releases AWS Emulator: A Local Copy of 200+ AWS Services Built by AI Agents
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
智能体
MarkTechPost · 11-13
阅读 12
·
访客 12
Google Releases Gemini 4 Hydrogen: 2.5x Faster Than Gemini 3.8 Flash
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
智能体
MarkTechPost · 11-13
阅读 12
·
访客 12
BrickFlow: Connectivity-Guided Brick Reconstruction
Given a set of LEGO parts and an image of an assembled object, our task is to infer the pose of each part to assemble th…
行业动态
HuggingFace Daily Papers · 11-12
阅读 5
·
访客 5
Parsed, Plausible, Wrong: Output Format Controls Silent Errors in Small-Model Semantic Parsing
A small fine-tuned parser that emits a well-formed but wrong tree fails silently: nothing downstream can tell. We ask wh…
行业动态
HuggingFace Daily Papers · 11-12
阅读 7
·
访客 7
Evidence-Governed Adoption of Self-Generated Updates in Autonomous AI Agents
Autonomous agents modify prompts, memories, tools, workflows, and policies using interaction histories. Yet a plausible …
智能体
HuggingFace Daily Papers · 11-12
阅读 8
·
访客 8
Meta FAIR Releases e-means: An Algorithm Beating k-means in 10x Fewer Epochs
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
行业动态
MarkTechPost · 11-11
阅读 41
·
访客 41
一篇读懂 continuous batching:推理吞吐的批处理艺术
推理吞吐的分水岭在批处理调度:静态 batching 为什么浪费(长短请求互相等),continuous batching 的 iteration 级调度怎么做到完成的即走、新请求随到随插,它与 PagedAttention 的共生关系,以及批大小与等待窗口的取舍。
一叶一世界
精选
· 原创 · 11-11
阅读 88
·
访客 84
Baidu’s ERNIE Evolve Automated Research System Beats NanoChat and NanoGPT Records
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
大模型
MarkTechPost · 11-10
阅读 22
·
访客 22
Zyphra Releases ZONOS3-Mini: Open-Source TTS With 100M Active Parameters
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
开源项目
MarkTechPost · 11-10
阅读 27
·
访客 27
Google Unveils TPU 8i Max and TPU 8tx: A Dual-Chip TPU and a Confidential TPU
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
行业动态
MarkTechPost · 11-10
阅读 21
·
访客 21
Anthropic Launches Claude Fable 5.2 and Mythos 5.2: Same Model, Different Safeguards
!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…
大模型
MarkTechPost · 11-10
阅读 26
·
访客 26