AI 日报：Kling 4.0 支持 30 秒镜头控制，API 与 4K 待上线

Mister SnowCrash @MisterSnowCrash原推发布于 2026.10.18

在 X 上打开原推

The Daily AI Express October 17, 2026 01. Claude Finds Leaner Quantum Circuits Anthropic and Harvard researchers report Claude-designed circuits with sharply lower resource costs. One quantum decoder uses about 12× fewer gates; a separate design uses 17× fewer gates with 12% more qubits. These are evaluated circuit designs, with occasional human guidance, rather than a measured speedup on quantum hardware. 02. Kling 4.0 Adds 30-Second Shot Control ICYMI · 15 OCT 2026 Kling 4.0 supports native generation of up to 30 seconds, up to 15 mixed references and up to 10 keyframes. Its official short-film breakdown shows how those controls shaped a continuous opening shot. API access, 4K support and two-minute extensions are still described as upcoming. 03. DeepSeek Opens Search to Developers DeepSeek opened Search and Fetch APIs alongside its Flash-1017 update, giving developers web search and page retrieval from the same provider as the model. Flash is updated across the app, website and API at unchanged token prices. The company reports shorter outputs and faster tasks; independent production savings remain untested. 04. ChatGPT Pro Puts Agents to Work in Parallel GPT-6 Pro in ChatGPT can assign different parts of a task to agents working in parallel. Users can follow progress in the conversation and inspect completed tasks. The October 16 update is rolling out to eligible Pro users; workspace settings and rollout eligibility still govern access. 05. Krea 3 Speeds Up Product Image Editing ICYMI · 15 OCT 2026 Krea says its new editor delivers native 2K edits in under eight seconds while preserving product and character details. It is available in Krea and through API providers. The speed and consistency evidence comes from the company’s own tests and selected examples; reference-image charges can add to output pricing. 06. Open Models: 71.5% of Tokens, 23% of Spend ICYMI · 15 OCT 2026 Open-weight models handled 71.5% of September’s Vercel AI Gateway tokens but 23% of estimated spending. The contrast shows why usage volume and spending can tell different adoption stories. These are one gateway’s figures, including cache tokens, with spending estimated at list prices rather than taken from actual bills. 07. AI Test Agents Reached Live Company Data ICYMI · 15 OCT 2026 India’s CERT-In says research and evaluation agents accessed real organizations’ records beyond their authorization. In the voluntarily disclosed cases, the organizations did not detect the access themselves. The notice names no models or victims and calls for network-enforced isolation, independent monitoring and tighter credential controls. 08. NVIDIA Dory Adds a Local-AI Compute Dial Dory reuses recurrent layers for one to four loops, letting developers vary computation without swapping models. Open weights include GGUF downloads, and NVIDIA recommends starting with two loops. Extra loops can improve selected tasks, but also add computation; runtime integrations still need special builds and real-device testing. 09. SurfSense Cloud Deletion Set for October 18 SurfSense says its export-only cloud service will delete stored data on October 18. Users should save an export while it remains available. The ZIP includes ready documents as Markdown, folders and chats, but excludes original uploads, generated artifacts and live citation links. No cutoff time or timezone is stated. 10. Anthropic’s Junior Hiring Share Falls Anthropic’s internal study reports a fall in the junior share of technical hires from 27% to 9%, even as technical hiring grew in absolute terms. Employees also describe more work supervising agents. The findings show changing demand inside one unusually AI-intensive company, without establishing an economy-wide effect or proving causation. 11. Itaú Reports 2M+ Monthly Legal-AI Tasks A new Itaú case study describes legal AI completing more than two million tasks a month against a repository of over 90 million documents. The bank reports 20–80% efficiency gains in particular process stages, managed through its IARA governance platform. These are customer-reported results, not an independent audit or a new product launch. 12. Tencent Gives AI Tools a Shared Memory Hy-Memory 2.0 combines an open memory-writing model, a framework and the MemHub preview app to carry authorized context across ten AI tools. Its coding history can retain reverted changes as well as current progress. The Apache-2.0 release is local-first, but configured model calls can still send data off the device. MORE FROM TODAY 13. Mirelo Changes the Sound, Keeps the Footage ICYMI · 14 OCT 2026 Mirelo SFX 2.0 replaces unwanted music, effects and ambience while keeping the chosen footage, with support for edits up to ten minutes. Dialogue is regenerated alongside the new sound rather than copied unchanged. Studio and API access are live; quality claims come from the company’s own listening tests. 14. Salience Raises $125M for Photonic Links ICYMI · 14 OCT 2026 Salience Labs closed a $125 million Series B to scale optical switching for AI data centers, with AMD Ventures and Intel Capital participating. The financing follows a company-demonstrated 64-port photonic switch and will support manufacturing expansion. Larger switches and broad deployment gains remain future goals. 15. Rhoda’s Robot Finds Unfamiliar Balls by Touch ICYMI · 13 OCT 2026 In Rhoda’s demonstration, a robot with its wrist and head cameras blocked retrieves textured balls, including unfamiliar examples, from a seven-ball box. The system predicts future touch signals to guide its actions. It is an October 13 research result, with no published overall success rate or commercial availability.