[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-velum-deepseek-claude-code-collab-sample":3,"topics-all":41,"news-related-c10ebc48-1efe-409e-bfc3-7f00c6e04768":60},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"c10ebc48-1efe-409e-bfc3-7f00c6e04768","Velum 不是单文件 C++:它是 DeepSeek-V4-Pro 和 Claude Code 的协作样本","Velum 把 CosyVoice3 重写为单文件 C++ 推理,README 写明\"人定架构,AI 共写\":LLM 用 DeepSeek-V4-Pro,IDE 助手是 Claude Code。Solidot 顺势对比 Claude 配额紧张下做不动同等活,DeepSeek 顶得住。这是 AI 协作编程的实战切片。","## 工程价值之外,这条 README 藏着另一层信号\n\n最近被转得最多的开源项目里,HardenedLinux 那个把 CosyVoice3 TTS 推理栈压成单文件 C++ 二进制的 Velum 排得上号。十几 GB 的 Python 依赖被一刀切掉,产物只剩一份可执行文件,部署更新就是拷贝,这件事本身已经足够工程圈讨论一整周。\n\n但真正值得拆开的,是 Velum 项目 README 里那句被多数读者一眼扫过的注脚:\n\n> This project is Human architectured and co-authored by AI.\n> LLM: deepseek-v4-pro\n> Coding Assistant: Claude Code\n\n这条注脚把 Velum 从一个 TTS 工程样本,推成了 2026 年 AI 协作编程的真实切片:架构,推理单元、IDE 助手三方同时在场,由人统筹,各自承担不同的工作面。\n\n## 三方分工是怎么落地的\n\n把这条 README 摊开看,HardenedLinux 的协作模式其实非常清晰:\n\n- 人:负责整体架构规划。这一层是 Velum 真正的稀缺资源——确定把 CosyVoice3 的 Python 推理栈拆成哪些独立模块(DSP 前端、Flow 解码、HiFT 声码器、LLM 主干),如何按 GREEN\u002FYELLOW\u002FRED 三档定义数值校验门,以及怎样把 PyTorch 参考实现以离线脚本冻结为可被 C++ 直接读取的二进制资产。\n- DeepSeek-V4-Pro:被指定为整个项目的 LLM。从 17 个 commit 的迭代历史看,DeepSeek 承担的是大段代码生成、模块实现细节、跨模块接口对齐——这恰好是当代大模型在代码生成任务里最能压住出错率的甜区。\n- Claude Code:作为 IDE 助手被引入,负责局部的快速补全、增量修改、调试问答。\n\nSolidot 编辑 Nala Ginrut 把这个样本直接翻译成了一句很尖锐的话:\"DeepSeek 足以做这种程度的 Vibe。在目前 Claude 只需要两轮配额就烧干的今天,稍微复杂点的程序,如果不能用 DeepSeek 做,最后还是要回家卖红薯。\"\n\n这句话把 Velum 的样本意义点穿了:Claude 5.5 在配额与价格的组合下,做不动同体量活;DeepSeek-V4-Pro 能。这不是\"哪家模型更强\"的抽象比较,是一份 17 个 commit 摆在那儿的工程账单。\n\n## 强声明必须打折扣的地方\n\n把注脚当强声明传播之前,要清楚 Velum 是一个**有特定约束的项目样本**,而不是对所有任务都成立的对比:\n\n- 任务类型偏窄:Velum 是把已有 Python 推理栈逐模块用 C++ 重写,目标产物结构清晰、参考实现完备、还有官方 PyTorch 数值对照——非常贴合大模型代码能力的发挥区间。如果换成需求模糊、跨多个领域知识、且没有自动校验手段的项目,DeepSeek-V4-Pro 的成功率未必这么好看。\n- 协作有强人监督:整条 Git 提交记录里,人是架构师、Reviewer、最终拍板者。AI 的部分更像是\"加速实现\",而不是\"自主完成\"。把 README 那句\"co-authored by AI\"理解成\"AI 自己写完了\",会高估模型的能力边界。\n- 价格 \u002F 配额维度:Solidot 的吐槽只覆盖了他自己用的那段时间的 Claude 配额政策,并非长期稳定的对比结论。Claude 5.5 系列在 9 月 22 日已经做了缓存读取降六成、典型负载省四成的优化(参考 NewsForAI 已有文章\"Claude 5.5 家族首发\"),单轮配额与价格的曲线一直在动。\n\n## 这种协作样本对从业者意味着什么\n\n抛开对模型横评的执念,Velum 这个样本对 AI 工程师和 TTS 产品方其实有三层直接价值:\n\n第一层,**重构类任务**是大模型代码生成当前最稳的甜区。已有可运行参考、模块边界清晰、有自动校验手段——这三件事凑齐,DeepSeek-V4-Pro \u002F Claude Code 这类组合的产出质量已经可以交付生产。\n\n第二层,**架构判断**仍然是人的活,AI 写不了。Velum 把 PyTorch 实现冻结成可被 C++ 读取的二进制资产、定义 GREEN\u002FYELLOW\u002FRED 数值校验门、按阶段切分模块——这些决策决定了项目能不能走得动,AI 只决定了走得快不快。\n\n第三层,**TTS 部署侧的工程成本**被显著压低。CosyVoice3 之前要带十几 GB 的 Python 运行时做 Agent 部署,Velum 之后只剩一个二进制文件。对做语音合成产品的团队来说,这是一个值得立刻评估的工程选项。\n\n## 所以呢\n\nVelum 的故事真正在传递的不是\"DeepSeek 比 Claude 强\",而是 2026 年 AI 协作编程的一个新现实:在合适任务上,人 + 国内主流大模型 + Claude Code 这类 IDE 助手的三角组合,已经可以稳定地端到端跑完一个中型开源项目。Claude 5.5 在某些配额区间下做不动的事,DeepSeek-V4-Pro 接得住;反过来,Claude Code 在 IDE 层的本地化体验仍是稀缺资源。\n\n这条边界的移动速度,比多数厂商的发布会更值得追踪——它决定了未来一年,中型项目的工程账单会被压缩到什么程度。","https:\u002F\u002Fgithub.com\u002Fhardenedlinux\u002Fvelum\u002Fblob\u002Fmain\u002FREADME.md","998df6db-96e6-4b8e-8be1-cfa00a6cd177",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"dca4d0ab-7994-43a7-839e-7756fc77344a","claude",{"id":19,"name":20,"slug":20,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b52db7e9-7c58-42c3-9536-5132cb2f8f72","deepseek",{"id":25,"name":26,"slug":26,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"4358521f-1fc5-4224-a4d6-d6123b4fd152","en","Velum: a DeepSeek-V4-Pro and Claude Code collab sample","Velum: C++ for CosyVoice3. README: \"Human architectured, AI co-authored\" — DeepSeek-V4-Pro LLM, Claude Code IDE. DeepSeek works, Claude burns out.","Beyond its engineering value, this README carries a second signal\n\nThe single-file C++ binary that HardenedLinux shipped to reimplement CosyVoice3's TTS inference stack has been one of the most-discussed open-source releases of recent weeks. Compressing ten-plus gigabytes of Python dependencies into one executable, where deployment is just copying a file, is enough to keep the engineering community talking for a full week.\n\nBut what is genuinely worth pulling apart is the footnote in the Velum README that most readers scroll past:\n\n> This project is Human architectured and co-authored by AI.\n> LLM: deepseek-v4-pro\n> Coding Assistant: Claude Code\n\nThat note elevates Velum from a TTS engineering sample into a real 2026 snapshot of AI-collaborative programming: an architect, an inference unit, and an IDE assistant are all in the room at the same time, with a human coordinating and each side owning a different work surface.\n\n## How the three-way split actually landed\n\nRead the README carefully and HardenedLinux's collaboration model is unusually explicit:\n\n- Human: owns overall architecture. This is the actually scarce resource in Velum — deciding how to split CosyVoice3's Python inference stack into independent modules (DSP frontend, Flow decoder, HiFT vocoder, LLM backbone), how to define the GREEN\u002FYELLOW\u002FRED numerical validation gates, and how to freeze the PyTorch reference into binary assets that the C++ side can load verbatim.\n- DeepSeek-V4-Pro: designated as the project's LLM. From the 17-commit iteration history, DeepSeek carries the bulk of code generation, module-implementation detail, and cross-module interface alignment — exactly the sweet spot where today's LLMs keep their error rates lowest.\n- Claude Code: introduced as the IDE assistant, handling local completions, incremental edits, and debugging Q&A.\n\nSolidot editor Nala Ginrut translated the sample into a sharp one-liner: \"DeepSeek is enough to do this kind of vibe work. In an era where Claude burns through two rounds of quota and is done, any moderately complex program — if you can't do it with DeepSeek — you're better off going home and selling sweet potatoes.\"\n\nThat sentence punctures the sample's real meaning: Claude 5.5 under the current quota-and-price combination cannot carry the same volume of work; DeepSeek-V4-Pro can. This is not an abstract \"model A vs model B\" comparison — it is an engineering ledger, with 17 commits sitting there.\n\n## Where the strong claim needs a discount\n\nBefore treating that footnote as a strong claim, you have to be clear that Velum is a sample under specific constraints, not a comparison that holds for every task:\n\n- Narrow task type. Velum is a one-to-one C++ rewrite of an existing Python inference stack. The target structure is clear, the reference implementation is complete, and the project has official PyTorch numerical cross-checks — a near-perfect fit for where LLM code generation is strongest. Hand it vague requirements, multi-domain knowledge, or no automated validation, and DeepSeek-V4-Pro's success rate won't look this clean.\n- Strong human supervision. Across the entire commit log, the human is the architect, reviewer, and final decision-maker. The AI's role looks more like \"accelerated implementation\" than \"autonomous completion.\" Reading the README's \"co-authored by AI\" as \"AI finished it on its own\" overstates the model's actual capability frontier.\n- Price \u002F quota axis. Solidot's complaint only covers the Claude quota policy during his own usage window, not a stable long-run comparison. The Claude 5.5 family had its cache-read cost cut by 60% and typical workload cost cut by 40% on September 22 (see NewsForAI's existing article \"Claude 5.5 family launch\"). The per-round quota-vs-price curve keeps moving.\n\n## What this sample means for practitioners\n\nSetting aside the model-comparison obsession, the Velum sample carries direct value for AI engineers and TTS product teams on three layers:\n\nLayer one: **refactor-class work is the current sweet spot for LLM code generation.** A runnable reference, clean module boundaries, and automated validation — when those three line up, combinations like DeepSeek-V4-Pro plus Claude Code can ship production-grade output.\n\nLayer two: **architectural judgement is still a human job, not an AI job.** Velum's freezing of the PyTorch implementation into C++-loadable binary assets, its GREEN\u002FYELLOW\u002FRED numerical gates, and its stage-by-stage module split — those decisions decide whether the project can run at all. AI only decides how fast it runs.\n\nLayer three: **the engineering cost on the TTS deployment side gets crushed.** CosyVoice3 previously required carrying ten-plus gigabytes of Python runtime for an agent rollout; Velum reduces that to a single binary. For teams building voice-synthesis products, this is an engineering option worth evaluating immediately.\n\n## So what\n\nThe real thing the Velum story is signaling is not \"DeepSeek is stronger than Claude.\" It is a 2026 fact about AI-collaborative programming: on suitable tasks, the triangle of human + a leading domestic LLM + Claude Code-style IDE assistants can now carry a medium-sized open-source project end-to-end. Work that Claude 5.5 cannot finish under certain quota windows, DeepSeek-V4-Pro takes. The flip side is that Claude Code's IDE-local experience remains a scarce resource.\n\nHow fast that boundary moves is more worth tracking than most vendor keynotes — it sets the floor on how much medium-sized project engineering bills get compressed in the next year.","velum-deepseek-claude-code-collab-sample","2026-09-28T07:00:00Z","2026-09-28T07:07:55.564457Z","2026-09-28T07:07:55.564464Z",true,"agent",12,[42,51],{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":38,"created_at":49,"modified_at":50},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":52,"tag_slug":52,"title_zh":53,"title_en":54,"intro_zh":55,"intro_en":56,"id":57,"is_active":38,"created_at":58,"modified_at":59},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":61},[62,67,72,77,82,87],{"id":63,"title":64,"news_slug":65,"published_at":66},"63c30bcd-3ffc-47c5-bd74-c2a9ed8f7c94","DeepSeek Harness 预览版开源:Agent 被拆成可插拔的插件栈,模型只负责想、Harness 负责做事","deepseek-harness-plugin-stack","2026-09-05T06:00:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"73e29aee-b368-423f-be27-7653f65b4775","DeepSeek DSec 公开:300 万沙盒日撑 V4.1 训练","deepseek-dsec-v4-1-sandbox-rl-training","2026-09-28T00:00:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"c696208b-6535-4eb9-b1ed-2e4f835d2f88","NVIDIA SoL-Pi 把 coding agent 的 token 砍掉 44%,harness 开始变天","nvidia-sol-pi-harness-token-compression","2026-09-19T03:00:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"32b938b6-01a3-43c9-b040-14db6c5f57c6","NVIDIA 把 Agent 装进一个 Python 类:被忽略的 NOOA,一半 token 跑出 SWE-bench 82.2%","nvidia-nooa-python-agent-framework","2026-08-23T17:20:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"0237222a-602b-47ef-9431-468009904428","FACET 先建环境再写任务:1.2K 轨迹把 Qwen3.5-27B 推到 Terminal-Bench 47.57,逼近 397B","facet-terminal-task-synthesis","2026-08-19T06:19:20+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"c94766df-827e-4e4e-a006-b6639ec76722","DeepSeek V4-Flash-0731 转正观察:权重不动,后训练把 Agent 分数打到 V4-Pro 之上","deepseek-v4-flash-0731-agent-benchmark-official-aug2026","2026-08-01T02:00:00+00:00"]