AI 时代，只要学得慢，就可以少学很多过时的东西。
作者丨李娜
编辑丨岑峰
一代人有一代人的鸡蛋要领。(雷峰网(公众号：雷峰网))以前是超市门口排队领免费鸡蛋，如今是 AI 社区里蹲邀请码、抢内测名额、攒免费额度。每逢一款新的 Agent 发布，总有人忙着寻找体验入口，研究怎么领积分、怎么薅到更多 Token，唯恐错过下一款改变工作方式的划时代产品。只是这届 AI 产品，有些还没等大家用熟，就已经开始准备退场了。今年 3 月才开启内测的腾讯 QClaw 在9月底宣布将在年底停止运营，从抢邀请码到领积分搬家，前后不过 199 天。类似的变化还有不少：10 月，另一款智能体创建平台元器也发布了停服公告。阿里则将悟空、QoderWork、MuleRun 等产品线重新整合进千问办公。科幻作家威廉·吉布森曾说：“未来已来，只是分布不均。” 在如今的 AI 应用市场里，另一种时间差也开始显现：一款新产品的热潮可能还没传遍所有用户，第一批尝鲜者已经收到了迁移通知。这种不确定性，也不只发生在刚推出几个月的新 Agent 身上。10 月 12 日，曾经红极一时的 AI 陪伴平台 Character.AI 宣布启动美国《破产法》第 11 章程序，寻求通过法院监督出售公司资产。这款仍拥有约 1200 万活跃用户的产品，目前还在正常运营，用户可以继续登录，与熟悉的 AI 角色聊天。但如果最终找不到能够维持平台运营的收购方案，Character.AI 仍将面临关闭。一边是层出不穷的新模型、新 Agent，一边是还没用上多久便开始调整、整合甚至退出的旧产品。在相关讨论区，有网友留下了一句颇有意思的评论：“AI 时代，只要学得慢，就可以少学很多过时的东西。”今年以来，大厂推出 Agent 的速度相当快，但部分产品还没经过半年验证，内部就开始了新一轮调整。我们梳理了腾讯、阿里、字节和百度今年公开推出的 16 个代表性 AI 产品及新产品形态，截至 10 月 19 日，其中 1 产品个已宣布未来停服，另有 5 个经历了产品线整合、停止新购或组织调整。变化最快的是字节 TRAE Work，从 6 月升级为新产品形态，到 8 月团队并入豆包体系，相隔只有 76 天；阿里 3 月推出的悟空，距离三条产品线启动整合也只有 107 天。阿里是一个典型案例。年初，QoderWork、悟空和 MuleRun 分别由不同团队推进，各自侧重桌面执行、企业协作和云端任务。随着能力扩展，办公场景逐渐重叠。财新此前援引阿里内部人士的说法称，早期多团队探索有其必要性，但到了现阶段，已经没有必要继续“重复造轮子”。7 月，三条产品线开始整合；8 月，千问办公正将原本分散的能力汇集到同一个入口。腾讯也在收拢产品线。3 月几乎同时推出的 QClaw 和 WorkBuddy，最初分别侧重微信远程操作电脑和全场景 AI 办公。7 月，QClaw 相关业务及部分团队被调整至 WorkBuddy 所属部门；9 月，腾讯宣布 QClaw 将于年底停运，并提供向 WorkBuddy 迁移的渠道。字节则在 8 月将 TRAE、扣子相关团队整合进豆包体系，原有编程产品线继续发展，办公 Agent 能力向豆包工作汇集。OpenClaw 热潮之后，究竟是桌面 Agent、云端任务，还是企业办公平台能够成为下一代入口，谁都没有确定答案。与其押注一条路线，不如多个团队同时推进，先抢占市场，再看哪款产品能跑出来。但随着第一轮产品落地，算力成本成了大问题， 维持多个没有明确商业模式的 Agent 产品并行跑，是对 GPU 资源的极大浪费。把零散的入口统一，本质上是为了集中算力、做高日活，向 B 端讲一个更完整的商业故事。除了研发维护、模型调用等成本，底层模型的迭代也在改变独立 Agent 的生存空间。SpaceX中国区大使、同时也是AI创业者的Kin在接受 AI 科技评论采访时谈到，Agent 的更替速度一直很快，并不是今年才突然加速。GPT、Claude 等基础模型不断升级，一些缺少独特优势的 Agent，原有功能可能很快就被新模型或综合性产品覆盖。Kin 提到，他所在的数十个 AI 微信社群里，只要新产品有亮点，大家仍然愿意第一时间尝试。真正的问题在尝鲜之后：一款 Agent 能否提供足够的长期价值，让用户持续打开，而不是体验几次就放在一边。对腾讯、阿里这样的公司来说，结束一个独立产品，并不意味着放弃原来的业务。QClaw 的部分能力可以由 WorkBuddy 承接，悟空、QoderWork 等产品也能被整合进千问办公。大厂有足够多的业务和团队，可以重新分配资源，继续寻找下一代 AI 入口。Character.AI 曾是 AI 陪伴赛道最受关注的公司之一。2021 年，两位前谷歌研究员 Noam Shazeer 和 Daniel De Freitas 创办了这家公司，凭借用户可以自由创建角色、与 AI 进行长期对话的产品形态迅速走红。到 2024 年，其月活跃用户已达到约 2000 万，估值一度达到 10 亿美元。这家曾经的明星 AI 独角兽并不缺用户。根据路透社援引的破产资料，其目前仍有约 1200 万活跃用户，但每月亏损高达 800 万至 900 万美元。2025 年，公司每月的算力支出仍在约 1000 万美元，而订阅业务月收入的峰值仅为约 280 万美元。即便加上广告，收入也始终难以支撑高昂的运营成本。除了算力开支，Character.AI 还面临创始团队离开、技术竞争及法律诉讼等多重压力。曾经让它迅速走红的高黏性用户群，最终没能支撑起一门可持续的生意。如今，公司只能寄希望于资产出售，为平台寻找新的出路。相比之下，大厂内部的 Agent 赛马，即使某款产品退出，原有技术和用户仍有机会进入新的产品体系；Character.AI 面临的却是整个平台能否继续存在的问题。两种处境背后，有一个容易被忽略的事实：一款 AI 产品即使还有人使用，也不代表公司一定有理由，或者有能力继续经营它。对 QClaw 用户来说，至少还有 WorkBuddy 可以迁往；Character.AI 的用户则需要等待破产程序的最终结果，才能知道自己熟悉的平台会不会继续存在。9 月 24 日，QClaw 宣布停运，备份和迁移窗口保留至次年 3 月。腾讯元器则将于 10 月 29 日停止创建新智能体及分发 API Key，11 月 正式停服，数据导出期限延续至次年1月。腾讯元器公布的 ADP 迁移指引显示，新版元器的知识库、工作流和 Widget 支持导出后直接导入，但智能体设置、自定义工具仍需手动配置；旧版工作流则需要参照原有逻辑重新搭建。由于两个平台支持的模型和插件并不完全相同，迁移后的效果也需要重新验证。部分产品还设置了数据保存期限。今年 7 月，豆包宣布下线用户自建智能体功能，允许用户在过渡期内查看、保存相关信息及历史对话。按照公告，10 月 15 日后，用户将无法再通过豆包查看或恢复这些数据。对于长期使用 Agent 的用户来说，产品更替并不一定意味着从头再来。AI 发烧友 Likka 从今年 4 月开始付费使用 Cursor。起初吸引他的，是 Cursor 能在同一个开发环境中使用不同模型，也方便随时查看和修改代码。但随着产品和服务发生调整，他开始尝试寻找替代工具。项目文件、AGENTS.md 和 Skills 都可以打包带走，真正让他难以切换的是使用习惯。他尝试过 Codex、Claude Code 等产品，目前却仍以 Cursor 为主力。“换到新的一个平台，你会不习惯，这可能也是一种沉没成本吧。”他甚至曾尝试搭建一套不依赖单一厂商的本地中转系统，希望避免因为平台停服或账号受限而丢失积累。但运行一段时间后，各类接口、配置和维护问题接踵而至，最终还是放弃了。另一位受访者 Kin 则更愿意主动承担迁移成本。他习惯将项目上下文、Skills 等资料备份到 GitHub 和云空间，并为常用工具准备替代方案。过去在 Codex、Claude Code 中完成的部分工作，如今已经被他转移到了 Grok Bot，出门在外也能通过手机和 iPad 继续处理。在 Kin 看来，只要新产品能像Grok Bot一样带来足够明显的效率提升，就值得尝试。两人的经历呈现了 Agent 用户面对产品更替时的不同选择：有人不断迁往更好的工具，有人仍然留在熟悉的平台。但无论哪一种，迁移成本都不只是一份可以导出的数据。从年初 OpenClaw 掀起热潮，到如今大厂陆续整合产品线，甚至连拥有千万用户的 Character.AI 也走到了破产重组这一步，AI 产品的去留，似乎比人们想象中更难预测。厂商忙着寻找下一个入口，用户一边追着新产品跑，一边还得考虑此前积累的工作流该怎么搬走。但至少在受访者看来，这些变化并没有消磨他们对新技术的热情。Likka 把眼下的 Agent 竞争比作一场长跑：有人往这边跑，有人往那边跑，谁也不知道通往下一阶段的那扇门究竟在哪里，但总不能因为不知道，就停下来不跑了。新产品还在不断冒出来。下一扇门在哪里没人知道，但大家显然还想继续找下去。
可独家畅览：
专家演讲PPT
大会报告全文
热门论文解读
学术新星访谈
扫描上方二维码
或点击「阅读原文」关注专区。
雷峰网原创文章，未经授权禁止转载。详情见转载须知。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
In the AI era, if you learn slowly, you can learn a lot less.
The author, Leena.
Edit Peak
A generation has an egg to collect. Formerly at the supermarket entrance lined up for free eggs, it's now the AI community's invitation number, insider spot, and free quota. Whenever there is a new Agent publish, there are people who are busy looking for an entrance to experience, looking at how to get points, how to get more Token, lest they miss the next landmark product that changes the way they work. It's just this AI product, some of which are ready to leave before you're ready. Tencent QClaw, which only opened the internal survey in March this year, announced at the end of September that it would stop operating at the end of the year, moving from the invitation code to the lead, but only 199 days later. Similar changes took place: in October, another AI agent created platform metaphone and publish suspended service announcements. Ali has re-integrated the product lines of Goku, QoderWork, MuleRun and so on. The science fiction writer William Gibson once said, "The future has come, but it is not evenly distributed." In today's AI application market, another time difference is beginning to appear: the heat of a new product may not reach all users, and the first tasters have been notified of the move. This uncertainty does not just happen to the new Agent, which has been in place for months. On 12 October, Character.AI, the once-redest-ever AI Accompaniment Platform, announced the initiation of Chapter 11 of the United States Bankruptcy Code, seeking to oversee the sale of corporate assets through court supervision. This product, which still has about 12 million active users, is now operational and users can continue to log in and chat with familiar AI characters. However, Character.AI will still face closure if it does not eventually find an acquisition programme that will sustain the operation of the platform. On the one hand, new models, new Agents, and on the other hand, older products that have begun to adjust, integrate and even exit before it takes too long. In the relevant discussion area, some netizens left an interesting comment: “Ai age, if you learn slowly, you can learn a lot less of the obsolete.” Since the beginning of the year, large plants have been launching Agent quite quickly, but some products have not been certified for six months, and a new round of internal adjustments has begun. We combed 16 representative AI products and new product patterns that were publicly launched this year by Tencent, Ali, byte and Baidu. As of 19 October, 1 product had declared its future decomposition, and 5 had undergone product line integration, new purchases or organizational adjustments. The fastest change was in byte TRAE Work, which was upgraded from a new product pattern in June to a team integrated into Doubao in August, which was only 76 days away; and Ali, which was launched in March, was only 107 days away from the initial integration of three product lines. Ali is a typical case. At the beginning of the year, QoderWork, Goku and MuleRun were driven by different teams, each focusing on desktop execution, corporate collaboration and cloud tasks. As capacity expands, the office landscape gradually overlaps. Zhong Zhong had previously quoted inside Ali as saying that there was a need for early multi-team exploration, but at this stage it was no longer necessary to continue “repeating the wheel”. In July, the integration of the three product lines began; in August, the office was bringing together previously dispersed capacities at the same entrance. Tencent is also closing the product line. QClaw and WorkBuddy, launched almost simultaneously in March, initially focused on WeChat remotely operated computers and the full-scale AI office, respectively. In July, the QClaw related operations and some of the team were repositioned to the WorkBuddy department. In September, Tencent announced that QClaw would stop at the end of the year and provide a channel for relocation to WorkBuddy. byte, for its part, integrated the TRAE and Coze teams into the Doubao system in August. The original programming product line continued to develop, and the Agent capacity was brought together for Doubao work. Neither the desktop Agent, the cloud task, nor the enterprise office platform can become the next-generation portal after the OpenClaw boom, and no one is sure. Rather than bet on a course, multiple teams would move forward simultaneously, taking over the market and seeing which product would come out. But as the first round of products land, the cost of calculating power becomes a big problem, and maintaining multiple Agent products that do not have a clear business model running in parallel is a huge waste of GPU resources. The purpose of unifying the fragmented entrances is essentially to focus on computing, high-speed work, and tell B a more complete business story. In addition to the costs of R & D maintenance, model calls and so on, the laterality of the bottom model is changing the living space of independent Agent. In an interview with the AI Science and Technology Review, Kin, Ambassador of the China District of SpaceX and also an AI entrepreneur, said that the turnover of Agent has been fast, not suddenly accelerating this year. Basic models such as GPT and Claude are constantly being upgraded, and some Agent, lacking unique advantages, may soon be covered by new models or integrated products. Kin mentioned that in his dozens of AI WeChat communities, as long as the new product has a bright spot, you are willing to try it first. The real question is after a taste: can an Agent provide enough long-term value to keep users open instead of being sidelined several times? For companies such as Tencent and Ali, ending an independent product does not mean abandoning the original business. Some of QClaw's capabilities can be taken over by WorkBuddy, and products like Goku and QoderWork can be integrated into the office. The large plant has enough operations and teams to reallocate resources and continue to look for the next generation of AI portals. Character.AI was one of the companies most interested in accompanying the track. In 2021, the company was founded by two former Google researchers, Noam Shazeer and Daniel De Freitas, and was quickly replete with product patterns that allowed users to freely create roles and engage in long-term dialogue with AI. By 2024, its monthly active users had reached approximately 20 million, with a one-time valuation of $1 billion. This once star AI unicorn is not short of users. According to the bankruptcy information cited by Reuters, it still has about 12 million active users, but losses of between $8 million and $9 million per month. In 2025, the firm ' s monthly computing expenditure was still about $10 million, while the peak of monthly revenue from subscriptions was only about $2.8 million. Even when advertising is added, it is always difficult to sustain high operating costs. In addition to spending money, Character.AI faces multiple pressures such as the departure of the founding team, technical competition and legal proceedings. The high-compulsive user base that had allowed it to move quickly failed to sustain a sustainable business. Today, companies can only hope to sell their assets and find new ways for the platform. By contrast, Agent racing within a large plant, even if a certain product exits, still offers the technology and users the opportunity to access a new product system; Character.AI faces the question of the continued existence of the entire platform. Two. Behind the situation is the easily overlooked fact that an AI product, even if used by others, does not mean that the company must be justified or able to continue to operate it. For QClaw users, there is at least WorldBuddy that can be moved; and Character.AI users need to wait for the final outcome of the insolvency proceedings to know whether their familiar platform will continue to exist. On September 24th, the QClaw announced the stoppage and the backup and migration window was kept until March of the following year. Tencent dollars will stop creating a new AI agent and distribute API Key on October 29th and officially suspend the service in November, with the data export deadline extending until January of the following year. The ADP migration guide published by the Tencent dollar instrument shows that the knowledge base, workflow and Widget of the new version support direct import after export, but AI agent settings, custom tools still need to be configured manually; the old version workflow needs to be re-engineered in line with the original logic. Since the models and plugins supported by the two platforms are not identical, the effects of the migration also need to be re-tested. Some products also have data retention periods. In July of this year, Doubao announced that offline users would build their own AI agent functionality, allowing users to view and preserve relevant information and historical dialogue during the transition period. According to the bulletin, after October 15, users will no longer be able to view or restore the data through Doubao. For users who have long-term Agent, product replacement does not necessarily mean starting from scratch. AI fever friend Likka has been using Cursor since April this year. At first, Cursor was able to use different models in the same development environment and to easily view and modify codes. However, as products and services changed, he began to try to find alternative tools. Project documents, AGENTS.md and Skills can all be packed and taken away, and what really makes it difficult for him is to use habits. He has tried products such as Codex, Claude Code, and is still focused on Cursor. "In a new platform, you're not used to it. It might be a sunk cost." He has even tried to build a local transit system that does not depend on a single manufacturer in the hope of avoiding loss of accumulation due to platform shutdowns or account limitations. After some time of operation, however, the various interfaces, configurations and maintenance problems followed and were eventually abandoned. Another interviewee, Kin, prefers to take up the costs of relocation. He is used to back-up items context and Skills to GitHub and cloud space and to preparing alternatives for commonly used tools. Part of the work done in Cordex, Claude Code has now been transferred to Grok Bot, and can be handled outside via mobile phones and iPad. In Kin, it seems worth trying as long as the new product can lead to sufficiently significant efficiency gains like Grok Bot. The experiences of both people present different options for Agent users in the face of product turnover: people are moving to better tools and people remain on familiar platforms. But in either case, the costs of migration are more than just an output. From the beginning of the year, OpenClaw was the hottest, to the day when large plants integrated their product lines, even Character.AI, with millions of users, came to the point of bankruptcy and reorganization, and the stay of the AI product seemed more difficult to predict. The manufacturers were busy looking for the next entrance, and the users were chasing new products while considering how to move the workflow that had accumulated. At least in the view of interviewees, however, these changes did not diminish their enthusiasm for new technologies. Likka compared the current Agent competition to a long run: people run this way, people run that way, no one knows where the door to the next stage is, but can't stop because they don't know. New products are coming out. No one knows where the next door is, but it is clear that everyone wants to keep looking.
It's an exclusive view:
Expert presentation PPT
General Assembly
The reading of popular papers
Academic New Star Interview
Scan 2D above
This post is part of our special coverage Syria Protests 2011.
It's an original Ray Peak article, which is forbidden to reproduce without authorization. For further information, see the information below.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.