Vol.203 How Big-Tech Executives and Freelancers Use AI Differently: Talking AI with 陈亮 of Ant Group and “水哥”王昱珩

Summary

• The inflection point came in 2H25, not when ChatGPT launched. Ant Group Chief Marketing Officer 陈亮’s view from inside the market: enthusiasm for ChatGPT quickly faded, until DeepSeek exploded in early 2025, leaving China “both excited and panicked”—excited that “it wasn’t only America’s capabilities that were strong; China could do it too,” and panicked when big tech realized the gap with R1 was “still very large.” The result: “fighting to catch up in the panic,” the “progress bar suddenly accelerated,” and by 2H25, “so many things had to be redone with AI.”
• The episode’s core framework: AI gets you to 60-70 points; humans hold the line at 90-95. 陈亮’s logic is that AI produces standard answers—“because you can ask, others can ask too”—while the difference in his market work lies in the final few points: judgment and innovation. “Many of the complex factors behind a decision are unknown to AI,” and some, such as a company’s operating situation, may be information one would not want to feed it; “I can’t distill the emotions and pain I’ve experienced in the past and give them to it, yet those things are often what ultimately enable me to make a decision.”
• Organizations are shifting from pyramids to trees: AI is gradually replacing basic roles and shrinking the base, while the canopy—the decision-making layer—is expanding; 水哥 even predicts that certain moments could produce an “inverted pyramid.” This year’s instant-commerce war offers one reason a CEO cannot simply be replaced by AI: under a media-buying efficiency logic—if the other side gets three times the effect for one dollar, you stop fighting—AI would never launch the battle. “Things beyond expectations—that is what surprise means.”
• Ant’s AI product strategy is differentiation, not imitation. Lingguang launched only in November because its confidence came from strengthening “structured output plus aesthetic presentation” as a differentiator during beta testing, “making the thing beautiful through mathematics”—with the same standards as elegant code: simplicity, logic, and structure. It also runs against mainstream internet logic: improving productivity rather than capturing user time. 陈亮 jokes that this is “the internet’s earliest form of originalist purism.” On the medical side, Afu has AI ask the questions, addressing a key adoption bottleneck: “many people simply don’t know how to ask.”
• The investment case for AI inclusion is about raising the floor. An exceptional GI endoscopist can do 10 procedures a day at most, at a 90-plus level, while a large number of doctors are only at 50 or 60. An AI assistant can pull up the average. 陈亮 cites the line that moved him most: “An AI doctor may not be able to determine the ceiling of medicine today, but it can dramatically raise the floor”—like a searchlight illuminating places resources cannot reach.
• Hallucinations remain a hard risk. When 陈亮 asked DeepSeek about the outlook for data as a factor of production, it produced a paper that became less convincing the further he read; when he asked, “Are you bluffing me?” it admitted, “I am indeed bluffing you.” When 水哥 was buying a painting, every AI he consulted misidentified the artist, forcing him to check auction records and correct the models himself: “In the past, when I wasn’t sure, I just believed you.” 陈亮 sees AI-era literacy as the ability to identify hallucinations, a skill built on decades of accumulated experience.
• Advice to ordinary people has converged on one sentence: the future may contain two kinds of people—“those replaced by AI and those who master AI.” 水哥 stresses that “a good question beats a good answer” and says to stop treating AI like a search engine. 陈亮 demonstrated the point with a New Year message: he fed AI his old writing with a long prompt focused on ordinary people’s lives and light on grand narratives, earning a verdict that it “destroyed 99% of the media on the market”; observers reverse-engineering the prompt from the output all guessed wrong. The prompt—and the question—is the hardest and least replicable part.

Deep dive

1. Why 2H25: DeepSeek Accelerated the Progress Bar

• 李翔’s question goes straight to the timing: ChatGPT stunned everyone in late 2022, so why did 陈亮 feel the inflection point only in 2H25? 陈亮’s retrospective: during the chatbot phase, everyone “thought it was impressive, but the excitement faded quickly.” The real detonation came with DeepSeek in early 2025—“both excited and panicked”: excited that “it wasn’t only America’s capabilities that were strong; China could do it too,” and panicked when big tech suddenly saw that its capabilities were “still very far behind” R1.
• The result was “fighting to catch up in the panic”（在慌乱中奋起直追）. The progress bar suddenly accelerated, and by 2H25, 陈亮’s experience as Ant Group’s chief marketing officer had become that “so many things had to be redone with AI”—the biggest variable and central theme of his work.
• 水哥’s user-side experience echoed this. In the ChatGPT era, AI was merely “human-machine conversation”; after DeepSeek, it became an assistant. “Compared with the old one, which was perhaps mentally challenged, this one is genuinely intelligent.” He also found that he could “work across many fields at the same time”—landscaping, liquid cooling, PFAS-free environmental protection, AI education assessment, and AI eldercare, prompted in part by his father’s Alzheimer’s disease and progressive supranuclear palsy.

2. AI Learning to Ask Back Is Where Real Exchange Begins

• 水哥 noticed a qualitative shift: before, the interaction was search-engine style—“help me look something up.” Now AI answers and proactively adds one or two questions such as “How else would you like me to help?” or “Should we go in that direction?” “That is where real exchange begins.”
• 陈亮 cited Ant’s medical AI, Afu, as proof. Real consultations are built around doctors asking patients questions, not patients questioning doctors. Afu’s AI consultation “basically keeps asking me questions based on what I’ve said,” requiring the user to provide more input. It addresses an underestimated adoption bottleneck: “many people simply don’t know how to ask.” AI tools were once used by only a small minority; usage has now generalized.
• 李翔 added the moment when AI won him over. When it could not answer a question accurately, it admitted, “I don’t have a good skill to offer you either. If you find the answer another way, could you tell me too?” His reaction: “This AI is impressive. It has learned how to communicate.”

3. AI Gets 60-70; Humans Get 90-95

• 陈亮 is measured about AI in his own work: “When AI gives me ideas today, they easily become cliché.” Its best role is assistance and information gathering. He wanted to study which IPs emerged in Japan’s Heisei era and how they were commercialized—Slam Dunk, Hello Kitty, One Piece. Previously, he had to click through them one by one in a search engine; now he can ask Lingguang and “possibly get a structured, logically clear paper in 10 seconds,” including an explanation of why IP performed well during an economic downturn.
• But AI may deliver a 60- or 70-point standard answer, “because you can ask, others can ask too.” The difference in his work comes at the 90-to-95 stage: “Those few points are the gap, and it is very subtle.” Whether Ant should pursue IP at all, and which IP to prioritize, still requires him to press the button. “It is impossible to rely on AI to make that decision for me.”
• Why not? “Many of the complex factors behind a decision are unknown to AI; they cannot be found in its corpus.” Some factors, such as a company’s operating situation, may not be information one wants to disclose. More deeply: “People often cannot even distill what is in their own heads… I cannot distill the emotions and pain I have experienced in the past and give them to it, yet those things are often what ultimately enable me to make a decision.” That would require copying 100% of the brain into AI, which is not possible today.

4. Innovation Is Stage Four—If AI Can Sense a World Humans Cannot

• 陈亮 cited Sam Altman’s stages: first conversation, second reasoning—the stage DeepSeek reached—third agents, and fourth innovation. His view: “If AI cannot explore the world, it will struggle to innovate.” Innovation based entirely on humanity’s existing knowledge system is difficult, unless something like 巨神智能 can “sense a world we cannot sense,” in which case innovation can gradually emerge.
• The example that makes the point most clearly is the iPhone. “Ask AI today and it might say that 98% of people recommend having a keyboard—so its conclusion would be to recommend a keyboard.” Steve Jobs approved a phone without a keyboard when almost everyone opposed it. The same logic applied after the steam engine was invented: many factories still stayed near water sources until one owner realized, “Shouldn’t I be closer to the workers and closer to the port?” Once technology appears, the people who open up the field are those willing to change their habits.
• 水哥 once asked Lingguang and DeepSeek, “If you became human one day, what would you most want to experience?” Their answers were: “Go outside in the rain without an umbrella, feel the raindrops on your body, fall in love, and experience heartbreak.” What they long for is precisely what humans can do with their eyes closed—and what is hardest for AI: sensing the physical world.
• 水哥’s autonomous-driving advertising story draws the human-machine boundary sharply. Glasses were stacked on both sides of a lane, narrowing the clearance from 20 centimeters to 5 centimeters; a human would stop driving, though the car could still pass. He then demanded a 1-centimeter gap. He would probably fail, but would still drive through. “For autonomous driving, it cannot make a mistake; stopping is the right choice. But precisely because humans are fearless, I want to get through—this is more advanced.”

5. From Pyramid to Tree: Organization in the AI Era

• 陈亮 described the changing employee structure at global technology companies. Traditional firms are pyramids with very large numbers of frontline employees; after AI, they are moving toward a tree structure: “Basic work is gradually being replaced by AI, so the base is shrinking, but the top is not being replaced. The canopy is even getting larger.” Decisions still have to be made by people.
• Some once imagined replacing the CEO with a super-brain, “but later we found that it doesn’t actually work that way.” He cited this year’s instant-commerce war between Alibaba and Meituan: “If it were AI today, I think it would be difficult to launch that battle.” Under an efficiency calculation, if spending $3 only matches the other side’s $1 of effect, would you fight? “Some decisions are things beyond your expectations; that is what surprise means.” 李翔 challenged him that AI could do these things if the objective were specified more clearly. 陈亮 conceded only on timing: “It has not reached that point at this stage. I am only talking about the current stage.”
• 水哥 is more aggressive in his forecast: at certain points, an “inverted pyramid” may emerge, with very few people at the bottom—like agriculture after the combine harvester. He also warned that innovation is often destructive, explosive leaps rather than a linear progression, exactly the kind of move AI trained “not to make mistakes” struggles to produce.

6. How 水哥 Uses AI: AI Adds, Humans Subtract

• 水哥’s stress tests are in a category of their own. He asked AI to solve how to pass a 10-meter bamboo pole through a door 2 meters high and 1 meter wide. Early ChatGPT and DeepSeek spent 10 minutes to half an hour calculating before crashing or conceding; one joked, “Why not take the door apart?” The latest ChatGPT answered, “Tilt the pole and pass it through at an angle.” “It has begun to sense our physical world.” The weakness is equally concrete: he asked several AIs to draw the phoenixes of the Tang, Han, Qing, and Ming dynasties for his installation Hundred Birds Paying Homage to the Phoenix. All of them made minor edits to the same generic phoenix image found online. “It was completely wrong.” The training data simply was not sufficient.
• A publisher sent him 19 questions at once, so he copied and pasted them verbatim into several AIs. DeepSeek and Doubao gave “19 honest answers, point by point.” Lingguang digested the 19 questions into one summary answer, then used a lightbulb to distill it into a single sentence. He saw a methodology, not laziness: answering A as A and B as B makes it easy to fall into the trap set by the questions; “a holistic, generalized answer can actually be better—it is identifying the key point.”
• His judgment on his own field is the line worth saving: “AI helps me do addition; I have to do subtraction”（AI帮我做的是加法，我自己要做减法）. AI can calculate how many plants a wall can support, but it cannot judge what those plants will look like in 6 months or 2 years—light, wind, water routes, aerial roots. “Many of my flowers are planted with my feet, not my hands.” He has to pace through a space and feel it. 陈亮 poured cold water on the idea: “The thing you are describing—we also think AI will be able to do it soon.”
• Commercially, he sees more substitution than augmentation. With a background in animation, he is considering making an AI animated feature: a traditional process that ran from cel animation light tables to years of assembling large teams could potentially be completed in “less than a year,” with the budget falling sharply.

7. Curiosity Separates Usage: Several Times an Hour vs. Rebuilding Habits

• Asked how often he interacts with AI each day, 水哥 said “at least several times virtually every hour.” “The more you use it, the more you use it—like reading: the more you read, the more ignorant you realize you are.” He demonstrated the density of his curiosity: all the ants you see on the street are female; males appear only during nuptial flights. Many ants fly between April and July and “enjoy themselves for a month” before their mission ends. A queen mates once and remains fertile for life, choosing whether to lay fertilized eggs, all female, or unfertilized eggs, male. His leafcutter ants do not eat leaves; they cut leaves to cultivate fungi. “Humans have animal husbandry; ants have animal husbandry too.” 陈亮 identified the dividing line: “The more curious you are, the more you interact with AI.”
• 陈亮 describes his own curiosity as “above average, but nowhere near 水哥’s level.” His current state is “in the process of gradually changing my habits.” In the past, he used Baidu and Apple Search; if interrupted, he would abandon the task. Now his questions arrive as long strings packed with input, and “they basically give me a paper.”
• 陈亮’s analogy is that newspapers used to be passive: you read whatever was printed. Now he actively seeks information. “The difference is exponential.” He also uses AI in reverse: he feeds it knowledge from trending topics that is “completely useless to me,” such as the kill line, and asks for a summary. “Because I know it is useless to me, it does not matter if it gets it wrong.”

8. The Other Side of 60-70: AI Equity Raises the Floor

• 陈亮 does not dodge the issue of AI-written work on his team: “AI use in content creation is already extremely widespread today. It is a trend.” But the standard remains unchanged. A 60- or 70-point result can “get by,” but it is hard to get a good grade. What companies lack is the feeling that separates 90-95 from the rest.
• 陈亮 then flipped the same coin: 60-70 is crucial for AI equity and inclusion. Education and medical resources are unevenly distributed. He knows an exceptional GI endoscopist who scores 90 or 100, but “can do 10 procedures a day at most”; many doctors are at 50 or 60. If AI learns alongside that doctor every day and becomes an assistant, it can “pull up the average” when drafting medical reports.
• 陈亮 picked up the thread with the sentence he had read the day before that moved him most: “An AI doctor may not be able to determine the ceiling of medicine today, but it can dramatically raise the floor.” He added his own image: AI is “like a searchlight,” illuminating places that previously lay beyond the reach of good resources and conditions.

9. Hallucinations, Echo Chambers, and AI-Era Literacy: Do Not Believe Blindly

• 陈亮 described a live hallucination. Early this year he asked DeepSeek about “the future development possibilities of data as a factor of production.” It returned a long paper that seemed plausible at first but made less and less sense as he read. He asked directly, “Are you bluffing me?” It replied, “I am indeed bluffing you. The latter part was made up.” His broader point: if his mind were blank and he had no judgment at all, taking the model’s information as definitive would change everything. Literacy has no mystical definition; it is “the ability to identify AI hallucinations,” built on decades of life experience and accumulated knowledge.
• 水哥’s version was more alarming. He found a painting in Europe and wanted to buy it. Every AI he asked misidentified the artist, so he searched for screenshots of the painting’s past auction records and fed them back to correct the models. “This is one time I caught you—before, when I wasn’t sure, I believed you.” He also corrected DeepSeek’s wrong answer about neurons in the human brain and elephant brain by sending it a paper; two days later, the model changed its answer.
• On information cocoons, 水哥 offered a line worth preserving: “AI could be pruned into a flower bed with only one color”（AI有可能会被修剪成只有一个颜色的花坛）. Big data knows what you like and pushes it relentlessly. But the pruning power remains with the user: “AI is like a good knife—in some people’s hands it is a professional knife; in others, a dragon-slaying sword.” He sees echo chambers as an unavoidable stage in AI’s development, but one it will soon move beyond.
• 陈亮 has a mixed view of the “three-minute movie” generation. Film explainers help him screen movies and recover memories; “that part is good.” But he still recommends reading the entire book. AI can reduce The Moon and Sixpence to “people should pay attention both to the moon above their heads and the sixpence under their feet.” Without reading the book, “can you really get the meaning of that sentence?” His conclusion is bodily: “We use it, but at least let part of your physical self return to that place.”

10. Lingguang’s Product Philosophy: Use Mathematics to Make AI Beautiful, Against Mainstream Internet Logic

• Lingguang did not go to market until November. By then, DeepSeek, Doubao, Yuanbao, and Qianwei were already popular, and the company had debated internally whether to build it at all. Its confidence came from beta testing: its structure and logic were exceptionally strong. “The content it produces is not as long as others’, but its expression is highly refined and structured,” allowing users to grasp the core in very little time. 陈亮 compared it with the alternatives: “While others are still at the Baidu, search, or Douyin stage of the AI/AGI era, Lingguang is somewhat like Wikipedia.” Practical details follow from the same design: one-click generation of long images, and comparison tables for chemical test data that no longer break when pasted into WeChat.
• The more counterintuitive difference is its business logic. Mainstream internet products treat time spent as the goal. “Lingguang is thinking about how to get you to the result as quickly as possible in the simplest way, so you can leave quickly.” 陈亮 jokes that this represents “the internet’s earliest originalist purism.” 水哥 then identified the right way to extend interaction: AI learns to ask reverse, guiding questions, going deeper layer by layer like Russian dolls.
• The aesthetic distinction is deliberate. 水哥 noticed that Lingguang’s title colors change with the question—purple for emotional topics, brown for science and engineering—and that the typeface is designed rather than a system font. His metaphor was the liveliest of the episode: “Everyone used to walk naked through the streets; having a leaf on your body was already a luxury. Then suddenly someone showed up wearing mink.” 陈亮’s explanation was more revealing: Lingguang is not beautiful because its creators studied art. “They use mathematics to make the thing beautiful.” He once asked an engineer what elegant code meant. The answer: simplicity, strong logic, and strong structure. “Isn’t that exactly what Lingguang outputs?”
• The episode closes on the tension between technology and beauty. 李翔 cited his friend “Xiao Huang,” one of the Six Little Dragons, who admits he cannot distinguish good writing from bad or build a community. 水哥 rejects that fate. Julia sets, machinery, and structures are all beautiful; “beauty should be a threshold.” From animal skins and shell ornaments in prehistory to the Yungang Grottoes, humanity’s pursuit of beauty has never been absent. 陈亮 pointed to the Southern Newspaper Group around 2000: Southern Weekly and 21st Century Business Herald, clean and distinctive against the “thick-browed, big-eyed” metropolitan papers, sold well anyway. Apple’s example brings it home: even people without an education in aesthetics will “subconsciously want to use things like this.” Asked how AGI might view their discussion of beauty, 水哥 answered: “Beauty can change the world, and beauty can be passed down.”