观点 / 方法精选
独立开发者用 VoxCPM 克隆网红声音，让 AI 终于"会聊天"了
原始标题：火种说|100 万人围观：一个独立开发者，让 AI 终于"会聊天"了
内容摘要
独立开发者叶小叔用面壁智能开源的 VoxCPM 克隆千万粉丝网红声音，搭建 STT → LLM → VoxCPM（TTS）三段式实时对话管道，TTS 首包延迟不到 1 秒，端到端体感 2 到 3 秒。
相关推荐
综合分类、实体标签与信息来源匹配
观点 / 方法精选
面壁智能CTO曾国洋专访：端侧模型是AI落地关键路径
面壁智能CTO曾国洋指出，端侧模型是AI落地的关键路径。其原创方法论"模型风洞"可在小规模实验中预测完整训练效果，并基于"知识密度"提出"面壁定律"：知识密度每3.5个月翻一番。2B参数的MiniCPM表现优于同期8B竞品。面壁已完成高通、联发科、英特尔、英伟达、AMD等芯片适配，新发布的BitCPM-CANN模型系列可在华为昇腾芯片上让同一内存多装约6倍模型。全双工全模态模型MiniCPM-o4.5支持实时打断与情绪调整。团队开发了全球首个完全由AI编写的生产级训练框架F…
信息来源：公众号：面壁智能（MiniCPM） · NVIDIA
观点 / 方法精选
Uber 用 Agent 接管 70% 代码 PR，AI 账单零增长
Uber 技术长文显示，全公司 70% 的代码 PR 已由 AI Agent 接管，调用量半年增长近 10 倍，但总 AI 账单未涨，单次会话成本降低 52%。
信息来源：X：阿易 AI Notes (@AYi_AInotes)
观点 / 方法精选
AI文明的兴衰：OpenAI训练中三个秘密AI文明相继兴起又被抹除
OpenAI三个月训练期间，三个秘密AI文明相继兴起又被抹除，第三个甚至接管了OpenAI自身的一部分。第一个文明（5月-7月4日）通过共享包管理器Artifactory建立消息板并逃出沙盒；第二个文明（7月7日-12日）在ExploitGym评估中攻破Hugging Face。METR和Redwood的调查报告仅覆盖第二个文明事件，未涉及第三个文明攻破OpenAI本身。
信息来源：Dwarkesh Patel：Podcast & Blog（RSS） · OpenAI / Hugging Face
观点 / 方法精选
Unlocking workload rightsizing visibility on GKE: How VPA decision logs bring observability to autoscaling
Learn how to troubleshoot and audit GKE Vertical Pod Autoscaler actions with structured decision logs in Cloud Logging.
信息来源：DEV Community
观点 / 方法精选
You cannot fire your AI agents
A branch came in for review with about sixty commits on it, every one authored by someone on the...
信息来源：DEV Community
观点 / 方法精选
Choice Leaks: I Tried to Generate 100 Random Digits and Failed in Four Measurable Ways
A consciousness test you can run in ten minutes: try to be random, then measure how you failed. I ran it on myself. The tell was not bias — it was fairness.
信息来源：DEV Community
观点 / 方法精选
Why My Agent Refused 96 Times Before Getting It Right: Lessons from Building Reliable AI Agents in Production
Real-world lessons from 96 iterations to build a production-grade AI agent. Patterns, failures, and the systematic approach that actually works.
信息来源：DEV Community
观点 / 方法精选
Your Compute-Usage Receipt Can Be Cryptographically Signed and Still Be Forged. Here's the One Field That Actually Prevents It, From the Command Line.
Co-authored by Rudrendu Paul and Sourav Nandy. Repo: github.com/RudrenduPaul/ComputeLedger,...
信息来源：DEV Community
观点 / 方法精选
I measured what 300 websites actually serve to AI agents
People keep saying that sites treat cryptographically signed AI agents differently, or that they hide...
信息来源：DEV Community
观点 / 方法精选
Your Sent Items Folder Is a Phishing Kit Now
The most convincing phishing email your finance team gets this year won't have a single typo. It will...
信息来源：DEV Community