Seek and Destroy is getting a new engine. The 1990s military action game, first released on the Amiga and later brought to DOS, is being rebuilt through an

While the chatbot still seems to lean left on many issues, its bias is changing.

AIを活用した採用支援ツール「OpenSeek」を手掛ける株式会社CAPERへの出資が発表されました。今後の展開が期待されます。

[株式会社CAPER] [画像1: https://prcdn.freetls.fastly.net/release_image/165302/1/165302-1-67825f6abee954829393ed0c8cc92b4d-1920x1080.png?width=536&quality=85%2C75&format=jpeg&auto=webp&fit=bounds&am...

DeepSeek may have originated in China, but it stopped being Chinese the minute it was released on Hugging Face with an accompanying paper detailing its development. Soon after, a range of developers, including the Beijing Academy of Artificial Intelligence (BAAI), scrambled to replicate DeepSeek’s success but this time as open source software. BAAI, for its part, launched OpenSeek, an ambitious effort to take DeepSeek’s open-weight models and create a project that surpasses DeepSeek while uniting “the global open source communities to drive collaborative innovation in algorithms, data, and systems.” If that sounds cool to you, it didn’t to the U.S. government, which promptly put BAAI on its “baddie” list. Someone needs to remind U.S. (and global) policymakers that no single country, company, or government can contain community-driven open source. ## The moment that keeps on going It’s increasingly common in AI circles to refer to the “DeepSeek moment,” but calling it a moment fundamentally misunderstands its significance. DeepSeek didn’t just have a moment. It’s now very much a movement, one that will frustrate all efforts to contain it. DeepSeek, and the open source AI ecosystem surrounding it, has rapidly evolved from a brief snapshot of technological brilliance into something much bigger—and much harder to stop. Tens of thousands of developers, from seasoned researchers to passionate hobbyists, are now working on enhancing, tuning, and extending these open source models in ways no centralized entity could manage alone. For example, it’s perhaps not surprising that Hugging Face is actively attempting to reverse engineer and publicly disseminate DeepSeek’s R1 model. Hugging Face, while important, is just one company, just one platform. But Hugging Face has attracted hundreds of thousands of developers who actively contribute to, adapt, and build on open source models, driving AI innovation at a speed and scale unmatched even by the most agile corporate labs. Hugging Face by itself could be stopped. But the communities it enables and accelerates cannot. Through the influence of Hugging Face and many others, variants of DeepSeek models are already finding their way into a wide range of applications. Companies like Perplexity are embedding these powerful open source models into consumer-facing services, proving their real-world utility. This democratization of technology ensures that cutting-edge AI capabilities are no longer locked behind the walls of large corporations or elite government labs but are instead openly accessible, adaptable, and improvable by a global community. ## It’s Linux all over again In many respects, the rise of open source AI mirrors the trajectory of Linux from decades past. What started as the passion project of a lone developer quickly blossomed into an essential, foundational technology embraced by enterprises worldwide. Linux won not because it had superior backing from governments or major corporations, though it eventually earned both. No, Linux initially won out precisely because it captivated developers who embraced its promise and contributed toward its potential. Its inherent openness enabled rapid, organic growth driven by practical necessity and persistent innovation from a dedicated global community. We are witnessing a similar phenomenon with DeepSeek and the broader open source AI ecosystem, but this time it’s happening much, much faster. AI adoption and innovation cycles now run in weeks and months rather than the years and decades typical of previous technology revolutions. Organizations that cling to proprietary approaches (looking at you, OpenAI!) or attempt to exert control through restrictive policies (you again, OpenAI!) are not just swimming upstream—they’re attempting to dam an ocean. (Yes, OpenAI has now started to talk up open source, but it’s a long way from releasing a DeepSeek/OpenSeek equivalent on GitHub.) ## You can’t stop us now This brings us to policymakers and the role governments imagine they can play. The current geopolitical approach to AI governance—epitomized by the restrictions imposed on BAAI—is at best shortsighted, and at worst, actively harmful. Open source isn’t subject to export controls or trade embargoes. It’s a pull request away, all day every day. Trying to block or slow the open source AI movement through policy edicts fundamentally misunderstands the decentralized, organic nature of technological evolution itself. Indeed, the war between countries and corporations to dominate AI is already over. It’s too soon to say that open source AI has won (most models we regularly use are proprietary), but directionally, all signs point to open source. That said, it remains to be seen just how open the big players are willing to be: Meta’s Llama 4, for example, is exceptionally impressive but Meta seems determined to stick to its “open enough” licensing stance. The recent history of technology underscores why policy-driven containment strategies inevitably fail. Open source ecosystems operate as living organisms, continuously mutating and adapting. Unlike centralized corporate projects, open source initiatives such as OpenSeek rapidly evolve, leveraging global collaboration at scale. Innovations that might take years in closed environments often happen within weeks when code, ideas, and data flow freely. AI policymakers would do well to understand this dynamic sooner rather than later. Restrictive policies won’t just fail to halt the spread of open source AI; they’ll inadvertently harm domestic companies and push innovation—and critical technological leadership—away. (President Trump is madly attempting to shift the balance of AI innovation away from the U.S. with tariffs that threaten to halt AI infrastructure spending by Microsoft, Google, Amazon, and Meta, among others.) The unstoppable nature of the open source AI movement—exemplified by DeepSeek and its global progeny—signifies a profound shift in technological development. No one can own this wave, no one can stop it, and no one can contain it. Embrace it, shape it, contribute to it, or risk being left behind. Open source AI isn’t just the future; it’s already reshaping the present.

截至 2026/05 ，coding agent 这个赛道在 GitHub 上已经分成几条不同的线： * **闭源旗舰** : Claude Code （ Anthropic 官方），Cursor 。体验最好，但贵 + 锁定。 * **开源通用** : OpenCode （ 164k+ Star ，Go 写的）、Aider （ Python ，VSCode 替代）、Cline 。 * **DeepSeek 原生** : DeepSeek-TUI （ Rust 写的，过去一周从 ~2k 涨到 24k+ Star ）。 * **GUI 切换器** : cc-switch （ 78k+ Star ，Tauri 写的，管理上面这些 agent 的 provider 配置）。 如果诉求是 **「我想要 Claude Code 同款的工程（ MCP / LSP / Plan-Agent-YOLO 三模式 / 真正的 sub-agent fan-out ），但走 DeepSeek V4 把成本压一个数量级」** ，目前 OpenCode + 手动配 DeepSeek 是一条路，但 OpenCode 的 harness 不是为 DeepSeek 优化的（ cache 命中率不显示、reasoning_content 没 replay ）。 DeepSeek-TUI 是另一条路，但是 Rust ，需要装 toolchain 。 第三条路是我两周前推上 GitHub 的 openseek —— **TS + Bun 写的 TUI agent ，27 providers ，第一公民支持 DeepSeek V4** 。目前 150 Star 下面分享一下做这个东西时踩过的几个工程坑，挑了 5 个对 DeepSeek 用户最有用的。 * * * ## 1. cache-aware prompt assembly （最重要） DeepSeek V4 的 prefix cache 把 cached tokens 算 ~10% 价。如果你不主动 claim 这个 discount ，你的成本优势会一夜砍半。 大部分手搓 agent loop 第一次在 turn 之间 format 不一致就把 cache 干掉了。 * * * ## 2. reasoning_content replay V4 thinking mode 会发 internal-monologue tokens ，用户看不到。非显然部分：**下一个 turn 你必须把这些 reasoning tokens 重新喂回模型** ，不然它会失去线索。 ## 3. 5 种 compaction ，不是 1 种 1M context 不等于 1M usable context 。OpenSeek 实现了 5 种： 策略 | 触发 | 用途 ---|---|--- `micro` | 每个 tool result 后 | 砍 noise （ grep 结果、文件 head/tail ） `auto` | 累计 token 阈值 | 自动 summary 老段 `reactive` | 模型主动请求 | 模型自己说「这段我不需要了」 `sessionMemory` | session 结束 | 跨 session 持久化 (extractMemories) `snip` | 用户手动 `/snip` | 精确剪掉一段 ## 4. 真的 MCP 客户端 (不是 stub) 支持 Linear / GitHub / Postgres / Browser 任何 MCP server 一行配置接入。 ## 5. LSP 跑对应语言的 language server (tsserver / rust-analyzer / pyright / gopls / clangd)，每次 edit 之后把 diagnostics 喂回模型。 * * * ## 整体 spec (v1.0) * 27 providers （ OpenAI-compat 19 个 + Anthropic 4 个 + Google 2 个 + ollama + 自定义） * 52 个工具 + 108 slash commands （ CCB 反编译参考 + 自写） * Plan / Agent / YOLO 三模式 + git worktree 隔离 * MCP client (stdio + SSE + websocket) * LSP feedback (5 种语言) * SessionMemory 跨 session 持久化 * HTTP/SSE Runtime API （ headless 跑） * SQLite-backed 任务队列（跨重启） * Cost meter + wallet 集成 文档: openseek.dev 仓库: github.com/LichAmnesia/openseek 欢迎 star 、提 issue 、聊 DeepSeek V4 用法。如果你跑过其他 TUI agent 觉得这个有问题，DM 我（ X @ShenHuang / V2EX 都行）。

截至 2026/05 ，coding agent 这个赛道在 GitHub 上已经分成几条不同的线： * **闭源旗舰** : Claude Code （ Anthropic 官方），Cursor 。体验最好，但贵 + 锁定。 * **开源通用** : OpenCode （ 164k+ Star ，Go 写的）、Aider （ Python ，VSCode 替代）、Cline 。 * **DeepSeek 原生** : DeepSeek-TUI （ Rust 写的，过去一周从 ~2k 涨到 24k+ Star ）。 * **GUI 切换器** : cc-switch （ 78k+ Star ，Tauri 写的，管理上面这些 agent 的 provider 配置）。 如果诉求是 **「我想要 Claude Code 同款的工程（ MCP / LSP / Plan-Agent-YOLO 三模式 / 真正的 sub-agent fan-out ），但走 DeepSeek V4 把成本压一个数量级」** ，目前 OpenCode + 手动配 DeepSeek 是一条路，但 OpenCode 的 harness 不是为 DeepSeek 优化的（ cache 命中率不显示、reasoning_content 没 replay ）。 DeepSeek-TUI 是另一条路，但是 Rust ，需要装 toolchain 。 第三条路是我两周前推上 GitHub 的 openseek —— **TS + Bun 写的 TUI agent ，27 providers ，第一公民支持 DeepSeek V4** 。目前 150 Star 下面分享一下做这个东西时踩过的几个工程坑，挑了 5 个对 DeepSeek 用户最有用的。 * * * ## 1. cache-aware prompt assembly （最重要） DeepSeek V4 的 prefix cache 把 cached tokens 算 ~10% 价。如果你不主动 claim 这个 discount ，你的成本优势会一夜砍半。 大部分手搓 agent loop 第一次在 turn 之间 format 不一致就把 cache 干掉了。 * * * ## 2. reasoning_content replay V4 thinking mode 会发 internal-monologue tokens ，用户看不到。非显然部分：**下一个 turn 你必须把这些 reasoning tokens 重新喂回模型** ，不然它会失去线索。 ## 3. 5 种 compaction ，不是 1 种 1M context 不等于 1M usable context 。OpenSeek 实现了 5 种： 策略 | 触发 | 用途 ---|---|--- `micro` | 每个 tool result 后 | 砍 noise （ grep 结果、文件 head/tail ） `auto` | 累计 token 阈值 | 自动 summary 老段 `reactive` | 模型主动请求 | 模型自己说「这段我不需要了」 `sessionMemory` | session 结束 | 跨 session 持久化 (extractMemories) `snip` | 用户手动 `/snip` | 精确剪掉一段 ## 4. 真的 MCP 客户端 (不是 stub) 支持 Linear / GitHub / Postgres / Browser 任何 MCP server 一行配置接入。 ## 5. LSP 跑对应语言的 language server (tsserver / rust-analyzer / pyright / gopls / clangd)，每次 edit 之后把 diagnostics 喂回模型。 * * * ## 整体 spec (v1.0) * 27 providers （ OpenAI-compat 19 个 + Anthropic 4 个 + Google 2 个 + ollama + 自定义） * 52 个工具 + 108 slash commands （ CCB 反编译参考 + 自写） * Plan / Agent / YOLO 三模式 + git worktree 隔离 * MCP client (stdio + SSE + websocket) * LSP feedback (5 种语言) * SessionMemory 跨 session 持久化 * HTTP/SSE Runtime API （ headless 跑） * SQLite-backed 任务队列（跨重启） * Cost meter + wallet 集成 文档: openseek.dev 仓库: github.com/LichAmnesia/openseek 欢迎 star 、提 issue 、聊 DeepSeek V4 用法。如果你跑过其他 TUI agent 觉得这个有问题，DM 我（ X @ShenHuang / V2EX 都行）。

Slashdot: News for nerds, stuff that matters. Timely news source for technology related news with a heavy slant towards Linux and Open Source issues.

DGDVがAI採用人事プラットフォーム『OpenSeek』を手掛けるCAPERに出資し、新しい採用プロセスの確立を目指します。

DGDV has announced its seed round investment in CAPER, an AI recruitment platform. CAPER offers solutions for comprehensive hiring processes, tackling the labor shortage challenge.