用不到百行纯标准库 Python 手撕 BPE 分词器：训练学合并表与编码贪心重放两个阶段拆开讲，小语料实测编码—解码往返一致，再用反例展示预分词正则为什么必不可少，最后对照 GPT-2 与 cl100k 的字节级 BPE、special token 与数字切分。

模型读到的不是字也不是词，而是 token。本文用 low/lower 语料演示 BPE 合并出词表的过程，解释中文为什么更费 token，以及它对计费、上下文长度和算术能力的影响，文末附代码示例与流程图。

Vector Quantization (VQ) is fundamental to discrete visual tokenizers that power modern autoregressive and masked image …

Tokenizer 的工作机制：BPE 分词、上下文窗口与计费逻辑

把 LLM 输出接进程序，坏 JSON 是头号工程痛点。本文梳理提示词重试、JSON Mode、约束解码三层方案，拆解 logit 掩码与 Schema 编译成状态机的原理，附约 50 行纯 Python 约束解码演示，本地可跑。

!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…

Agent 工程系统学习第 3 章：上下文工程取代 prompt engineering 成为核心技能。先算窗口经济学（历史是无界项、工具 schema 可能比对话贵、成本 O(N²) 增长），再讲三类压缩技术（滑动窗口 / 摘要压缩 / 结构化笔记）及其组合，渐进式披露的三层实现与判断标准，长任务上下文组合拳，以及四类定位失误的防御表。

Agent 工程系统学习第 1 章：从 HTTP 协议层讲透 LLM API——Chat Completions 协议与 role 语义、Function Calling 的"模型选择/代码执行"分工与三大常见错误、SSE 流式手写解析器（含 tool_calls 分块拼接）、token 计量与前缀缓存工程、重试/超时/幂等的错误分类纪律、多模态输入成本。附零框架多轮工具 Agent 实现作业。

!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…

Predicting the future evolution of a scene is a fundamental capability for world modeling. Recent work has shown that op…

This week, OpenAI released GPT-6.1 Sol. It upgrades GPT-6 Sol, the mid-tier model in the GPT-6 family. OpenAI’s claim is…

Keyword-matching benchmarks can credit small models for tool use they never perform. We document such a false positive i…

Recent video-based world models pair the scalability of autoregressive (AR) prediction with the visual quality of diffus…

Palash Debnath 的全本地开源语音工作台 VoiceStudio（当日涨星 +3,274、★43,728、AGPL-3.0）：把 17 个 TTS 与 7 个 ASR 引擎抽象成可插拔引擎层，覆盖克隆/设计/配音/听写/有声书并内置 MCP。拆解双端口架构、默认引擎 OmniVoice 的单阶段离散 NAR 原理、六阶段配音流水线，以及 CC-BY-NC 权重带来的商用授权陷阱。

Vectorize 开源的智能体记忆系统 Hindsight（当日涨星 +4,463、★37,121、MIT）：以世界事实/经验/观察/心智模型四网络替代扁平 RAG，LongMemEval 从同骨干全上下文的 39.0% 拉到 83.6%、最强 91.4%。拆解 TEMPR/CARA 分层、四路检索+RRF+重排、反思持久化机制，附五个落地场景与竞品争议。

Supersonic Labs, a small AI lab from Brazil, has released Julia 1. It is a compact decision model, not a chatbot. You pa…

田, 晏林* 2026-09-26 17:01:00 来源：量子位 GitHub现在最火热的大模型开源小蜂鸟Colibrì是个啥？ 闻乐 发自 凹非寺 量子位 | 公众号 QbitAI 25GB笔记本硬跑744B GLM-5.2，32GB…

Kyutai has released **Voice of Reason**, 2 open-weight speech-to-speech models that solve math problems out loud. Both s…

GGUF, GPTQ, AWQ, EXL2, and EXL3 solve the same problem in different ways. This guide separates file containers from quan…

Standard subword tokenizers either treat every orthographic variation of a word (such as hello, Hello, HELLO, and Héllo)…

We study length inflation in on-policy distillation (OPD), where student responses can become excessively long and even …

Visual signals require compact yet sufficient representations for robust downstream prediction. Convolutional sparse cod…

]

A core goal of efficient reasoning is to improve the accuracy-efficiency frontier. However, jointly improving reasoning …

Specialist distillation effectively transfers domain expertise to student models via teacher-generated reasoning traject…

We introduce StepAudio 3 Music, a large-scale, long-form music generation model that supports explicit musical planning …

We introduce StepAudio 3 Gen, a general-purpose audio generation model that supports zero-shot text-to-speech (TTS), voi…

从零微调 GPT-2 的完整流程：数据准备、训练与效果评估