Krea 插件让 Grok 支持草图转动画
Krea 宣布其插件已接入 Grok,用户可在 Grok 内把草图转换成卡通动画,只需向 Grok 提出请求即可。
#产品更新#多模态#视频
Krea 宣布其插件已接入 Grok,用户可在 Grok 内把草图转换成卡通动画,只需向 Grok 提出请求即可。
#产品更新#多模态#视频
fx v0.0.16 This version has better prompt caching and smaller downloads & binaries. • 41% fewer uncached input tokens • About 60% smaller libfx npm downloads • 2.8% smaller CLI binary Changelog → https://t.co/9n2qtJfh4E https://t.co/GjeEEDT8ze
#现象/趋势#Agent
#评测/基准#视频#开源生态
OpenAI 发布 GPT-Rosalind-6 系统卡附录,该模型与 GPT-6 Astra 使用同一套权重,未新增训练数据或后训练,仅通过可信访问提供给经过审核的科研机构、企业和政府合作方。在最高推理预算下,MedChemBench v7 平均评分 49.6%(GPT-Rosalind-5.5 为 28.9%),GeneBench Pro v13 单次尝试成功率 37.7%(前代为 16.7%),LifeSciBench Gold 平均分 61.3%(前代为 44.0%)。按 Preparedness Framework,该发布在生物与化学领域定为 High 能力、未达 Critical 阈值,网络安全领域定为 Critical,AI 自我改进低于 High。
#安全/对齐#OpenAI#数据/训练
Figma Weave 将 AI 生成与编辑工具整合为基于节点画布的可复用工作流,并新增与 Figma 画框并行的使用方式。工作流支持分支探索、同一提示词跨多个 AI 模型对比结果,内置色彩校正、合成、色阶、裁剪、修剪、绘画等编辑节点。Weave 工具已开放社区发布,队友无需操作节点即可直接运行打包好的流程。

#产品更新#Agent#多模态
Baseten 工程博客称,约 770GB 的 GLM-5.3 检查点增量权重同步从平均约 36 秒降到 15 秒,请求暂停从 6 秒降到 0.8 秒。新设计以 GPU 上的权重作为父副本,不再在 CPU 内存保留完整副本,并用 Adler-32 校验和按变更字节做常数时间更新。省下的主机内存用于 KV cache 卸载,在并发 48 时缓存命中率从 2% 升到 92%,单 GPU 输出吞吐提升 3.2 倍;Kimi K3 的 8.4GB 增量同步为 75 秒、暂停 4.1 秒。

#教程/实践#部署/工程#推理
Google 首颗搭载 TPU 的数据中心卫星经过约三周在轨测试与校准后进入正式运行阶段,工程师本周首次向卫星发送提示指令,由预装的 Google DeepMind Gemma 4 模型创作了一首关于太空的俳句。该卫星可产生最高约 1 千瓦电力,因 TPU 功耗较高每次只能启动芯片约 20 分钟;Google 接下来将评估 TPU 在温度变化和宇宙辐射下的可靠性,若验证稳定,计划于 2027 年再发射两颗采用更先进芯片的后续卫星。
#行业动态#Google#部署/工程
星巴克计划在汽车餐厅(Drive-thru)引入AI个性化菜单,会员驾车抵达时系统可识别身份,并根据消费记录和偏好自动展示推荐饮品、简化支付流程。公司预计从2027财年开始测试后台AI模型,用于辅助员工准确录入订单,目前尚未公布测试门店数量、AI供应商和全面推广时间表。星巴克管理层披露,约一半顾客使用Drive-thru服务。
#行业动态#Agent
富士胶片计划在2026财年至2028财年投资6400万美元扩建日本神奈川县小田原市工厂,将磁带产能提升至当前约五倍,以应对云服务商、数据中心和企业客户增长的需求。扩产重点是LTO-10磁带,单卷原生容量40TB、压缩后最高100TB,传输速度每秒400MB。生成式AI带来的海量数据长期归档需求,加上磁带可物理隔离以降低勒索软件风险,使这项技术重新受到关注。
#行业动态#数据/训练
We’re renaming our Max plans: Max 5x → Max Max 20x → Ultra Max is still $100/month, and Ultra is still $200. If you’re a subscriber, there’s nothing you need to do. We’re also removing the 5-hour limit for verified Ultra accounts, starting today with most existing subscribers. Newer accounts keep a temporary limit until they’re verified, which helps us prevent fraud and abuse. You'll get an email when it's removed. Weekly limits remain unchanged. View our plans here: https://t.co/niRpyGF5Me
#产品更新#Anthropic
#模型发布#推理#开源生态
#模型发布#推理#开源生态
#模型发布#推理#开源生态
#模型发布#推理#开源生态
#开源/仓库#推理#部署/工程
hiding amongst the greats. see you at https://t.co/HfcmDb4hBz https://t.co/QF6it1JqTR
#行业动态
Cleo Abram (15M followers across all platforms) just released a new video about @neuralink. She visited the company's Austin HQ. "This is Neuralink’s Next-Gen R1 surgical robot. Its main job is to put a computer chip into people's brains, and this is the first time they're unveiling it." https://t.co/wzTuUq2Or0
#行业动态#具身智能
Drop 2/15 Are we actually measuring the things we ultimately want our voice AI models to achieve? The best AI voices are becoming hard to tell apart. From each other. Not from humans. 🚀 Introducing Turing TTS: an evaluation that measures how often human listeners mistake AI-generated speech for real human speech. We tested leading TTS models against real human speech. Even the best model fooled listeners only ~10% of the time. With Turing TTS, we hope to contribute to a broader discussion within the research community about how we measure progress in voice AI. If one of our goals is to develop speech that is indistinguishable from human speech, how well do our current evaluation methods capture progress toward that goal? We don’t claim that Turing TTS is the right way forward, or that human indistinguishability should be the goal of every TTS system. Rather, we hope these findings encourage the community to collectively examine whether blind pairwise human evaluations of naturalness are sufficient, and what else we might need to measure. Results are now live on @voicearena_ai.
#评测/基准#语音#Google
reddit basically one of the last places i search for human results and now google is shredding that. does the open web even exist in ten years?
#其他
#评测/基准#Agent#编码