新智元报道
带出Transformer一作的「AI祖师爷」，来中国了！
他的名字，叫Kevin Knight。
今天，这位自然语言处理领域的泰斗级科学家现身广州，为TGC 2026科技增长大会打头阵。
这场大会主题的核心是：AI is Profitable，一开场就把分量拉满了。
同场阵容里，还有Google、OpenAI等全球平台，以及阿里千问、腾讯WorkBuddy等模型与应用团队。
从AI如何运作，到它的发展是否正在加速，Kevin Knight在开场连抛五问。AI for Work，是他给出的下一次飞跃方向之一。这也把现场的关注点带向了企业：模型能力持续突破，哪些日常业务会率先迎来变化？
全场还有一个特殊的AI主角——行业首个Pro Agent，Navos Pro。
由全球领先的AI营销科技公司钛动科技发布，这位新生代Agent配齐了三大「Pro」：Professional、Proactive、Profitable，意味着它是带着KPI上岗的：专业干活、主动找机会，最后还得在真实生意里交出增长结果。
从出海营销起步的钛动，长期帮助品牌开拓海外市场，业务覆盖市场洞察、创意制作、广告投放与效果优化。
在主题演讲中，钛动科技创始人兼CEO李述昊，将下一代Agent归纳为三个Pro：
Pro Agent = Professional + Proactive + Profitable。
专业的工作能不能做好？机会能不能主动发现？行动之后有没有商业回报？Navos Pro先在全球增长营销场景中回答这三道题。
Professional：深入业务，才能把工作做对。
同一款商品，进入不同国家，用户需求、文化语境、内容习惯和媒体环境都可能变化。一段视频为什么跑不动，该换受众还是换内容，都需要专业判断。
李述昊在演讲中强调，基础模型持续变强，专业应用仍需要结合行业训练、专属数据和实际业务经验。
Navos背后的钛极专业大模型，覆盖推理、理解、生成三类任务。
32B问答推理模型在2026年1月SuperCLUE-Mkt广告营销榜单排名第一，27B视频理解模型在2026年7月SUPERCLUE-ADSVU测评中以86.43分位列总榜第二；5B视频生成模型则面向营销视频内容生产。
视频生成就是一个例子。现场展示中，钛极围绕镜头、字幕、情绪、商品构造样本，并采用面向业务结果的强化学习对齐。素材表达与商品卖点、营销目标之间的关系，也进入了模型训练。
专业能力还需要数据支撑。
钛动已服务数十万品牌出海，覆盖全球200多个国家和地区。不同产品、市场、渠道的差异，也给专业模型提出了具体要求：既要理解客户的业务背景，也要吸收专家经验，并用行业基准检查结果。
如今，钛动拥有庞大的数据资产规模。28亿+广告策略、6.1亿+创意、1亿+商品、300万+应用、2.1亿+达人相关数据，并积累了200多个行业基准、10万多个实战案例和9年的出海Know-how。
这些数据还要经过清洗、构造、专家标注，形成专业样本、知识与评价依据。内容是什么、实际表现怎样，两类信息形成双标签，并近实时回流，进入后续判断。
落实到产品上，Navos给出了四大专业能力：数据库、Agent、工作流和互动。
数据库支持趋势、竞品、选品、素材与投放洞察；Agent通过70多个MCP、100多个Skill，提供工具连接和专业任务能力。
工作流把Skill、模板与数字员工组合起来，将选品、素材创作、广告诊断等复杂工作接成执行链路；互动则覆盖意图识别、数据调用与结果反馈，让过程可见、结果可追溯。
今年9月，钛动还牵头组织ECCV 2026 MARS2 Workshop，联合清华、牛津等高校与研究机构，讨论多模态推理、慢思考与Agentic Systems等问题，继续打磨AI理解内容和商业决策的能力。
Proactive：懂了业务，还要主动发现机会。
从Chatbot、Copilot走向Proactive Agent，AI开始参与判断下一步，也开始接过一部分发现问题、推动工作的职责。
前提是，它得先懂这家公司。此次升级的一项重点，是理解企业的「品牌DNA」：品牌属性、调性、使用者偏好，在长期使用中逐步积累。后续分析和策略，才能更贴近企业自己的目标。
用得越久，越懂这家企业；知道它想做什么，才有条件判断哪里值得行动。
钛动把主动性拆成了三个具体场景：盯账户、盯商品、盯素材。
- 盯账户：要关注预算消耗、获客成本与广告回报，发现异常后给出预算、出价等调整建议。
- 盯商品：要跟踪市场热点、竞品变化和潜力商品，帮助企业发现新的生意机会。
- 盯素材：则要识别内容疲劳、爆款衰退和达人表现变化，为卖点调整、素材迭代提供依据。
这套工作从持续感知、发现与诊断，推进到行动建议和反馈学习。具体执行根据用户授权，由Agent或人工继续推进。
感知范围覆盖媒体侧、用户侧与内部系统，目标是7×24持续跟踪。这一环节日均涉及10亿+感知数据、2亿+感知频次。
拿素材巡检来说，发现表现衰退之后，后续还要判断调整卖点、替换素材还是改变投放安排，再把执行结果带回下一轮分析。主动性的价值，就体现在这一串连续工作中。
企业可以更早看到问题和机会，减少一次次提问、拆任务、补背景的工作。Navos希望承担的，也更接近一个懂业务的领航员。
Profitable：把机会变成行动，再用回报检验。
黄仁勋此前说过，有用的AI已经到来，并且已经能够赚钱（Useful AI has arrived, and it is profitable）。
TGC 2026把「AI is Profitable」写进主题，钛动则把这一要求放到了客户的业务结果上。
企业做营销，提出一个好主意只是其中一步。创意要转成素材，素材要接受投放测试，后续还要筛选有效方案、调整策略，再放大表现好的部分。
钛动给出的路径很具体：批量创编、规模测试，筛选有效方案，再迭代策略、放大收益。
一次点击、一次曝光、一次任务完成，都需要放回整段生意里检查价值。结果回来之后，洞察和创意方向也要跟着调整。
今年7月，钛动成为支持ChatGPT广告的技术合作伙伴。公司披露的首批客户测试中，一个出海品牌获客成本显著下降，ROI达到原来的2.5倍。这是特定客户的测试案例，也提供了观察AI商业价值的具体样本。
而要持续改进，AI还得学会利用这些反馈。
李述昊在演讲中介绍了Navos RSI：用Agentic RL推动Agent Harness持续自迭代。
这里的Agentic RL，是面向整段任务轨迹的强化学习：任务怎样规划、工具怎样调用、多智能体怎样协作，都进入评估。
系统根据业务增量、过程质量与效率成本等信号优化策略和Skills，再经过评测、灰度验证，写回工作体系；出现异常时，可以回滚。
奖励和约束也随任务一起设计。ROI、成交额等业务变化，规划是否合理、工具调用是否准确，以及预算、时延和品牌安全，都在演讲列出的考虑范围内。
执行任务、评估结果、调整方法、验证迭代，构成了下一轮工作继续改进的依据。
三个Pro也由此接了起来：专业地处理业务，主动识别并推进机会，再用实际回报检验效果。
从Navos Pro出发，李述昊想回答的，是一个更大的问题：AI已经能写代码了，能不能把生意做大？
他的选择依然明确：「只做结果导向（Result-Driven）的AI for Work」。
这句话的分量，落在「结果导向」四个字上。
现场，李述昊把企业采购AI的变化，概括为从「成本转移」走向「价值创造」。
过去，企业可以按人数、天数购买服务，按席位订阅软件，按调用量使用AI。如今，客户开始追问：这些投入，最终换回了多少业务ROI？
钛动希望交付的，是能执行具体业务、带来可衡量结果的企业生产力。
这种变化已经实打实发生了，Anthropic最近就公布过一个数字：截至8月，他们自家员工调用的交互式Agent运行时长，已经相当于人类员工总工时的6倍之多。
运行时长不能直接换算成生产力，但Agent进入工程、销售、法务等日常工作的趋势，已越来越清晰。
工作交出去之后，企业还要看交付：营销能否带来更好的转化，研发能否缩短交付周期，管理能否减少协调损耗。
这也对应了钛动关注的三类投入：AI for Growth（营销）、AI for Innovation（研发）、AI for Administration（管理）。
在钛动的战略里，重算「企业三费」，就是用AI提高企业客户在营销、研发、管理三类投入上的回报。
围绕这三类投入，钛动估算，AI for Work对应着数万亿美元的潜在市场空间。这是有机会被AI优化的投入规模，里面装着企业每天都在付出的真金白银。
营销，是钛动先接受检验的地方。
产品去哪个市场，卖给什么人，用什么内容，在哪个平台投放，预算怎样分配，每个判断都会影响后面的结果。
反馈来得相对快，策略、执行与投入回报也更容易连起来。
这背后，是三层底层架构，即Agent Harness、专业模型和专业数据组合。
专业数据提供行业经验与评价依据；专业模型处理推理、理解、生成任务；Agent Harness组织任务规划、多智能体分工、工具调用和技能使用，再把评估后的改进带回下一轮执行。
模型训练与工程组织各自解决一部分问题，合起来，才有机会把专业工作持续做下去。
今年8月，钛动与腾讯WorkBuddy达成生态合作，将专业增长能力以Skill形式接入通用办公智能体。围绕营销积累的能力，已经有了向外复用的实际入口。
进入新业务，还需要继续补齐对应的行业数据、工具接口、业务规则和评价标准。
在研发端，营销获得的用户反馈和市场变化，有机会更快进入产品需求，帮助企业判断什么值得开发、什么需要调整。需求研究、产品开发、测试验证如何衔接，最终要检查研发周期、交付质量与投入产出。
在管理端，AI可以进一步探索预算分析、采购协同、进度跟踪和异常发现。反复确认状态、来回协调的时间能减少多少，资源能否安排得更及时，都需要具体衡量。
这两类能力，是钛动面向企业客户的下一步探索。营销中形成的专业模型、数据工程与任务组织经验，为它们提供基础。
大会上，李述昊首次提出AEE，即Agentic Enterprise Effectiveness：
AEE = Agent数量 × 活跃使用程度 × 产出质量。
制造业用OEE观察设备综合效率，AEE则尝试观察企业的Agent工作体系。
企业拥有多少Agent、用得是否频繁、产出能否交付，要放在一起看。
营销回报提高多少，研发推进加快多少，管理损耗减少多少，三笔账各有指标，共同指向企业投入产出的改善。
正如李述昊在演讲结束时所言：「出海营销是AI生产力的第一战场。用AI帮助全球企业赚钱，才是我们的星辰大海」。
Pro Agent已来，AI的演进也到了换标尺的时候。
这个问题，Kevin Knight在开场演讲中已经给出了一个尺度：他把补全句子、网络研究、编写软件等任务放到「自主工作时长」上衡量，讨论AI能否接下更长、更复杂的任务；在他列出的机器人、生物医学研究、AI for Work、娱乐和AI运营一家企业五个方向中，AI for Work看任务是否完成，AI运营一家企业则看利润。从语言到行动，评价AI的标尺正从「能力有多强」转向「工作能否交付、价值能否衡量」。
而钛动的回答，也不只写在Navos Pro一款产品里。
Agent一旦从单点工具变成企业工作体系，考验的就不再只是模型能力，是谁能把专业能力、真实场景和商业回报连成闭环。这个闭环不能只在产品演示里成立，还要在真实业务里跑通；而真实业务，本来就连接着平台、模型、应用、资本与企业客户。
于是，TGC 2026把全球商业平台、模型与应用团队、投资人、出海企业聚到了一起。技术、资金与业务需求在同一个场域里碰面，钛动也把目光投向更多AI创业团队。
同期启动的「Pro Agent Global Challenge全球挑战赛」，面向全球征集优质AI项目，关注Agent演进、AI营销、AI内容入口与算力基础设施等方向，并开放资本、产业与全球增长资源，帮助潜力团队把产品带进真实业务。
从钛动发布Navos Pro，到Kevin Knight讨论「AI能否接下更长的工作」，再到寻找下一个Pro Agent，一条线索逐渐清晰：让AI从语言走向行动，再让行动接受真实业务的检验。
钛动希望连接更多团队，把Useful、Profitable AI带向全球。这份邀请很直接：有好产品，就来真实业务里试一试。
AI的下一张成绩单，企业会用自己的账本来打分。
编辑：桃子
不喜欢
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。
New Ji-won.
"Major Ai" by Transformer, coming to China!
His name is Kevin Knight.
Today, this Thai fighter in the field of natural language processing appeared in Guangzhou to spearhead the TGC 2026 Science and Technology Growth Conference.
At the heart of the theme of the conference is that AI is capable, and the beginning of the session is full.
There are global platforms like Google, Open AI, and models and application teams like Ali, TencentWorkBuddy.
From how AI works to whether its development is accelerating, Kevin Knight opens the door and asks. AI for Work, one of his next leaps. It also leads the focus on the field to the enterprise: model capacity continues to break down and what day-to-day business will lead to change?
There's also a special AI lead - the first Pro Agent, Navos Pro.
With the world’s leading AI marketing company, titanizing technology publish, the new generation Agent has three “Pro”s: Productional, Proactive, Profitable, which means that it takes KPI on board: professional work, proactive opportunity, and, ultimately, growth results in real business.
The titanium movement, starting with marketing at sea, has long helped brands to develop overseas markets, covering market insight, creative production, advertising and optimization.
In the keynote address, the founder of titanium-activated technology and CEO Lee Shu Zheng summarized the next generation of Agent into three Pros:
Pro Agent = Professional + Proactive + Profitable。
Can you do your professional work? Do you have a chance? Any business returns after the operation? Navos Pro first answered these three questions in a global growth marketing scene.
Professional: In order to do business well.
The same commodity, entering different countries, can change user needs, cultural contexts, content habits and media environments. A video needs professional judgement as to why it cannot run and whether it should change audience or content.
In his presentation, Li Shih-ho stressed that the base model continued to grow stronger and that professional applications still needed to be combined with industry training, proprietary data and operational experience.
The titanium professional LLM behind Navos covers three types of tasks: reasoning, understanding and generation.
32B Question and Answer Logic Model ranked first on the SuperCLUE-Mkt Advertising List in January 2026, and 27B Video Interpretation Model ranked second on the SUPERCLUE-ADSVU Master List in July 2026; 5B Video Generation Model was oriented towards marketing video content production.
Video generation is one example. In the field presentation, titanium is concentrated around shot, subtitle, emotional, commercial construction sample and is aligned with enhanced learning oriented towards operational results. The relationship between asset's expression and the point of sale, marketing objectives has also entered model training.
Professional capacity also requires data support.
titanium has served hundreds of thousands of brands, covering more than 200 countries and territories worldwide. Differences in products, markets, channels and specific requirements for professional models include understanding the client ' s business background, drawing on expert experience and using industry benchmarks to check results.
Today, titanium has huge data assets. The $2.8 billion+advertising strategy, 610 million+creative, 100 million+commodity, 3 million+appliance, 210 million+person-related data, and the accumulation of more than 200 industry benchmarks, more than 100,000 active cases and nine years of sailing Know-how.
These data are also cleaned, constructed, labelled by experts, and form the basis of professional sample, knowledge and evaluation. What is the content, what is the actual performance, and the two types of information are double-labeled and re-flowed in near real time into subsequent judgement.
In terms of products, Navos offers four major professional competencies: databases, Agent, workflow and interactions.
The database supports trends, competitions, selections, asset and insights; Agent provides tool connectivity and professional mission capacity through over 70 MCPs and over 100 Skills.
workflow combines Skill, templates and digital staff, linking complex tasks such as selection, asset creations, advertising diagnostics, etc., to the implementation chain; interaction covers intent identification, data call and feedback on results to make the process visible and results traceable.
In September, titanium also led the organization ECCV 2026 MARS2 Workshop, which brings together universities and research institutes such as Tsinghua and Oxford to discuss multimodular reasoning, slow thinking and Agentic Systems, and continue to sharpen AI ' s ability to understand content and business decision-making.
Proactive: Understand business and proactively identify opportunities.
From Chatbot to Copilot to Proactive Agent, AI became involved in the diagnosis of next step and began to take over some of the responsibility for identifying problems and moving things forward.
It must understand the company first. One of the priorities of this upgrade is to understand the company's "brand DNA": brand attributes, modulations, user preferences, which accumulate over time. Follow-up analysis and strategies will bring enterprises closer to their own goals.
The longer it takes, the better to understand the enterprise; knowing what it wants to do, the better to judge where it is worth doing.
The titanium initiative is broken down into three specific scenarios: targeting accounts, commodities, and asset.
- Focusing on accounts: attention should be given to budget consumption, client cost and advertising returns, and adjustments should be made to budgets, bids, etc., when anomalies are detected.
- Focusing on commodities: Following up on market hot spots, competitions and potential commodities to help businesses identify new business opportunities.
- Focusing on asset: The content fatigue, the collapse of the explosive funds and changes in human performance are to be identified to provide a basis for the point of sale adjustment, asset.
The package moves from continuous perception, discovery and diagnosis to recommendations for action and feedback learning. Implementation continues on the basis of user authorization by Agent or manual.
Perceptions cover media, user and internal systems, with a target of 7 x 24 for continuous tracking. Every day, the link involves 1 billion + sensory data, 200 million + sensory frequency.
On the basis of the asset inspection, it was then decided whether to adjust the selling points, replace asset or change the placement arrangements, and then take the results back to the next round of analysis. The value of initiative is reflected in this continuum of work.
Businesses can see problems and opportunities earlier, reducing the number of times they ask questions, dismantle tasks, and supplement the background. Navos would have liked to have been closer to a pilot who understood the business.
Profitable: Turning opportunities into action and applying return tests.
Wong In-hoon has said that useful AI has arrived, and it is profitable.
TGC 2026 included "AI is available" in the theme, while titanium placed this requirement on the business results of clients.
Businesses do marketing and come up with a good idea is just one step. The idea is to convert to asset, which is subject to a drop test, and then to screen for effective programmes, adjust strategies and magnify the good parts.
titanium gives a specific path: batch creation, scale testing, screening for effective programs, iterative strategies, magnifying benefits.
One click, one exposure, one mission completed, all need to be put back in the business to check the value. When the results come back, insight and creativity will be adjusted.
In July this year, titanium became a technical partner in support of ChatGPT advertising. In the first customer tests disclosed by the company, the cost of taking a trip to sea was significantly reduced and ROI was 2.5 times higher. This is a client-specific test case and provides specific sample observations of the commercial value of AI.
For continuous improvement, AI has to learn to use these feedbacks.
In his presentation, Li Shujun introduced Navos RSI: Using Agenic RL to promote the continuation of Agent Harness.
The Agenic RL here is an intensive learning exercise for the entire trajectories of the mission: how the mission is planned, how the tools are mobilized, how the multi-AI agent works.
The system is then evaluated, greyscale validated and returned to the work system based on signal optimization strategies such as business increments, process quality and efficiency costs, and Skills; in the event of an anomaly, it can roll back.
Incentives and constraints are designed along with missions. Business changes such as ROI, turnover, reasonableness of planning, accuracy of tool call, budget, latency and brand security were all considered in the presentation.
The implementation of mandates, the assessment of assessments, the adaptation of methodologies, and the validation of iterative patterns form the basis for continued improvement in the next round.
Three Pros are also involved: professional handling of business, proactive identification and promotion of opportunities, and real return testing of results.
From Navos Pro, what Li Shuxiang would like to answer is a bigger question: can AI be able to write codes, can it be bigger?
His choice remains clear: "Only AI for Work for Results"
The weight of this sentence falls on the words "result orientation".
On the ground, Li Shuxiang sums up the changes in the company's procurement of AI, from “cost transfer” to “value creation”.
In the past, businesses were able to purchase services in number of people, days, subscription to software by seat, and use AI as called. Today, clients are asking: how many businesses did these inputs eventually get back?
titanium is the type of enterprise productivity that can deliver specific operations with measurable results.
This change has taken place, and Anthropic has recently published a number: As of August, the interactive Agent, which was called by their own employees, was running for as long as six times the total number of human employee hours.
While convert cannot become productive directly for long operating hours, the trend of Agent into day-to-day work such as engineering, sales, law enforcement and so forth is becoming clearer.
Once the work is delivered, the enterprise also has to look at delivery: marketing can bring about better transformation, research and development can shorten the delivery cycle and management can reduce coordination losses.
This also corresponds to three types of inputs of interest to titanium: AI for Marketing, AI for Information, AI for Administration.
In the titanium strategy, the recosting of the "business three" is the use of AI to increase the returns of business clients on marketing, research and development, and managing three types of inputs.
With regard to these three types of inputs, titanium is estimated to be the equivalent of trillions of dollars in potential market space for AI for Work. This is the opportunity to be optimized by AI in terms of the size of the input, which contains real money and silver that companies pay every day.
Marketing is where titanium is tested first.
Each judgement affects the results when products go to markets, sell to people, use what, put on which platform, how budgets are allocated.
Feedback comes relatively quickly, and strategies, implementation and returns on inputs are more easily linked.
Behind this is the three-layer bottom structure, namely, Agent Harness, professional models and professional data sets.
Professional data provide the basis for industry experience and evaluation; professional models deal with reasoning, understanding, generation of tasks; Agent Harness organizational mission planning, multi-AI agent division of labour, tool call and use of skills, with subsequent improvements brought back to the next round of implementation.
A combination of model training and engineering organizations solve part of the problem and provide opportunities to sustain professional work.
In August this year, titanium entered into ecological cooperation with TencentWorkBuddy to connect professional growth capacity in the form of Skill to General Office AI agent. The ability to accumulate around marketing already has a physical entry point to reuse.
Entry into new operations will also require continued completion of corresponding industry data, tool interfaces, business rules and evaluation standards.
At the R & D end, user feedback and market changes from marketing offer opportunities for faster entry into product demand to help enterprises judge what is worth developing and what needs to be adjusted. Needs research, product development, testing and validation are linked, and ultimately examine research and development cycles, delivery quality and input outputs.
At the management level, AI could further explore budget analysis, procurement synergies, progress tracking and anomalies. The extent to which the state of repeated confirmation and the time for back-to-back coordination can be reduced and the timeliness of resource allocation needs to be measured.
These two types of capabilities are titanium-oriented next step exploration for business clients. Professional models, data engineering and mission organization experience developed in marketing provide a basis for them.
In the General Assembly, Li Shu Xian first introduced AEE, which is known as Agency Enterprise Effectology:
AEE = Agent quantity x Active use x Output quality.
The manufacturing industry uses OEE to observe integrated efficiency, while AEE attempts to observe the company's Agent work system.
The number of businesses with Agent, the frequency of use and the delivery of outputs must be seen together.
The increase in marketing returns, the acceleration of R & D advances, the reduction in management losses, and the indicators for each of the three accounts collectively point to improvements in the output of the enterprise ' s inputs.
As Li Shujun said at the end of his speech, "Marketing is the first battle of AI productivity. It's our starship to help global businesses make money."
Pro Agent has come, and the evolution of AI is about to change the ruler.
Kevin Knight, in his opening speech, has given a measure of this: He put the completion of the sentence, network research, software development, etc., on the measurement of "self-working hours", and discussed whether AI could take over longer and more complex tasks; in the five directions he listed as robotics, biomedical research, AI for Work, entertainment and AI operating a business, AI for Work and AI operating a business, it looked at profits. From language to action, the evaluation of the AI scale is moving from "how powerful" to "the ability to deliver and measure the value of the job".
And titanium answers are not just in Navos Pro.
Once Agent has moved from a single-point tool to an enterprise work system, the test is no longer just modelling capacity, but who can link professional competence, real scenes and business returns. This closed circle cannot be built only in product demonstrations, but also in real business, which is inherently linked to platforms, models, applications, capital and business clients.
As a result, TGC 2026 brought together global business platforms, models and application teams, investors and offshore enterprises. Technological, financial and business needs meet in the same field, and titanium is directed towards more AI entrepreneurial teams.
The Global Challenge Competition, launched in the same period, is aimed at the global collection of high-quality AI projects, focusing on Agent evolution, AI marketing, AI content access and computing infrastructure, and opening up capital, industry and global growth resources to help potential teams bring products into real business.
From titanium publishNavos Pro to Kevin Knight to discuss whether "AI can take over a longer job" to looking for the next Pro Agent, a clue is becoming clear: let AI move from language to action, then let it be tested for real business.
Titanium wants to connect more teams and bring Useful, Profitable AI around the globe. This invitation is direct: if you have a good product, try it in real business.
AI's next report card, and the enterprise will score its own bill.
Edit: Peaches
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.