DeepSeek 永久下调其旗舰 AI 模型价格 75%

[00:00] Welcome to Daily Tech briefing. Here are the most important AI and technology stories for today.欢迎收听每日科技英文速读。下面是今天最值得关注的 AI 与科技新闻。---[00:08] Everyone is navigating AI security in real time - even Google. Google Cloud COO Francis de Souza says security can no longer be an afterthought for AI. He warns about shadow AI where employees use consumer tools without oversight. The average time from breach to handoff has dropped from 8 hours to just 22 seconds. De Souza argues companies need an AI-native, fully agentic (具有自主行动能力的) defense. But Google itself faces criticism. Developers reported surprise five-figure bills after compromised API keys accessed Gemini. And even after deleting a key, attackers can still use it for up to 23 minutes because revocation (吊销撤销) spreads slowly. The gap between what platforms prescribe and how fast they adapt is still wide open.所有人都在实时摸索 AI 安全——Google 也不例外。Google Cloud 首席运营官 Francis de Souza 表示，安全不能再是 AI 的附加项。他警告说员工在无监管情况下使用消费级工具形成影子 AI。从入侵到攻击交接的平均时间从 8 小时缩短到仅 22 秒。但 Google 自身也面临批评：有开发者因被入侵的 API 密钥访问 Gemini 而收到上万美元账单。即使立即删除密钥，攻击者仍可在 23 分钟内继续使用。平台开出的药方与其自身的适应速度之间仍有巨大差距。---[00:50] DeepSeek makes permanent 75% discount on its flagship AI model. DeepSeek has made its 75% discount on V4-Pro permanent. This is a huge move in the AI pricing war. First-party pricing now sits at roughly 18 cents per million tokens blended (混合的). That is about 3 times cheaper than Gemini 3.1 Pro Preview. It is 12 times cheaper than GPT-5.5. And it is 19 times cheaper than Claude Opus 4.7. The community calls this intelligence too cheap to meter. The price cut changes the entire cost-performance frontier (前沿) for AI inference. Developers can now run frontier (前沿)-level models at a fraction of previous costs.DeepSeek 永久下调其旗舰 AI 模型价格 75%。DeepSeek 已将 V4-Pro 75% 的折扣永久化。这是 AI 价格战的重大举措。自营定价混合约每百万 token 18 美分，比 Gemini 3.1 Pro Preview 便宜约 3 倍，比 GPT-5.5 便宜 12 倍，比 Claude Opus 4.7 便宜 19 倍。社区将其称为智能便宜到不计量。这次降价彻底改变了 AI 推理的性价比边界。---[01:30] I tried Amazon's Bee wearable (可穿戴设备) and I am both intrigued (感到好奇的) and creeped out. Amazon's Bee is an AI wrist wearable (可穿戴设备) that records, transcribes, and summarizes your conversations all day. It acts like a personal assistant that never forgets. During a business call Bee gave a helpful summary. It even figured out during a movie night that people were analyzing a Tarantino film. But the device raises serious privacy concerns. Bee needs access to your location, photos, contacts, calendar and notifications. All data is stored in the cloud. The reviewer says it has promise for professionals but feels too invasive for personal life.我体验了亚马逊的 Bee 可穿戴设备，既好奇又不安。亚马逊 Bee 是一款 AI 腕戴设备全天候记录转录并总结对话，像永不遗忘的个人助理。商务电话中 Bee 提供了有用摘要，甚至能识别出观影之夜大家在讨论塔伦蒂诺电影。但它引发严重隐私担忧：需访问位置照片通讯录日历和通知，所有数据存云端。评测者认为对专业人士有潜力但用于个人生活过于侵入。---[02:05] Xreal, Google's smartglasses partner, thinks it has finally cracked the industry. Xreal is launching Project Aura, its latest smart glasses with OLED displays. The glasses connect to a small puck-shaped computer that fits in your pocket. They offer Google Maps integration, VR YouTube and hand-tracking painting apps. Xreal CEO Chi Xu says the industry is at an inflection point (转折点). Meta's Ray-Ban partnership showed there is demand. But almost every company is still losing money. Xreal hopes to break even (盈亏平衡) next year and is working on an IPO expected before end of 2026.Google 智能眼镜合作伙伴 Xreal 认为行业难题终于被攻克。Xreal 推出最新智能眼镜 Project Aura，配备 OLED 显示屏。眼镜需连接可放入口袋的小型冰球状电脑。支持 Google 地图集成、VR YouTube 和手势追踪绘画。CEO 徐驰表示行业处于转折点，Xreal 希望明年盈亏平衡，计划 2026 年底前 IPO。---[02:39] DeepSeek launches Reasonix, a native coding agent with high caching (缓存) and low cost. DeepSeek has launched Reasonix, a native coding agent designed for high caching (缓存) efficiency and low operational cost. It builds on the company's aggressive pricing strategy. The coding agent market is getting very competitive. OpenAI recently shipped Codex updates with Appshots and remote computer use. Claude Devs expanded auto mode to the Pro plan and added Sonnet 4.6 support. DeepSeek's entry shows that the coding agent space is no longer just for US companies.DeepSeek 推出 Reasonix——原生高缓存低成本编程 Agent。DeepSeek 推出 Reasonix 原生编程 Agent，以高缓存效率和低成本为设计目标。编程 Agent 市场竞争激烈：OpenAI 为 Codex 增加 Appshots 和远程电脑使用；Claude Devs 将自动模式扩展到 Pro 计划。DeepSeek 入场表明编程 Agent 不再是美国公司独占领域。---[03:16] Memory now accounts for nearly two-thirds of AI chip costs. High-bandwidth memory (高带宽内存（HBM）) has grown from 52% to 63% of total AI chip spending between early 2024 and late 2025. This covers chips from Nvidia AMD Google and Amazon. HBM spending grew from 12 billion dollars in 2024 to 32 billion in 2025. Memory supply remains tight and prices are rising. Microsoft's 190 billion dollar FY2026 capex (资本支出) outlook includes about 25 billion from higher component prices. Meta also raised its 2026 capex (资本支出) by 10 billion citing higher component costs.内存已占 AI 芯片成本的近三分之二。从 2024 年初到 2025 年底，高带宽内存在 AI 芯片总支出占比从 52% 增至 63%。涵盖 Nvidia AMD Google 和 Amazon 芯片。HBM 支出从 2024 年 120 亿美元增至 2025 年 320 亿美元。内存供应紧张价格上涨，微软 2026 财年 1900 亿美元资本支出中约 250 亿来自组件涨价。---[03:56] Apple introduces PICO, an AI image codec (编码解码器) that beats traditional standards. Apple has introduced PICO, Perceptual Image Codec (编码解码器). It is the first learned codec (编码解码器) optimized for the human visual system. PICO provides 2.3 to 3 times bitrate (比特率) savings over AV1 and VVC. It beats other AI-based codecs by 20 to 40 percent. On iPhone 17 Pro Max it encodes 12-megapixel images in 230ms and decodes in 150ms. That is faster than most ML codecs on a V100 GPU. PICO also guarantees cross-platform robustness which is rare for learned codecs.Apple 推出 PICO——超越传统标准的 AI 图像编码器。Apple 推出 PICO（感知图像编码器），首个针对人类视觉系统优化的学习型编码器。比 AV1 和 VVC 节省 2.3 到 3 倍码率，比基于 AI 的竞品好 20% 到 40%。在 iPhone 17 Pro Max 上 230ms 编码 1200 万像素图像 150ms 解码，比多数在 V100 GPU 上运行的 ML 编码器更快。---[04:36] New study reveals constraint decay (约束衰减) in LLM agents for backend code. A new paper introduces constraint decay (约束衰减). When structural requirements pile up for backend code generation LLM agents get significantly worse. The study tested agents across 8 web frameworks on 100 tasks. Top configurations lost 30 points on average in test pass rates from simple to fully specified tasks. Weaker ones dropped close to zero. Agents did well with minimal frameworks like Flask but struggled badly with convention-heavy ones like Django and FastAPI. Main problem was the data layer: incorrect database queries and ORM (对象关系映射) runtime errors.新研究揭示 LLM Agent 在后台代码中的约束衰减现象。新研究论文提出约束衰减概念。后台代码生成结构性要求增加时 LLM Agent 表现显著下降。覆盖 8 种 Web 框架 100 个任务后，顶级配置平均通过率下降 30 个百分点。Agent 在 Flask 等简洁框架表现不错但在 Django 和 FastAPI 等高度约定化框架问题严重。主要瓶颈在数据层的错误数据库查询和 ORM 运行时错误。---[05:14] 📚 今日单词：constraint- n. 约束，限制条件- [kEn'streInt]- 词根：源自拉丁语 constringere 由 con-（一起）+ stringere（拉紧）组成。---[05:19] 词组/例句- structural constraint 结构性约束- Modern AI agents must learn to work within structural constraints while maintaining flexibility. 现代 AI Agent 必须学会在结构性约束内工作同时保持灵活性。---[05:38] That's all for today. See you tomorrow, and thank you for listening.以上就是今天的内容。感谢你的收听，我们明天见。前往小宇宙评论区与主播互动