AI Search Evolves, Manus Breaks Out, Agents Rise, and Education Changes: Guokr Founder 姬十三’s 2025 Tech Outlook
Summary
AI search is not simply replacing traditional search; it is restructuring “finding links” into “getting the full picture directly.” 姬十三 likens his first experience with Perplexity to using a touchscreen phone for the first time: “Surprising, yet so natural.” In practice, the two divide the work this way: Baidu and Google still handle URLs, site-specific information, and simple facts, while complex questions go to AI. Traditional search has not yet declined in absolute volume; the likely end state is a hybrid of AI search and traditional search.
DeepSeek R1-style “deep thinking plus web search” has pushed search products into the territory of real-time research tools. In a test of “how to treat depression,” traditional search returned information to read, while DeepSeek and Quark delivered explanations close to complete articles. Tracking the Russia-Ukraine battlefield, users can also ask AI to compare coverage from media in different countries and with different value systems. Input is entering “low-altitude flight mode”: “Just type whatever comes to mind,” while the model handles intent clarification and summarization.
Google and Baidu’s core problem is not catching up technologically, but cannibalizing their own search profit pools. Koji believes that once a complete AI answer occupies the top of the page, users will no longer need to scroll past keyword ads. This is closer to “a decisive battle over business models.” Perplexity has charged for memberships since Day One and begun testing ads; whether it can make AI-native advertising work will determine whether incumbents can escape the innovator’s dilemma.
AI SEO is becoming a new market in which brands compete for weight in model answers, while making the line between advertising and fact harder to see. Next Ad generates ads automatically from conversational context, while Profound helps brands gain exposure and more positive treatment in model responses. After DeepSeek took off, Koji also spoke with many Chinese brands. His question is worth preserving: “Will this process make the world better?” 姬十三 hopes ads and organic results will eventually be clearly labeled, while noting that organic search results already carry marketing intent: “If the water is too clear, there will be no fish.”
Search entry points will continue to fragment, and distribution plus an existing user base may be more commercially valuable than standalone model capability. 姬十三 cites an a16z ranking placing Quark sixth globally among AI mobile apps by MAU. Koji believes Quark’s longstanding “search plus tools” strategy, without the baggage of traditional advertising, makes its shift toward AI search and eventually agents smoother. 姬十三 expects phones to remain the center of computing over the next 5 years, while smart speakers, glasses, cars, and Xiaohongshu’s “Dian Dian” take over interactions by scenario.
Manus offers a powerful but still inconclusive case study for general-purpose agents. At a briefing the day after launch, Haiku Cloud disclosed an internal comparison claiming that Manus beat roughly three-quarters of the agent startups in Y Combinator’s Winter 2025 batch on corresponding vertical tasks, supporting its belief in “Less structure, more intelligence.” But 姬十三 preserves the key risk: if each of 8 steps has a chance of getting stuck, can the final deliverable still be complete?
DeepSeek has cut costs, improved capability, and repaired confidence among founders and investors, potentially making 2025 the year AI applications explode in number. Pine is using a voice agent to dispute bills on behalf of U.S. consumers, while Vozo reached $1M in annualized revenue within 6 months of founding by synchronizing lip movements, timing, and emotion in video translation. 姬十三 therefore expects both software and hardware applications to explode, alongside a “re-rating of Chinese assets.”
AI education may shift its priority from the “head” to the “hands and heart,” a change with more long-term meaning than making children memorize more knowledge. 姬十三’s framework is that human memory and intelligence are unlikely to surpass AI, while hands-on creation, organizing teams, bringing things from 0 to 1, and maintaining healthy emotions will appreciate. Koji firmly chooses the “happy and confident” philosophy over classes that are “strict and hypercompetitive”; 姬十三 says he will encourage children to use future tools early and become “an AI-native generation.”
Deep dive
1. AI Search First Rewrites the Division of Labor Rather Than Immediately Erasing Traditional Search
姬十三 recalls his first experience with Perplexity as “like using a touchscreen phone for the first time”: the product form was unfamiliar, which made it surprising, but it was also closer to what users expected from search results, making it “so natural.” Koji estimates that 1.5 years, or even 2 years, had passed between his first use and the recording. During that period, the product’s structure and content continued to evolve, but the biggest change was that users had started treating it as an everyday tool close at hand.
The two draw nearly identical boundaries between use cases: Baidu and Google are faster for finding a specific URL, locating information on a known site, or answering simple facts such as “Who wrote this book?” AI search is better when a question is just opening up and requires a full picture or an extended argument. 姬十三 retains one practical constraint: “Overall, it still takes quite a while.”
姬十三 does not agree that traditional search has disappeared: its share or absolute volume “has not declined so far.” The future is more likely to combine AI answers with traditional results, with AI becoming standard across all search products. But the advertising burden in the traditional experience is real. Koji quotes a user’s description: “You have to climb over 3 mountains of ads before you can find the actual answer.”
2. Reasoning Plus Web Search Upgrades Search into Real-Time Research
After DeepSeek layered R1’s reasoning model onto web search, 姬十三 saw a clear improvement in intent understanding and the overall experience. When he was looking for alternatives because of service instability, Quark struck him as “extremely fast and extremely impressive.” Kimi uses K1.5, while Doubao and Quark rely on their own reasoning-model stacks, so they did not connect to DeepSeek R1.
Koji ran a side-by-side test with “how to treat depression.” Baidu supplied multiple pieces of information from across the web; DeepSeek and Quark organized the background, recommendations, and cautions into an article, giving readers an initial but comprehensive understanding in one pass. “Deep thinking plus web search” therefore does more than shorten an information list: it takes over reading, abstraction, and summarization.
姬十三’s stronger example is breaking international news. After a White House argument, he wanted to know what had changed on the Russia-Ukraine battlefield over the previous 2 days. AI could aggregate the latest reports and show the positions of media from different countries and with different value systems. Traditional search can also find the news, but organizing the material into a comparative framework is “really cumbersome.”
3. The Prompt Barrier Is Falling, but Good Answers Still Require Active Pushback
Koji observes that AI search lets users ask vague, colloquial questions. 姬十三 adds that before DeepSeek, there were still many courses equating “learning AI” with learning to write prompts. Now models can fill in what users left unsaid or ask follow-up questions to clarify, lowering the expressive barrier for ordinary users.
Koji calls his own state “low-altitude flight mode”: his brain is not running at full speed, so he throws whatever comes to mind at AI, yet often feels that it “understands what I’m thinking pretty well.” He also assumes by default that AI is an all-knowing, all-powerful, highly intelligent, highly empathetic partner. 姬十三 does not set expectations infinitely high; he asks AI to try first whenever something comes up, then decides whether to use the result.
The lesson from working together is that models tend to flatter users and are not inclined to volunteer surprising answers. When brainstorming, 姬十三 asks AI to “praise enthusiastically and criticize harshly” at the same time. When generating titles, he asks for 10 first, then says “another 10” if they are not good enough, repeating until something surprising appears. Koji calls it “an AI workhorse that never complains,” but prompts and follow-up questions still set the ceiling.
4. Google and Baidu’s Real Barrier Is the Income Statement
Koji places both companies in a classic innovator’s dilemma. Their commercial foundations are traditional search and keyword-linked advertising, which created value in the tens or hundreds of billions. If they put a complete AI answer at the top of the page, users will no longer scroll down to see ads, meaning product evolution would directly undermine the companies’ foundations. Their current responses therefore look “awkward, even a little clumsy.”
Koji believes the problem is not that large companies cannot build AI search, but that the difficulty lies in finding a new monetization model. If advertising ultimately works, Google and Baidu can migrate more easily. If AI search can rely only on paid subscriptions over the long term, they will have to make a real choice between the new experience and old revenue.
5. AI SEO Builds New Weights and Creates New Trust Costs
Perplexity put high-frequency and advanced search behind a paid membership from Day One, unlike the historically free Google and Baidu. It subsequently began recommending ads based on search intent. Next Ad is exploring AI-native advertising, allowing chatbots such as iAsk to generate ad content in real time from context, user intent, and creative assets; the ads can appear beside the answer or potentially inside the body.
The AI SEO described by 姬十三 is no longer a contest for PageRank; it is an effort to influence how models evaluate brands. Brands such as Coca-Cola will want models to provide descriptions that are more positive, more complete, and more favorable than competitors’. Profound, backed by Peter Thiel, serves multiple consumer and financial companies with the explicit goal of increasing brand visibility and the quality of brand expression in AI answers.
After DeepSeek took off, Koji spoke with many Chinese brands. Business owners were especially focused on “how to appear more often in DeepSeek’s answers.” This means models are establishing a new distribution of value weights, and influencing the answers themselves will become an independent marketing budget and startup market.
Koji is unsure whether it would make the world better if post-SEO answers and ads were placed together “more seamlessly and in a way that is harder to distinguish.” 姬十三 wants major model companies to label ads and organic results clearly, while noting that traditional organic rankings also carry marketing intent and content costs: “If the water is too clear, there will be no fish.” Even a genuinely ad-free VIP version would not automatically eliminate bias.
6. Entry Points Will Diffuse into Scenarios, and Hardware Survives Only When It Carries Real Life
Looking 5 years ahead, 姬十三 does not expect phones to disappear: “The pocket itself is not going away.” Phones will remain the most important computing center; smart speakers may own the home, glasses the outdoors, and cars the driving scenario. These interactions may sometimes look like search and may sometimes become forms no longer called search.
Continuous conversation on the Xiaoai Speaker Pro means children do not need to wake it again every turn. Huawei’s assisted driving systems and overseas FSD give vehicles multimodal context through video, voice, and other inputs. 姬十三’s causal chain is straightforward: the richer the situational information a device captures, the more complete the “input file” and context search receives, and the more likely the experience is to surpass an isolated query on a phone.
WeChat, Xiaohongshu, and Douyin have already eaten into traditional search entry points. Xiaohongshu has also launched “Dian Dian,” a standalone app connected to a reasoning model. Neither speaker has developed a habit of opening it, but it could serve users who already rely primarily on Xiaohongshu search. Koji cites an a16z ranking placing Quark sixth globally among AI mobile apps by MAU; its “search plus tools” route also avoids Baidu-style advertising baggage.
The novelty of hardware still fades quickly. 姬十三 bought 2 pairs of Meta Ray-Ban glasses, and both eventually sat unused: “A middle-aged person’s life apparently doesn’t have that many exciting moments.” But when he used Vision Pro to capture the space of his old home, he imagined re-entering that memory with his children 20 years later. Spatial video felt “more emotionally rich” than 2D photos, and the future mode of viewing struck him immediately.
7. AI Is First Eating Learning and Document Workflows
Both speakers use Monica frequently because one entry point can switch among multiple models. Koji also has 8 models answer the same question simultaneously for comparison, and uses Claude through Monica when he does not have a separate subscription. If one model stalls or produces a poor answer, he can switch immediately without maintaining a separate workflow for every provider.
Koji received usage credits for ChatGPT Deep Research, while 姬十三 admitted he “couldn’t bring himself to pay $200.” Before recording, Koji asked it to research 姬十三’s recent activities and recommend interview topics. A regular chatbot could also do the job, but Deep Research tends to deliver “better, more comprehensively,” making it suitable for tasks where users want to be lazy but still demand coverage.
姬十三 relies on Notion AI because it is embedded in the workflow: after writing a draft in low-altitude mode, he can polish it directly; midway through an article, he can ask for a structure; when a single transition feels awkward, he can use the context to generate 5 versions. It can also draw on historical documents to reuse knowledge and style. Recraft generates the cover and illustrations for “Crossroads” through a fixed filter, and its daily allowance of 50 free images covers his routine needs.
姬十三 believes AI’s biggest effect on him is not any isolated productivity task, but the speed at which he learns unfamiliar subjects. When an unfamiliar structure appeared at a Xiaomi Auto product launch, his team could immediately ask AI to explain it for a beginner. Guokr has not yet found a breakthrough internal pattern, but he observes that “the earliest AI users are often the people with stronger creative abilities.” As long as the output remains high quality, relying on AI represents stronger learning and creative capability.
8. Manus’s First-Principles Value Is Compressing Multistep Tasks into a Single Delegation
What impressed 姬十三 most remains the launch demo: upload a ZIP containing dozens of résumés, and the agent processes the files on its own before providing relevant summaries. Tasks that previously required users to break the work into multiple steps and invoke multiple tools can now be completed “in one step.” That is why he still called the product impressive despite seeing flaws in the first version.
Koji disclosed that he is an adviser to Manus and an early angel investor in a company on Xiaohongshu. He experienced the product’s journey from idea to private testing to launch. Manus shared the agent’s execution process at the pixel level, which 姬十三 saw as a growth insight from product managers and an elegant viral feature, rather than a marketing sin deserving condemnation.
The real problem was that distribution ran out of control. After DeepSeek, self-media accounts rushed to tell the story of “Chinese AI achieving world-class growth,” and invitation codes quickly became a traffic event. 姬十三 speculates that the team originally intended the invite-only system to control the early user profile and expectations. Once demand exploded, the mix of users and use cases created a great deal of noise for the product team.
9. General-Purpose Agents Have a High Ceiling, but Reliability Decays Along the Task Chain
姬十三 does not avoid the most valuable criticism: are general-purpose agents inherently inferior to vertical agents on specific tasks? Vertical agents can design workflows and optimize specifically around important scenarios. If every vertical can be handled more deeply, the question of where a general-purpose product’s highest-frequency scenarios lie still requires practical proof.
At a small briefing the day after launch, Haiku Cloud disclosed an internal comparison. It tested Manus against every agent startup funded by Y Combinator’s Winter 2025 batch on their corresponding vertical tasks, and Manus beat three-quarters of them. That result became the team’s core evidence for believing in general intelligence.
The team’s formulation is “Less structure, more intelligence.” Its contrarian judgment resembles DeepSeek’s: rather than continuously stacking manually designed workflows, trust the model to organize tasks on its own. Which route will work is “hard to derive in theory; only practice can reveal the truth.”
The risk is that if a task contains 8 steps and each step has some probability of getting stuck, final completeness can fall rapidly. 姬十三 therefore calls Manus “the most cutting-edge application product this year,” apart from models such as DeepSeek, while identifying multistep success rates as a core problem that must be fixed. Koji also said the product was almost named “Dudu,” with the slogan “Just Do Do It,” but the name was abandoned because the dodo was thought to be nearly or already extinct.
10. AI Education Must Reorder the “Hands, Head, and Heart” Rather Than Run a Knowledge Competition
姬十三 admits he does not have a ready-made answer: everyone believes AI will change children’s futures, but school education has not yet made the corresponding turn. His principle is clear, however: new tools cannot be avoided. Once his child is old enough, he will encourage the use of ChatGPT, Doubao, or Kimi; “he will become part of an AI-native generation.”
姬十三 breaks traditional education into the “hands, head, and heart.” After AI, the priority of the “head” may decline because human memory and intelligence are unlikely to surpass machines. The “hands” mean creating things, organizing teams, and bringing something from 0 to 1. The “heart” means maintaining healthy emotions, handling crises, and embracing unfamiliar change. Both will become more important.
Koji believes the mobile internet has already exposed adults to psychological strain through fragmentation and information overload. Children without a gradual adaptation process may face even sharper contradictions. His practical choice is therefore specific: between a famous extracurricular program emphasizing strictness, homework, and competition, and an institution whose goal is “happiness and confidence,” he firmly chose the latter, because sustained positive feedback gives children more reason to believe, “I can do this well.”
11. DeepSeek Cuts Costs, Raises Capability, and Repairs Market Confidence at the Same Time
姬十三 sees DeepSeek as a turning point for 2025: lower costs and higher capability will directly unlock applications while indirectly affecting founders, capital, and U.S.-China relations. He expects “the vast majority” of the impact to be positive, but acknowledges that the world may also become more divided and polarized; technology diffusion will not automatically eliminate political friction.
What moved 姬十三 was another parallel chain: DeepSeek benefited from the open-source community and then returned its成果 to that community. Even amid tense national relations, “the transmission of the spark of knowledge” continues climbing toward the summit; knowledge still appears to move across borders and prejudice. This is a longer-term infrastructure effect beyond the application boom.
姬十三 believes DeepSeek may have a deeper impact than was visible at the time: it lowers development costs, raises Chinese founders’ and investors’ confidence, directs more capital into early-stage projects, and drives a “re-rating of Chinese assets.” He places DeepSeek and Ne Zha alongside one another as signals that confidence had returned since the start of the year.
12. Agents and Multimodal Systems Have Entered a Revenue-Ready Application Window
姬十三 believes agents are entering a mature window as costs fall, reasoning chains improve, and open-source efforts multiply. His analogy is Li Auto’s ONE: even if ONE quickly stops being Li Auto’s main product, it first proved that there was real demand for “a refrigerator and a television inside a large vehicle.” Manus may likewise first show the market that general-purpose agents meet a real need, with later products completing the iteration.
Pine is turning audio capability into a concrete agent service. When U.S. consumers believe their credit-card bills are unreasonable, Pine calls on their behalf to seek a refund or price adjustment. The idea was not new, but older models spoke “stiffly or awkwardly.” Once capability crossed the threshold, the same need moved from a concept to a deliverable product.
Vozo, as described by Koji, demonstrates the speed of multimodal commercialization. The team reached $1M in annualized revenue within 6 months of founding, moving from video “remixing” into translation. The challenge is not merely translating text, but synchronizing the lip movements, the duration of each phrase, and emotional shifts in languages such as Japanese with the original video. That requires both foundational breakthroughs and extensive engineering refinement.
姬十三 therefore expects AI applications to “surge” in number in 2025, with both software and hardware potentially entering an explosive phase. Model capability still matters, but the more tradable shift is that needs everyone could imagine but few could execute well are beginning to turn into real products and revenue across verticals.