[{"data":1,"prerenderedAt":709},["ShallowReactive",2],{"news-\u002Fnews\u002Fopenai-sora-preview":3,"news-prev-\u002Fnews\u002Fopenai-sora-preview":168,"news-next-\u002Fnews\u002Fopenai-sora-preview":215,"news-related-\u002Fnews\u002Fopenai-sora-preview":296},{"id":4,"title":5,"body":6,"category":153,"date":154,"description":155,"draft":156,"extension":157,"meta":158,"navigation":159,"path":160,"seo":161,"source":162,"stem":163,"tags":164,"__hash__":167},"news\u002Fnews\u002Fopenai-sora-preview.md","OpenAI 展示 Sora:AI 生成 60 秒视频,物理模拟接近真实",{"type":7,"value":8,"toc":145},"minimark",[9,13,17,20,23,26,29,105,108,111,114,117,120],[10,11,12],"p",{},"OpenAI 在 2024 年 2 月 15 日展示 Sora 视频生成模型,能根据文字提示生成长达 60 秒的视频,物理模拟能力接近真实,引发行业震动。该模型当时未公开发布,仅向红队和部分创作者开放测试。",[14,15,16],"h2",{"id":16},"事件详情",[10,18,19],{},"Sora 是 OpenAI 的文生视频模型。一段 prompt 进去,出来最长 60 秒的视频,画面里的人物走动、物体遮挡、光影变化基本符合物理规律。发布会放出的几个 demo——东京街头漫步的女性、纸船在浪花里起伏、几只长毛象在雪地行走——质量明显高于同期 Runway Gen-2 和 Pika。",[10,21,22],{},"技术上 Sora 用了 Diffusion Transformer 架构。把视频拆成时空 patch(类似语言模型的 token),在 patch 上做扩散生成,统一处理图像和视频。这种方式让模型能处理不同时长、分辨率、宽高比的素材,生成时也能保持长视频的时间连贯性。OpenAI 没公开训练数据,但外界普遍推测用了大量合成数据和带版权的影视素材,这也埋下了后来的版权官司。",[10,24,25],{},"模型发布后没有立即开放。OpenAI 采取红队测试策略,先让安全研究人员和少数创作者试用,逐步公开。普通用户等到 2024 年 12 月才正式用上 Sora,且仅限 Plus 和 Pro 用户。这个过程里 Google Veo、快手可灵、Runway Gen-3 都相继发布,逼着 OpenAI 加快进度。",[14,27,28],{"id":28},"关键信息",[30,31,32,45],"table",{},[33,34,35],"thead",{},[36,37,38,42],"tr",{},[39,40,41],"th",{},"项目",[39,43,44],{},"参数",[46,47,48,57,65,73,81,89,97],"tbody",{},[36,49,50,54],{},[51,52,53],"td",{},"模型名",[51,55,56],{},"Sora",[36,58,59,62],{},[51,60,61],{},"首次展示日期",[51,63,64],{},"2024 年 2 月 15 日",[36,66,67,70],{},[51,68,69],{},"最大生成时长",[51,71,72],{},"60 秒",[36,74,75,78],{},[51,76,77],{},"架构",[51,79,80],{},"Diffusion Transformer(时空 patch)",[36,82,83,86],{},[51,84,85],{},"公开发布",[51,87,88],{},"2024 年 12 月(Plus\u002FPro 用户)",[36,90,91,94],{},[51,92,93],{},"定价",[51,95,96],{},"Plus 用户有限量,Pro 200 美元\u002F月",[36,98,99,102],{},[51,100,101],{},"同期竞品",[51,103,104],{},"Runway Gen-3、Pika、可灵、Google Veo",[14,106,107],{"id":107},"我们的观察",[10,109,110],{},"我们当时看了所有 demo,第一反应是\"这下视频行业的版图要变\"。但用上一段时间后判断更冷静了:60 秒是个噱头,真正能用的片段大多在 10 秒以内,长视频里仍会出现人物变形、物体凭空消失这类问题。物理模拟\"接近真实\"是营销话术,复杂场景(多人物互动、精细手部动作)还是会露馅。",[10,112,113],{},"定价和算力是更大的门槛。生成一段视频的成本远高于生成图片,Plus 用户的配额很少,真正做商业化内容得用 Pro。这个成本结构让 Sora 短期内很难替代商业拍摄,更多是用在概念验证和素材草图阶段。",[10,115,116],{},"国内可灵、即梦跟得很紧,可灵在人物动作一致性上已经不输 Sora,而且有国内合规优势。视频生成这个赛道,技术差距在缩小,真正的分水岭是版权和算力成本谁先解决。",[14,118,119],{"id":119},"相关工具",[121,122,123,131,138],"ul",{},[124,125,126,130],"li",{},[127,128,56],"a",{"href":129},"\u002Ftools\u002Fvideo\u002Fsora"," — OpenAI 的视频生成模型",[124,132,133,137],{},[127,134,136],{"href":135},"\u002Ftools\u002Fvideo\u002Frunway","Runway"," — 老牌 AI 视频工具",[124,139,140,144],{},[127,141,143],{"href":142},"\u002Ftools\u002Fvideo\u002Fkling","可灵"," — 快手出品的视频生成模型",{"title":146,"searchDepth":147,"depth":147,"links":148},"",2,[149,150,151,152],{"id":16,"depth":147,"text":16},{"id":28,"depth":147,"text":28},{"id":107,"depth":147,"text":107},{"id":119,"depth":147,"text":119},"产品发布","2024-02-15","OpenAI 2024年2月15日展示 Sora 视频生成模型,能根据文字提示生成长达 60 秒的视频,物理模拟能力接近真实,引发行业震动。",false,"md",{},true,"\u002Fnews\u002Fopenai-sora-preview",{"title":5,"description":155},"https:\u002F\u002Fopenai.com\u002Fsora","news\u002Fopenai-sora-preview",[165,56,166],"OpenAI","AI视频","tKTYmFxh7CKXsrw9pjY4eF8MLxW9MqJe_gXZ28Bm5TY",{"id":169,"title":170,"body":171,"category":202,"date":203,"description":204,"draft":156,"extension":157,"meta":205,"navigation":159,"path":206,"seo":207,"source":193,"stem":208,"tags":209,"__hash__":214},"news\u002Fnews\u002Fweekly\u002F2024-02-11.md","AI 周报 · 2024-02-11 当周权威信源更新",{"type":7,"value":172,"toc":199},[173,176,180,196],[10,174,175],{},"本期汇总 2024-02-11 当周来自 Hugging Face Blog 的 1 条更新。内容为信源标题与摘要的客观转述，完整信息请点原文链接。",[14,177,179],{"id":178},"_1-from-openai-to-open-llms-with-messages-api-on-hugging-face","1. From OpenAI to Open LLMs with Messages API on Hugging Face",[121,181,182,185,188],{},[124,183,184],{},"来源：Hugging Face Blog（原文日期 2024-02-08）",[124,186,187],{},"分类：技术更新",[124,189,190,191],{},"原文：",[127,192,193],{"href":193,"rel":194},"https:\u002F\u002Fhuggingface.co\u002Fblog\u002Ftgi-messages-api",[195],"nofollow",[10,197,198],{},"原文标题如上，详情见来源链接。",{"title":146,"searchDepth":147,"depth":147,"links":200},[201],{"id":178,"depth":147,"text":179},"AI 周报","2024-02-11","2024-02-11 当周 AI 行业权威信源更新汇总，共 1 条，来源：Hugging Face Blog。",{},"\u002Fnews\u002Fweekly\u002F2024-02-11",{"title":170,"description":204},"news\u002Fweekly\u002F2024-02-11",[210,211,212,213],"AI周报","每周更新","权威信源","AI资讯","ZX9TYs2m0UqDZlxn7cKHt84q4aY6zrp6HspUtEaFOTg",{"id":216,"title":217,"body":218,"category":202,"date":288,"description":289,"draft":156,"extension":157,"meta":290,"navigation":159,"path":291,"seo":292,"source":238,"stem":293,"tags":294,"__hash__":295},"news\u002Fnews\u002Fweekly\u002F2024-02-18.md","AI 周报 · 2024-02-18 当周权威信源更新",{"type":7,"value":219,"toc":283},[220,223,227,240,243,247,261,263,267,280],[10,221,222],{},"本期汇总 2024-02-18 当周来自 OpenAI News 的 3 条更新。内容为信源标题与摘要的客观转述，完整信息请点原文链接。",[14,224,226],{"id":225},"_1-video-generation-models-as-world-simulators","1. Video generation models as world simulators",[121,228,229,232,234],{},[124,230,231],{},"来源：OpenAI News（原文日期 2024-02-15）",[124,233,187],{},[124,235,190,236],{},[127,237,238],{"href":238,"rel":239},"https:\u002F\u002Fopenai.com\u002Findex\u002Fvideo-generation-models-as-world-simulators",[195],[10,241,242],{},"We explore large-scale training of generative models on video data. Specifically, we train text-conditional diffusion models jointly on videos and images of variable durations, resolutions and aspect ratios. We leverage …",[14,244,246],{"id":245},"_2-disrupting-malicious-uses-of-ai-by-state-affiliated-threat-actors","2. Disrupting malicious uses of AI by state-affiliated threat actors",[121,248,249,252,255],{},[124,250,251],{},"来源：OpenAI News（原文日期 2024-02-14）",[124,253,254],{},"分类：产品发布",[124,256,190,257],{},[127,258,259],{"href":259,"rel":260},"https:\u002F\u002Fopenai.com\u002Findex\u002Fdisrupting-malicious-uses-of-ai-by-state-affiliated-threat-actors",[195],[10,262,198],{},[14,264,266],{"id":265},"_3-memory-and-new-controls-for-chatgpt","3. Memory and new controls for ChatGPT",[121,268,269,272,274],{},[124,270,271],{},"来源：OpenAI News（原文日期 2024-02-13）",[124,273,254],{},[124,275,190,276],{},[127,277,278],{"href":278,"rel":279},"https:\u002F\u002Fopenai.com\u002Findex\u002Fmemory-and-new-controls-for-chatgpt",[195],[10,281,282],{},"We’re testing the ability for ChatGPT to remember things you discuss to make future chats more helpful. You’re in control of ChatGPT’s memory.",{"title":146,"searchDepth":147,"depth":147,"links":284},[285,286,287],{"id":225,"depth":147,"text":226},{"id":245,"depth":147,"text":246},{"id":265,"depth":147,"text":266},"2024-02-18","2024-02-18 当周 AI 行业权威信源更新汇总，共 3 条，来源：OpenAI News。",{},"\u002Fnews\u002Fweekly\u002F2024-02-18",{"title":217,"description":289},"news\u002Fweekly\u002F2024-02-18",[210,211,212,213],"6SfshGXVTUG_u9ZYHo-B_8Ze1MaVC6Glj-UOzam82kQ",[297,439,566],{"id":298,"title":299,"body":300,"category":153,"date":428,"description":429,"draft":156,"extension":157,"meta":430,"navigation":159,"path":431,"seo":432,"source":433,"stem":434,"tags":435,"__hash__":438},"news\u002Fnews\u002Fanthropic-claude-35-sonnet.md","Anthropic 发布 Claude 3.5 Sonnet,代码能力超过 GPT-4o",{"type":7,"value":301,"toc":422},[302,305,307,310,313,316,318,393,395,398,401,404,406],[10,303,304],{},"Anthropic 在 2024 年 6 月 20 日发布 Claude 3.5 Sonnet,在编码和推理基准测试上超过 GPT-4o,同时推出 Artifacts 功能,可在对话中实时生成和预览代码、网页、图表。该模型免费向 Claude.ai 用户开放。",[14,306,16],{"id":16},[10,308,309],{},"Claude 3.5 Sonnet 是 Anthropic Claude 3.5 系列的第一款模型。发布当天,它在 SWE-bench(软件工程基准测试,让 AI 修 GitHub 上的真实 bug)上拿到 31.2% 的成绩,超过 GPT-4o(约 22%)和自家上一代 Claude 3 Opus。GPQA(研究生级推理)和 MMLU(知识问答)等基准上同样排在第一梯队。",[10,311,312],{},"价格反而比上一代低。Claude 3.5 Sonnet 的 API 输入 3 美元\u002F百万 token,输出 15 美元\u002F百万 token,和 Claude 3 Sonnet 同价,但能力接近甚至超过更贵的 Claude 3 Opus。",[10,314,315],{},"Artifacts 是这次发布的另一个亮点。用户让 Claude 写一段 React 代码、一个 SVG 图表或一份 HTML 页面,模型会在对话窗口右侧生成一个独立面板,实时渲染结果,可以交互、可以复制。之前这类需求要来回复制代码到本地跑,现在对话里直接看效果。",[14,317,28],{"id":28},[30,319,320,328],{},[33,321,322],{},[36,323,324,326],{},[39,325,41],{},[39,327,44],{},[46,329,330,337,345,353,361,369,377,385],{},[36,331,332,334],{},[51,333,53],{},[51,335,336],{},"Claude 3.5 Sonnet",[36,338,339,342],{},[51,340,341],{},"发布日期",[51,343,344],{},"2024 年 6 月 20 日",[36,346,347,350],{},[51,348,349],{},"SWE-bench 得分",[51,351,352],{},"31.2%(当时最高)",[36,354,355,358],{},[51,356,357],{},"API 输入价格",[51,359,360],{},"3 美元\u002F百万 token",[36,362,363,366],{},[51,364,365],{},"API 输出价格",[51,367,368],{},"15 美元\u002F百万 token",[36,370,371,374],{},[51,372,373],{},"新功能",[51,375,376],{},"Artifacts(对话内实时渲染)",[36,378,379,382],{},[51,380,381],{},"上下文窗口",[51,383,384],{},"200K token",[36,386,387,390],{},[51,388,389],{},"免费可用",[51,391,392],{},"是(Claude.ai 有用量限制)",[14,394,107],{"id":107},[10,396,397],{},"我们日常写代码主要用 Claude 3.5 Sonnet,几个对比结论:写前端组件它比 GPT-4o 强,尤其是 React + Tailwind 的组合,一次出的代码改两行就能跑;长上下文处理(200K)对读整个仓库很有用,GPT-4o 当时还是 128K。中文写作它不如国产模型接地气,但英文输出明显比 GPT-4o 干净,少翻译腔。",[10,399,400],{},"Artifacts 这个功能我们用了一段时间,体验比想象中好。改一个图表的配色,它边写边渲染,你能立刻看到效果。这个交互方式后来被不少产品借鉴,Cursor 的 Composer、v0 都有类似设计。",[10,402,403],{},"定价策略值得注意。Anthropic 把中端型号(Sonnet)卖到接近旗舰的能力,价格还压着 OpenAI,这是在抢开发者市场。Claude 3.5 Opus 和 Haiku 之前没出,2024 年 10 月才补上 3.5 Haiku,Opus 至今没出 3.5 版本。",[14,405,119],{"id":119},[121,407,408,415],{},[124,409,410,414],{},[127,411,413],{"href":412},"\u002Ftools\u002Fchat\u002Fclaude","Claude"," — Anthropic 的对话式 AI 助手",[124,416,417,421],{},[127,418,420],{"href":419},"\u002Ftools\u002Fcoding\u002Fcursor","Cursor"," — AI 代码编辑器,支持切换到 Claude 3.5 Sonnet",{"title":146,"searchDepth":147,"depth":147,"links":423},[424,425,426,427],{"id":16,"depth":147,"text":16},{"id":28,"depth":147,"text":28},{"id":107,"depth":147,"text":107},{"id":119,"depth":147,"text":119},"2024-06-20","Anthropic 2024年6月20日发布 Claude 3.5 Sonnet,在编码和推理基准测试上超过 GPT-4o,同时推出 Artifacts 功能,可在对话中实时生成和预览代码。",{},"\u002Fnews\u002Fanthropic-claude-35-sonnet",{"title":299,"description":429},"https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fclaude-3-5-sonnet","news\u002Fanthropic-claude-35-sonnet",[436,413,437],"Anthropic","AI编程","YNpSEVyGXxoeLs4EYq9596u4EkM39BEPp7_8srryhco",{"id":440,"title":441,"body":442,"category":153,"date":555,"description":556,"draft":156,"extension":157,"meta":557,"navigation":159,"path":558,"seo":559,"source":560,"stem":561,"tags":562,"__hash__":565},"news\u002Fnews\u002Fopenai-gpt-4o-launch.md","OpenAI 发布 GPT-4o:实时语音对话+多模态,免费用户可用",{"type":7,"value":443,"toc":549},[444,447,449,452,455,458,460,530,532,535,538,540],[10,445,446],{},"OpenAI 在 2024 年 5 月 13 日的春季发布会上发布 GPT-4o 模型,支持实时语音对话、图片和视频理解,免费用户可用。这是 OpenAI 首次把旗舰模型免费开放,API 价格比 GPT-4 Turbo 降了一半。",[14,448,16],{"id":16},[10,450,451],{},"GPT-4o 的 \"o\" 代表 omni(全能)。它在单一模型里处理文本、音频和图像三种输入,不再像之前那样把语音对话拆成三段(语音转文字、模型处理文字、文字转语音)。端到端处理的好处是保留了语气、情绪和背景音,响应也更快——平均延迟 320 毫秒,最低 232 毫秒,接近人类对话的反应速度(人类对话平均约 200-300 毫秒)。",[10,453,454],{},"发布会上 OpenAI 演示了几段实时对话。模型能听出呼吸声和笑声,在对话里即兴唱歌,还能通过摄像头看用户画的方程图,实时讲解解题步骤。这些演示一度引发争议,因为部分画面被指经过剪辑,但实际能力基本属实。",[10,456,457],{},"免费用户可以使用 GPT-4o,但有用量限制,达到上限后会切换回更小的模型。Plus 用户(20 美元\u002F月)上限更高,Team 和 Pro 用户依次递增。API 端,GPT-4o 的输入价格 5 美元\u002F百万 token,输出 15 美元\u002F百万 token,比 GPT-4 Turbo(输入 10 美元、输出 30 美元)便宜一半,速度也快了一倍。",[14,459,28],{"id":28},[30,461,462,470],{},[33,463,464],{},[36,465,466,468],{},[39,467,41],{},[39,469,44],{},[46,471,472,479,486,494,502,509,516,522],{},[36,473,474,476],{},[51,475,53],{},[51,477,478],{},"GPT-4o(\"o\" = omni)",[36,480,481,483],{},[51,482,341],{},[51,484,485],{},"2024 年 5 月 13 日",[36,487,488,491],{},[51,489,490],{},"支持输入",[51,492,493],{},"文本、音频、图像、视频",[36,495,496,499],{},[51,497,498],{},"语音响应延迟",[51,500,501],{},"平均 320ms,最低 232ms",[36,503,504,506],{},[51,505,389],{},[51,507,508],{},"是(有用量限制)",[36,510,511,513],{},[51,512,357],{},[51,514,515],{},"5 美元\u002F百万 token",[36,517,518,520],{},[51,519,365],{},[51,521,368],{},[36,523,524,527],{},[51,525,526],{},"相比 GPT-4 Turbo",[51,528,529],{},"降价 50%,速度提升 2 倍",[14,531,107],{"id":107},[10,533,534],{},"我们当时第一时间用上了 GPT-4o 的语音模式。它最大的变化不是\"能对话\",而是\"能听见情绪\"。你叹气它知道你不耐烦,你笑它也跟着笑。这种交互密度是之前语音助手做不到的,但延迟实测比官方数据高,网络差的时候会卡顿。",[10,536,537],{},"免费开放这件事的意义被低估了。之前免费用户只能用 GPT-3.5,和 Plus 差一档,现在免费用户和付费用户用同一个模型,差别只在用量。这个动作直接把竞品的免费档位卡死了——Google 后来跟进把 Gemini 免费开放,Anthropic 也放宽了 Claude 的免费额度。",[14,539,119],{"id":119},[121,541,542],{},[124,543,544,548],{},[127,545,547],{"href":546},"\u002Ftools\u002Fchat\u002Fchatgpt","ChatGPT"," — OpenAI 的对话式 AI 助手,GPT-4o 的主要载体",{"title":146,"searchDepth":147,"depth":147,"links":550},[551,552,553,554],{"id":16,"depth":147,"text":16},{"id":28,"depth":147,"text":28},{"id":107,"depth":147,"text":107},{"id":119,"depth":147,"text":119},"2024-05-13","OpenAI 2024年5月13日发布 GPT-4o 模型,支持实时语音对话、图片和视频理解,免费用户可用。这是 OpenAI 首次将旗舰模型免费开放。",{},"\u002Fnews\u002Fopenai-gpt-4o-launch",{"title":441,"description":556},"https:\u002F\u002Fopenai.com\u002Findex\u002Fhello-gpt-4o\u002F","news\u002Fopenai-gpt-4o-launch",[165,563,564],"GPT-4o","多模态","iIO4NZVUQbQPfdqLW6R9fOE27swfdyB0SY3OZdYSTnc",{"id":567,"title":568,"body":569,"category":153,"date":698,"description":699,"draft":156,"extension":157,"meta":700,"navigation":159,"path":701,"seo":702,"source":703,"stem":704,"tags":705,"__hash__":708},"news\u002Fnews\u002Fsuno-v3-ai-music.md","Suno v3 让 AI 音乐出圈:输入文字生成完整歌曲",{"type":7,"value":570,"toc":692},[571,574,576,579,582,585,587,663,665,668,671,674,676],[10,572,573],{},"Suno 在 2024 年 3 月推出 v3 模型,输入一段文字描述就能生成带人声的完整歌曲,质量接近 Demo 级别。AI 音乐创作从概念验证进入实用阶段,普通人不需要乐理也能做出一首能听的歌。",[14,575,16],{"id":16},[10,577,578],{},"Suno 是一家波士顿的 AI 音乐创业公司,产品形态很简单:输入一句\"给我写一首关于加班的摇滚歌\",30 秒后给你两首完整的歌——有人声、有伴奏、有副歌,每首大约 2 分钟。v3 是 2024 年 3 月发布的版本,相比 v2 在人声清晰度和音乐结构上有明显提升,中文歌词的处理也比之前顺畅。",[10,580,581],{},"v3 之前,Suno 只在 Discord 里运行。v3 发布后上线了独立网站和付费方案:免费用户每天 10 首歌(50 积分),Pro 计划 10 美元\u002F月(2500 积分,约 500 首),Premier 30 美元\u002F月(10000 积分)。商业用途需要订阅 Pro 以上版本,免费版生成的作品不能商用。",[10,583,584],{},"版权是绕不开的问题。Suno 不公开训练数据,但生成的作品风格能明显听出流行音乐的痕迹。2024 年 6 月,美国唱片业协会(RIAA)代表环球、华纳、索尼三大唱片公司起诉 Suno 和 Udio,指控他们未经授权使用受版权保护的录音制品训练模型。案件至今仍在审理。",[14,586,28],{"id":28},[30,588,589,597],{},[33,590,591],{},[36,592,593,595],{},[39,594,41],{},[39,596,44],{},[46,598,599,607,615,623,631,639,647,655],{},[36,600,601,604],{},[51,602,603],{},"模型版本",[51,605,606],{},"Suno v3",[36,608,609,612],{},[51,610,611],{},"发布时间",[51,613,614],{},"2024 年 3 月",[36,616,617,620],{},[51,618,619],{},"单次生成",[51,621,622],{},"2 首 × 2 分钟左右",[36,624,625,628],{},[51,626,627],{},"免费额度",[51,629,630],{},"每天 10 首(50 积分)",[36,632,633,636],{},[51,634,635],{},"Pro 价格",[51,637,638],{},"10 美元\u002F月(2500 积分)",[36,640,641,644],{},[51,642,643],{},"Premier 价格",[51,645,646],{},"30 美元\u002F月(10000 积分)",[36,648,649,652],{},[51,650,651],{},"商用要求",[51,653,654],{},"Pro 及以上订阅",[36,656,657,660],{},[51,658,659],{},"主要竞品",[51,661,662],{},"Udio、网易天音",[14,664,107],{"id":107},[10,666,667],{},"我们试过用 Suno 给一个产品宣传视频配背景音乐,生成的成品直接能用,省了买版权音乐的钱。但要注意两点:一是人声歌词容易出怪腔怪调,中文歌的咬字偶尔会像外国人在唱;二是结构重复,你生成 10 首摇滚,前奏和副歌的走向会撞车,听多了能听出\"AI 味\"。",[10,669,670],{},"工具定位很关键。Suno 适合做草稿、做内部 demo、做不需要严肃版权的物料,不适合直接做发歌发片。一旦涉及商用,版权风险就得自己背——RIAA 的诉讼结果没出来之前,大公司不敢用 Suno 做正式音乐资产。",[10,672,673],{},"AI 音乐对独立音乐人是个双刃剑。一方面它能快速把灵感变成可听的 demo,不必先找编曲;另一方面低门槛生成会挤压低端外包市场,做 BGM、做短视频配乐的人最先受冲击。",[14,675,119],{"id":119},[121,677,678,685],{},[124,679,680,684],{},[127,681,683],{"href":682},"\u002Ftools\u002Faudio\u002Fsuno","Suno"," — AI 音乐生成工具",[124,686,687,691],{},[127,688,690],{"href":689},"\u002Ftools\u002Faudio\u002Felevenlabs","ElevenLabs"," — AI 语音合成,可做配音",{"title":146,"searchDepth":147,"depth":147,"links":693},[694,695,696,697],{"id":16,"depth":147,"text":16},{"id":28,"depth":147,"text":28},{"id":107,"depth":147,"text":107},{"id":119,"depth":147,"text":119},"2024-03-22","Suno 2024年推出 v3 模型,输入文字描述就能生成带人声的完整歌曲,质量接近 Demo 级别,AI 音乐创作从概念验证进入实用阶段。",{},"\u002Fnews\u002Fsuno-v3-ai-music",{"title":568,"description":699},"https:\u002F\u002Fsuno.com","news\u002Fsuno-v3-ai-music",[683,706,707],"AI音乐","创作工具","Hj5D6VoqFlIzUC89Azpj6uFtR3526SYf8Dlz1P8Epn0",1794695479678]