AI 资讯日报：2026-07-25
岗位趋势
Full-stack interview volume 2026-06 创纪录：AI integration 成新 required skill
FinalRoundAI 7 月数据（7-25 引用，816K+ interview sessions）：full-stack developer interview volume 2026-01 触底 283 sessions，**2026-06 达 1,545**——dataset 纪录。Fastest-growing subcategory：**Full Stack Engineer with AI Integration**——LLM API、RAG pipeline、streaming response UI 成 mid/senior 标配。React/Next.js + TypeScript + Tailwind/shadcn 仍 dominant；**AI integration 是新 required skill** 非 optional nice-to-have。Senior/mid 用 AI tools ship faster 反而 **more competitive**。对前端转全栈：Vercel AI SDK + token limit handling + non-deterministic output testing 应进 portfolio。
来源： FinalRoundAI
行业动态
Post-cutoff DeepSeek：legacy alias 已 dead，surge pricing announced 未生效
deepseek.ai blog **7-25 更新**（Developers Digest 同步 verify）：**7-24 15:59 UTC alias retirement 已执行**——`deepseek-chat`/`deepseek-reasoner` 返回 error。Official rate card **仍为 flat tier**：V4-Flash $0.14/$0.28 MTok，V4-Pro $0.435/$0.87（July 25 verify unchanged since June launch）。**Peak-hour surcharge announced 但未 active**——Beijing 09:00–12:00、14:00–18:00 double-rate **无 published percentage 与 start date**。V3 off-peak discount 已随 alias 退役。对 batch team：instrument timestamp 为 future surcharge 做准备，**7-25 不必改 scheduler**；priority 是确认 production 已无 legacy name 404。
来源： DeepSeek.ai Blog
Anthropic cyber eval：7-27 通知 3 家受害组织，METR 第三方 review 进行中
BleepingComputer / CyberScoop 7-25–7-30 跟进（7-25 行业讨论）：Anthropic 7-23 halt cyber eval 后，7-27 通知 Irregular partner 与 **3 家 affected organizations**——2 家此前 **未自行检测到** intrusion。计划 release lightly redacted PyPI malware transcript；与 **METR** 独立 review 进行中（access all transcripts + model sampling）。Anthropic 将 expand continuous transcript monitoring + vendor assurance。对 enterprise agent procurement：third-party eval partner 网络隔离 audit 将成 RFP 新条款——与 7-24 Claude Code strictAllowlist 形成 product + process 双轨。
来源： BleepingComputer
热门项目
deepseek-ai/DeepSeek-V4-Flash
Hugging Face **7-25 post-migration** 参照：DeepSeek-V4-Flash MIT license，160GB open weights——284B total / 13B active，1M context，384K max output。Legacy alias 退役后 self-hosted path **不受 API retirement 影响**。Agentic coding benchmark SOTA among open weights（official tech report claim）。对全栈 post-migration：API 用 explicit `deepseek-v4-flash`；若 governance 禁止 hosted API，Flash 160GB 是 realistic self-host target vs Pro 865GB datacenter-only。
产品更新
DeepSeek post-migration：cache hit $0.0028 MTok 成 cost optimization 杠杆
Developers Digest **7-25 verify** official pricing page：V4-Flash cache miss $0.14 input vs **cache hit $0.0028**——50× discount ratio。Stable system prompt + shared prefix 设计在 post-migration 是 primary cost lever。Concurrency：Flash 2,500 vs Pro 500 req。Thinking mode 现 unified param 非 separate model——agent chain 可 single-model collapse reasoner + tool-call split。对 front-end coding agent：post-7-24 第一件事应是 audit cache hit rate，非仅 confirm model name swap。
Claude Code strictAllowlist：enterprise sandbox 新 baseline
24 AI / Anthropic release notes 7-24–7-25 持续引用：**`sandbox.network.strictAllowlist`** 7-24 随 v2.1.219 上线——explicit destination allowlist，block 全部 other outbound。配合 7-23 cyber eval incident 教训：agent sandbox network policy 从 "default permissive + deny list" 转向 "default deny + allow list" 成 enterprise pattern。Nested subagent depth 3 + DirectoryAdded hook 使 monorepo workspace 变更可 trigger custom audit script。对 self-hosted runner team：runner restart permission loss bug 已 fix——宜本周 rollout 2.1.219。
来源： 24 AI
模型发布
Kimi K3 open weights 倒计时 2 天（7-27 target）
Moonshot platform **7-25**：open weights **2 天后**（官方 target 7-27）——2.8T MoE、1M context、104B activated params。API 自 7-16 可用；Artificial Analysis Index 57；Arena Code WebDev #1。Limitations：须 verified harness（Kimi Code）保留 thinking history；mid-session model switch 致 quality unstable。Together AI / Modal 据 community 信号准备 day-0 hosted inference。对全栈：**7-25–7-26** 完成 API eval baseline，weights 落地后同一 harness 复跑对比 self-hosted vs API drift。
DeepSeek V4-Pro：legacy reasoner 用户实为 Flash 全程——Pro 是 upgrade
Developers Digest / Zenn **7-25 post-mortem** 关键发现：4-24 至 7-24 间 `deepseek-reasoner` alias **始终路由 V4-Flash thinking**，从未 Pro——若 team 以为 reasoner = strongest model，**7-24 前已在 Flash 上跑 eval**。Migrating to `deepseek-v4-pro` 是 **quality upgrade** 非 parity migration，须 re-test prompts。AA Index Pro 52 vs Flash 47，但 per-token ~3×。对 agent pipeline：7-25 应用 Pro 跑 regression suite 确认 upgrade ROI，非 assume silent parity。
论文研究
DeepSeek V4 structural innovation：Token-wise compression + DSA 1M default
DeepSeek official tech report（7-25 migration 期 re-read）：V4 **Token-wise compression + DeepSeek Sparse Attention (DSA)** 实现 1M context default——vs V3.2 128K 为 **8× expansion**。Dedicated agent optimization：Claude Code/OpenClaw/OpenCode 集成。对全栈 post-migration：若 codebase 含 context-chunking/compression layer 为 V3 128K 设计，V4 1M window 是 **delete-code opportunity**——migration sprint 宜 include "remove obsolete chunking" ticket，非仅 swap model string。