109: How Digital Life Kazik Uses AI: Automating Anything Repeated 3 Times
Summary
- Digital Life Kazik’s most replicable method is simple: “If anything makes me repeat it 3 times, I will definitely RPA or AI-ify it.” He starts by asking, “What do you find a waste of time and don’t want to do?” and works backward from real pain points: mining topics across 30+ communities, producing morning briefs from 20+ information sources, and screening 2,400 event applicants have all been turned from manual labor into reusable workflows. For organizations, AI’s first payoff is not replacing jobs but shifting attention from mechanical collection to judgment, experience and relationships.
- Chat Fund showed that building an AI product is not the same as building a sellable, scalable business. The team used GPT-3.5, embeddings, fine-tuning and a homegrown toolchain to build a fund Q&A product and even entered the second batch of algorithm filings. But financial clients capped the price below RMB200K while demanding private deployment, exclusive promotion of their own funds, differentiation around their strengths and 2–3 people on-site for customization. The unit economics did not work, so Kazik cut his losses in August 2023. “Technical moats” ultimately lost to procurement systems and non-standard delivery.
- Digital Life Kazik’s high-frequency output rests on a boundary: content industrialization can only raise the floor. The system scores several hundred to roughly 1,000 pieces of information every day; humans review only the top 50 and select 10 for the morning brief. AI also summarizes and analyzes community chats and comments, cutting topic research from 5–6 hours of platform browsing per person per day to roughly 1 hour for Kazik. But product experience, opinions and final copy remain human work: “If AI can experience things for me, why wouldn’t I just use the brand’s brief?”
- The clearest value today comes from narrow, deep Agents; general-purpose Agents remain constrained by long-chain success rates. Kazik breaks an Agent into planning, tool execution and memory. If each link succeeds 90% of the time, the compounded rate quickly falls below production requirements. When he tested Manus, only 2 of the first 10 tasks succeeded before the servers were adjusted; after he understood its boundaries, the rate rose to roughly 5–6 out of 10. By contrast, Deep Research for research, Cursor for programming and o1 Pro for complex grouping are already genuine productivity tools.
- DeepSeek’s biggest contribution was not a single benchmark but expanding China’s AI user pool from Kazik’s estimated 100M people toward a mass market that could add roughly 800M more. Its fully displayed reasoning process showed first-time users that AI could actually think, making them more willing to revise their prompts after a failure instead of concluding that “AI is nothing special.” But continued competition in math, code and research also makes capability gains harder for ordinary users to perceive, creating tension between mass adoption and product elitism.
- Model upgrades will keep consuming thin applications, but products can still survive through multi-model orchestration, engineering optimization and user mindshare. Jasper worked before ChatGPT by packaging marketing prompt templates, then came under pressure as GPT, Claude and other general models improved. Perplexity and Cursor represent the idea that “the shell has its own use.” Manqi noted at the end that when GPT-4o’s image generation opened up, OpenAI did not simultaneously open the API or the free tier, making the paid ChatGPT plan a temporary exclusive supplier. Applications operating on a model company’s main road face especially high investment risk.
- Kazik is not building an AI startup not because he is bearish on demand, but because he cannot see a moat outside the reach of the major platforms while the mobile entry point remains unchanged. He describes startup success in this generation as shifting from mobile internet’s “20/80 split” to “99/1”: startups lose across compute, data and talent, and, by his account, the most expensive intern can “apparently” earn roughly RMB2,000 a day. He would rather be “the doorman at the entrance to AI’s temple” than “burn investors’ money” without conviction.
- AI’s practical meaning for ordinary people is a new deal in a weak labor market, but the path has to be “industry plus AI,” not “AI plus industry.” Kazik has seen civil engineers, shipbuilders, film producers and unemployed master’s graduates from China’s 985 and 211 universities use AI to move into new roles. He has also seen readers enter Kimi internships or take on AI projects inside traditional organizations after years of practice. “Giving some people some opportunities to keep surviving” matters more to him than AGI narratives.
Deep dive
1. Kazik Entered AI Through User Experience Design, Not an Algorithm Lab
Kazik graduated in 2017 but started working in 2014. Believing university courses were too far removed from mobile-internet practice, he entered UI, interaction design and user research early, later moving into product management and management roles.
Nearly a decade of his career was spent in finance. His early exposure was to “decision-making AI”—recommendation systems, risk controls and decision trees—which was “completely unrelated” to today’s generative AI. He handled the interaction and copy corresponding to algorithmic capabilities, not the models themselves.
In 2019, he helped build an insurance product from scratch, his first systematic exercise in thinking about product form, user experience and business models. He later brought that product discipline into AI: conduct user research, break down pain points, then choose the technology rather than starting with a model and searching for a use case.
2. An Obsession with Investing Led Him into Public-Fund Data
Kazik once reviewed markets every night from 10 p.m. to 3 a.m.: sector moves, leaders and “central” names, financial metrics, earnings reports and fund-manager comments. At 9 a.m. he checked pre-market news, then prepared to trade at 9:25. “Apart from finishing the company’s work, I was trading stocks and buying funds.”
After losing money in stocks, he gradually moved his capital into funds. By his account on the show, he had been almost fully invested in funds since 2021 and still achieved cumulative returns of roughly 120–130 percentage points after 3 years of a bear market. That personal interest led him to join a friend’s startup in 2022.
The team built “Jiuquan’er,” a public-fund data platform with roughly 50K daily active users and several million total users. Its goal was to use objective data to resist the sales logic of recommending whichever fund had risen the most, and let fund investors see the actual performance of funds and managers.
3. The Bear Market Exposed the Fund Data Platform’s Business Model
The platform did not sell funds and held no sales license, so it had no commission revenue. The remaining options were advertising and To C value-added services. A product team of roughly 50 people plus a 10-person media team was badly mismatched with realizable revenue.
From 2022 to 2024, the fund market stayed depressed. Users stopped caring about funds, while fund companies cut brand budgets from the tens-of-millions and millions of yuan range to “there’s no budget this year; the marketing budget is zero.”
The team still fought to break even every year but had “already lost sight of hope.” By March 2024, the company had been cut to a dozen-plus people and shifted toward a previously acquired private-fund license. Kazik wanted to keep building products, so he left and began running Digital Life Kazik full-time.
4. ChatGPT Turned Natural-Language Interaction from a Gimmick into a Productivity Tool
While studying “emotional design” in 2018, Kazik had already considered natural-language interaction: tell a computer, “turn the rectangular border into 3 pixels, then move it 10 pixels to the right,” and the system might not even know whose frame of reference “right” referred to.
Manqi used Luo Yonghao’s TNT as an example and asked whether building PPTs and Excel through speech was too far ahead of its time. Kazik’s answer was that an idea being ahead of its time was useless unless it landed in the right era. Given the technology then, it was “just cool, but had no efficiency whatsoever.”
Roughly 1 week after ChatGPT launched, the team connected a bot to a WeChat group, although the group “blew up” the next day because of an account issue. Kazik felt it was fundamentally different from Siri and Xiaoai: natural-language interaction had changed at the base level, and so had the assumptions behind product interaction.
5. “Slack Time” Turned Company-Wide AI Adoption into a Clear Incentive
Before the 2023 Lunar New Year, the team had already required the entire company to use GPT-3.5. Developers found code completion mediocre, but it could translate existing Python into Go, after which engineers would revise it—getting around the team’s shortage of Go developers.
The design team was required to use Midjourney and learn Stable Diffusion. A poster schedule that had taken 4 days was compressed to 2, but Kazik promised not to shorten it further for 6 months: “Once you finish it in half a day, the remaining 1.5 days are entirely yours.”
The system became known as “slack time.” Operations staff also had to submit 10 brainstorming directions before proposing a plan, while time saved by AI inside a fixed schedule was not filled with new tasks. Once mandatory learning was paired with a guaranteed benefit, the team showed unusually high enthusiasm.
6. A $70 Fine-Tuning Mistake Forced the Team to Understand Embeddings
At first, the team understood nothing about embeddings, fine-tuning or tool calls. Because fund NAVs changed daily and ChatGPT fabricated data, Kazik once proposed “fine-tuning it once a day.” Looking back, he calls the idea “extremely stupid and unprofessional.”
They actually fine-tuned a model on roughly 1,000 data points, spending $70, then realized that neither the cost nor the update mechanism worked. At the time, the internet offered little beyond prompt tutorials, so the team studied the OpenAI API documentation page by page and eventually found embeddings better suited to database retrieval.
The fine-tuning experience was not wasted. The team used small models such as GLM-6B for intent recognition, text segmentation and JSON conversion, routing questions like “compare 2 funds” and “find funds with positive annual returns every year” to query and SQL tools.
Mature function calling did not yet exist, so they built their own retrieval, data-assembly and generation pipeline. They still chose GPT-3.5 to organize the answers because, in tests in April and May 2023, Kazik believed no other model matched its assembly ability.
7. Chat Fund’s Moat Was Fund Methodology, Not Ownership of a Foundation Model
Chat Fund aimed to answer any fund question at roughly the level of a financial adviser with 2–3 years of experience. The team believed its real advantages were fund know-how, proprietary processed data and reporting methodology—not which general model it called.
Development was handled mainly by Kazik, the technical director and limited engineering resources—roughly “1.5 people.” Kazik worked on fine-tuning, prompts and financial methodology, while the technical lead handled the backend APIs and engineering pipeline.
The product later entered the second batch of algorithm filings in 2023. Kazik said it “should have been” the first approved filing in finance and was in the same batch as SenseTime and Baidu. The filing showed that a small team could build quickly, but it did not automatically create a commercial business.
8. The RMB200K Procurement Ceiling Destroyed Financial To B Unit Economics
Kazik spoke with fund companies, banks and brokerages across the market and found that a standardized solution could sell for at most roughly RMB200K. Above that, the client had to enter a higher-level internal review, and customers were unwilling to push a complex approval process for a product seen as a “wrapper.”
Manqi pointed out that large clients already preferred private deployment and on-demand customization. Their requirements also made business sense: recommend only their own funds, highlight their strengths against competitors, embed the product in official accounts and mini-programs, and absorb concurrency and token costs.
For the vendor, every client required rebuilding the database, prompts, UI and engineering interfaces, sometimes with 2–3 people providing ongoing service. The price was fixed while delivery was highly non-standard, leading Kazik to conclude that it was “probably a loss-making business.”
He did not continue because of sunk costs. In August 2023, he cut the business directly: “Even if we lock in the client, with such a small cost base, how much future ROI can we get? I don’t think it’s reliable.”
9. Pika and Sora Pushed a Financial Product Manager toward AI Content and Film
In August 2023, Kazik produced a trailer for The Wandering Earth 3, attracting attention in the film industry. He met Guo Fan in September that year, while the collaboration truly began in November 2024. He said AI content for The Wandering Earth 3 is worth anticipating, and the film remains scheduled for release during the 2027 Lunar New Year holiday.
When Pika went viral in November 2023, he said he was the only person in China with an internal-test account. His launch review entered financial and market circles, and the next day he conducted 13 consecutive roadshows for fund managers and researchers.
His biggest breakout came on February 16, 2024. Having already worked on AI video for 6 months, he wrote an overnight article on Sora titled “Reality No Longer Exists,” which drew more than 3M views. Kazik attributes it to “99% luck and 1% to not sleeping that night.”
10. Five Modalities Iterating at Once Forced Content Production into High Gear
Kazik has long divided AI into text, images, video, audio and 3D. In 2023, the areas he could write about continuously were mainly text and images; video was close to PPT, audio was still mostly SVC voice conversion, and 3D remained crude.
ChatGPT’s voice feature appeared during the October 2023 National Day holiday. The shock was that the model could pause, say “uh” and begin to show emotion. Products such as Tripo 1.0 did not appear until the end of that year. Only after 2024 did the 5 modalities, multimodality and reasoning models truly accelerate together.
After Kling launched in June 2024, competition in AI video intensified abruptly. By the time of the interview, major models and products were updating so quickly that Kazik felt not writing about an update such as QwQ-32B would mean he was “not professional.”
When working part-time, he set himself a cadence of 1 update every 3 days or less. After going full-time in March 2024, he raised it to 4–5 posts a week or even 2 a day. His normal schedule ran from 3 a.m. to 9 a.m.; when OpenAI released a voice model late at night, he sometimes worked past 5 a.m.
11. Missing DeepSeek’s Launch Exposed the Brutality of the Self-Media Attention Market
Kazik admitted he wrote nothing on the day DeepSeek-V3 or DeepSeek-R1 launched. The official account had already received extensive reposts, he judged that another article would get no traffic, and he wanted to sleep, so he passed. He later classified it as a “mistake” that could damage user trust.
As AI products multiply, users’ thresholds have also been raised by “mind-blowing” narratives. Ordinary people struggle to perceive a 3–5 point benchmark improvement and are more likely to reject incremental progress with an intuition such as “9.11 is still bigger than 9.8.”
o1 and o3 may be highly meaningful for math, molecules and new materials, but those are not scenarios most consumer users can evaluate or care about. Kazik’s content anxiety therefore shifted from “there is no update to write about” to “the update matters, but the public cannot see it directly.”
12. DeepSeek’s Real Breakout Was Adding Users with No ChatGPT History
Manqi believed DeepSeek-R1’s creative-text ability was relatively strong, making it easier for ordinary people to try directly. Kazik called Kimi K1.5 and Zhipu Zero “math champions” and “specialist champions”: potentially excellent at math but insufficiently general. He also mentioned hallucination issues in models such as K1.5 without attributing that judgment directly to DeepSeek-R1.
By Kazik’s rough estimate, roughly 100M people in China were genuinely using AI before DeepSeek, and products had been competing repeatedly for the same pool. After DeepSeek broke out, “800M users could appear out of nowhere,” with greater dispersion by age, region, industry and technical fluency.
For this group, the first AI product was DeepSeek rather than ChatGPT. Content about prompts, poetry, making money and combining AI with Jianying therefore became popular again in 2025. Tencent Yuanbao even promoted itself by painting rural walls, producing mass-market narratives such as “ask Yuanbao about sow care.”
Manqi agreed that “DeepSeek is China’s ChatGPT moment,” while noting that the impact gap between models such as Kimi K1.5 comes not only from capability but also from whether ordinary users can feel the difference in low-barrier tasks such as writing poetry and classical Chinese.
13. Visible Reasoning Changed AI’s Psychological Position for Users
When traditional models made mistakes, users often said “AI is nothing special” and left. Manqi noted that o1 did not fully display its chain of thought, while DeepSeek did. Kazik was unsure whether this was deliberate design and added that DeepSeek’s early open-source models already contained reasoning traces.
Kazik observed a subtle shift: when dissatisfied, users no longer blamed only the model; they also questioned their own prompting and became willing to learn how to collaborate with AI. That psychological shift encouraged them to share results with friends.
o1 already had reasoning ability but remained concentrated in tech circles because of access friction and the $20 monthly price. DeepSeek’s website and app were easier to access. Whether the reasoning process was shown further shaped ordinary users’ perception of model capability.
14. In Real Production Testing, o1 Pro Could Still Beat the Breakout Model
In a real script workflow related to a film project, Kazik tested o1 Pro, DeepSeek-R1-Lite-Preview, Kimi K1.5 and the then-unreleased Zhipu Zero together, rather than conducting a content benchmark.
His judgment at the time was that o1 Pro was “simply too strong” in literary quality, while the other models performed much worse. He did not retest the formal DeepSeek-R1 release in the same workflow, so he did not equate public attention with production capability.
This was the disagreement the show repeatedly preserved: DeepSeek’s value lies in mass access, visible thinking and user expansion, but in demanding scripts, programming or math, breakout attention alone cannot establish comprehensive leadership.
15. “The Doorman at AI’s Temple” Is Closer to Kazik’s Self-Image Than Technical Authority
Kazik does not describe himself as an algorithm expert but as “a doorman at the entrance to AI’s temple.” He simply plays with many tools, knows which ones might solve ordinary users’ PPT, Excel, voice-over and small-tool problems, and guides people through the door.
The account’s early slogan was “Pass one lamp to ten thousand lamps until all ten thousand lamps shine.” He later dropped it because it sounded too much like a religious leader. The simpler description now is “sharing some very new and cool AI know-how.”
His goal is not next year’s valuation or revenue. It is to make money through influence while reaching people and projects previously beyond his access: collaborations with Guo Fan, Xiaomi’s Wang Chuan, Huace Film & TV, CCTV-6 and Xuexi Qiangguo form a continuing feedback loop for creation.
16. Human Imperfection Became a Trust Signal in the AI Content Era
Kazik sees LatePost as a benchmark for professionalism and authenticity and wants “not a single point” of the content to be wrong. Yet his articles often contain typos. The direct causes are a convenient but error-prone cross-device WeChat keyboard and his reluctance to proofread long drafts repeatedly after finishing them.
One reader interpreted the typos as a “declaration that a human wrote this,” unexpectedly matching Kazik’s judgment: as AI-generated content floods the market and language becomes increasingly uniform, human traces and individual personality become scarce.
He therefore insists that final expression retain a human feel, but does not package typos as a professional virtue. Manqi also warned that human drafts miss errors too; genuine respect still comes from research, fact-checking and responsibility for readers’ time.
Kazik once attempted an AI copyright deep dive, spending nearly 1 month interviewing elite law firms, policy participants and data-security and IP lawyers and accumulating 70K–80K Chinese characters of material. The article drew only slightly more than 10K views but was reposted more than 3,000 times, showing that depth and mass distribution do not always align.
17. The HKR Method Gives Technical Tutorials a Distribution Lever
Kazik adopted HKR from Yingshi Jifeng: Happiness, Knowledge and Resonance—something interesting, something useful and something emotionally resonant. He tries to satisfy all 3 rather than merely explain a new model’s parameters and features.
When promoting the open-source voice-cloning tool F5-TTS, he did not write a software manual. Instead, the day after Fu Hang won a championship, he recreated Fu’s voice to deliver a stand-up routine. The hot topic attracted attention; the tutorial captured the underlying demand.
When Hailuo AI launched, Kazik used Xiang Zuo’s nose-touching performance as the hook, recreated a specific segment with the model and compared the characters’ emotions. The article drew more than 100K views and roughly 19K reposts. The capability became a watchable joke rather than an abstract description.
Resonance is the hardest part of HKR. Kazik admires Yingshi Jifeng’s ending in which it sent a fan’s wish into space, because it was not a mechanical attempt to “elevate the theme” but a sincere moment that could make people’s “scalps tingle.”
18. National Narratives Create Resonance—and Amplify Backlash
Kazik admits that he often uses The Wandering Earth, sixth-generation fighter jets, Nezha, Black Myth: Wukong and domestic AI as emotional anchors. Readers consequently accuse him of “promoting domestic products too much.”
His explanation is not a traffic strategy but personal experience. Born in 1995 and raised in Anqing, Anhui, he grew up at a time when KFC and McDonald’s were reserved for birthdays or good grades. Later he watched new-energy vehicles and Chinese products go overseas, so the emotion is genuine.
In AI video, he especially treats Kling’s June 6 launch as a “light of domestic technology.” By his observation, Hailuo had the highest DAU at the time of the interview and Kling ranked second. Runway slowed markedly after Gen-3 and Gen-3 Alpha, while domestic products including Hailuo, Kling, Jimeng and PixVerse updated faster.
Manqi’s response is worth preserving: the benefit of personal expression is having a clear opinion, but clear opinions inevitably attract equally clear opposition. Part of being a self-media professional is accepting that there will be “resonance and counter-resonance.”
19. Manus’s Controversy Began with Scarcity Allocation, Not Product Capability
On the night Manus launched, Kazik first saw a video in an AI group, then contacted its CMO through SevenJoy to obtain an invite code. He emphasized that the review was unpaid and that he had not known the team previously.
He believes the biggest marketing mistake was sending invite codes only to major influencers. A normal beta has 2 parties—the product and the user—but invite codes created a third transfer layer, quickly producing scalpers, account trading and screenshots of codes selling for RMB50K on Xianyu.
A same-name cryptocurrency rose at the same time and was mistakenly thought to be connected to the team, reinforcing the impression of a “scam.” Kazik believes the many new users brought in by DeepSeek had not experienced the widespread beta queues of 2023 and were especially hostile to being unable even to connect and try the product once.
Compute scarcity was real. The show mentioned a photo of someone wearing a T-shirt claiming Manus burned roughly $1.04M in cloud-token costs during its first 14 days. Kazik estimated that a simple task could consume tokens in the millions. Manqi said Manus later partnered with Tongyi; Kazik believed that could help reduce compute pressure in the domestic version.
20. Ordinary People Should Start Learning AI with What They Do Not Want to Do
Whenever someone asks, “What AI should I learn?” Kazik first asks back: “What do you find a waste of time and don’t want to do?” Only once the answer is specific does tool selection become clear.
A legal professional who does not make videos and has no plan to become a creator has no reason to learn AI video out of anxiety. Kazik calls “learn anything as long as it’s AI” a “disguised form of lazy coping”—using tactical diligence to replace strategic choice.
Once the pain point is stated as “I don’t want to make PPTs,” “I need to validate a product demo” or “I don’t want to build models manually,” it can be decomposed into clear tasks. Kazik’s principle is not to become omnipotent but to first identify the time that can be saved.
21. More Than 30 Communities Became a Continuously Updated Demand Database
Kazik runs more than 30 private communities with roughly 15K members. Every night at 8 p.m., the system captures chat records and asks AI to summarize them and select the 10 most valuable pain points as a reserve of tutorials and topics.
After each article is published, he reviews the like-to-read ratio, shares and comments, caring most about shares. One or two hundred comments are captured and then analyzed by AI for readers’ positive and negative reactions.
A developer on the team built the chat capture system. Kazik built the comment-analysis tool as a Tampermonkey plugin using ByteDance’s coding tool. The technology is not complicated; the key is turning scattered feedback into continuous data.
The topic system also catches “low-follower viral articles.” If an account that normally receives only several hundred to 1,000 views suddenly gets 50K, it enters the candidate pool. Metrics such as a like-to-read ratio above roughly 1%–2% are then used to determine whether it offers genuine reference value.
22. A Three-Times-a-Day Information Pipeline Compresses Noise into Decision-Ready Input
The 8 a.m. morning brief covers the previous 24 hours. Behind it, the system pulls from 20+ data sources and several hundred to roughly 1,000 pieces of information, then uses a general model to grade them S/A/B/C based on AI relevance and importance.
After the model ranks them, humans review only the top 50 and select the 10 most important for the community. AI handles the volume of reading; humans retain fact-checking and priority judgment.
At 8 p.m., an internal “briefing” adds roughly 20 new events from the day and includes low-follower viral posts from official accounts, Xiaohongshu and Douyin. At midnight, Twitter is checked again to confirm whether breaking information requires an immediate topic change.
When a major event occurs, Kazik begins testing and writing at 8 p.m. If not, he combines reserve topics with current events from the past 2 days. The system does not decide his content; it provides a continuous information feed and candidate topics.
23. Content Industrialization Raises the Floor but Cannot Manufacture the Ceiling
Without the system, the team spent 5–6 hours a day browsing Twitter, Weibo, Xiaohongshu and Bilibili. The information density led interns to leave in September and October and left creators almost no time for offline interaction.
Automation began in December, produced a demo in January and entered use in February, followed by continuous revisions. Kazik now spends roughly 1 hour a day on topic selection. He says explicitly, “It cannot possibly raise my ceiling.” The real benefit is time to record podcasts, meet people and manage the team.
“Content industrialization” is suited only to input, clustering and candidate ranking because these are fundamentally about collecting large volumes and extracting signals. Final experience, judgment and expression remain non-standard products; excessive industrialization turns content into garbage.
Product experience especially cannot be outsourced to a model. After obtaining a tool, he must run it himself to develop feel and an independent conclusion. Kazik asks, “If AI can experience it, why would I use something someone else has already integrated? Why wouldn’t I just use the brand’s brief?”
24. GPT and Claude Play Different Roles in Real Work
Kazik relies most on 2 models day to day. GPT handles serious analysis and hallucination control; Claude occasionally handles creativity and programming. He often criticizes OpenAI, yet admits that when he has to “do something major,” GPT remains his most-used model.
He has long paid for the $200-per-month tier, mainly for o1 Pro and Deep Research. The $20 Plus plan could also use Deep Research at the time, but offered only roughly 10 runs versus about 150 on the $200 tier.
Kazik uses Deep Research 2–3 times a day on average and sometimes reruns it when dissatisfied with the framework. Regenerating a report after each revision counts as a new run. 150 runs are roughly equivalent to 5 per day, which is basically enough for his work.
His cost comparison is direct: roughly RMB1,400 a month, cheaper than an intern, while he can do other things after posing the question. If a report materially shortens research, it is not a conspicuous-consumption purchase.
25. A “Cross-Shaped Research Framework” Turns Deep Research into a Researcher
Kazik starts horizontally: what competitors, comparable companies and products exist at the same time, and what are their respective strengths, weaknesses, business models and profitability? He then goes vertical: what happened from the company’s founding to today, why it launched a particular product and how its strategy changed.
The conventional approach requires reading large volumes of research reports, media coverage and historical material. Early scattered information on companies such as Alibaba is especially difficult to find. Ordinary AI search and Perplexity are still not deep enough in his view.
Once he gives Deep Research the cross structure, output requirements and research subject, it can produce a roughly 20K-word report in about 30 minutes. After comparing it with friends in finance, Kazik rated it at “the level of a researcher with around 3 years of experience,” while stressing that a senior person must still identify hallucinations.
He also uses it to fill in periods he lived through but never understood in detail, such as mobile-data pricing and the shift from 3G to 4G, then draws analogies between those historical mechanisms and the current AI industry.
26. Deep Research Writing a Novel Demonstrated the Value of Long Context for Logic
Kazik believes the hardest part of a novel is not prose but logic, foreshadowing and “Chekhov’s gun.” If an object appears in Act 1 without serving a later purpose, it should not appear; conversely, a key prop must run through the structure.
In his experiment, Deep Research first compiled 20K–30K Chinese characters of background on 16th-century alchemy, then wrote a roughly 30K-word novel based on the setting. By his account, Deep Research uses o3 underneath, and its long-range structure is clearly better than models that only make individual sentences sound beautiful.
In the finished work, the teacher introduced in Act 1 becomes the adversary in Act 5. The duke who initially recruits the protagonist changes his mind and personally sends the protagonist to prison in Act 4. The philosopher’s stone evolves from a clue in a book into foreshadowing for the summoning of an ancient Cthulhu god.
The structure is still conventional and hardly groundbreaking, but it is readable. A writer friend compared it with A Song of Ice and Fire, leaving Kazik to ask: “That is A Song of Ice and Fire. You’re asking it to do George R.R. Martin’s job right out of the gate?”
27. AI Capability Is a Gradient of Gray, Not “Replacement or Completely Useless”
Kazik sees unrealistic expectations as the biggest usage error. No model reaches 100% accuracy, and no model becomes equivalent to a 10-year veteran employee the day it launches. Even if Deep Research performs like a 3-year researcher, humans must still check for hallucinations.
But failing to reach master level does not mean having no value. AI may not be able to write A Song of Ice and Fire, but it might reach the level of “60% of web fiction.” It may not solve an entire job, but it can take over 60% of it, leaving genuinely available “slack time.”
AI video went through the same expectation reset. When Sora launched, the fantasy of “one click to generate a movie” was too high; after disappointment, users labeled it garbage. The industry later found that while it could not make films, it could make short dramas.
Kazik cites The Strange Incident in the Xing’an Mountains as the only AI short drama at the time to achieve positive profit-sharing returns. The market was willing to pay, showing that production value had already emerged within the capability boundary.
28. Video Production Keeps the Human Core and Gives Material Labor to AI
Kazik uses AI to research and assist ideation for short-video scripts, but humans still write the final version. Complex effects in editing are mostly done by hand because automation is not necessarily faster than an editor at this stage.
AI is better suited to B-roll. Supplementary footage that once required time to search, purchase and check for copyright can be generated with tools such as Kling and Jimeng. Special sound effects can also be added, while ordinary effects do not need AI because Jianying’s asset library is already sufficient.
Music usually comes from existing licensed material. AI music is considered only for videos with a strong narrative or plot. Kazik’s standard is not whether something can be AI-ified, but actual efficiency, quality and copyright cost.
The clearest red line is not using a digital avatar or AI-generated voice in the finished video: “The IP I’m building needs to have a human feel.” In his view, a personal IP is a trust business. Mass-produced digital humans may make traffic ROI work, but they weaken long-term trust.
29. Digital Avatars and Virtual Idols Are Completely Different Businesses
Manqi used Hatsune Miku to ask why a digital human could not become an IP. Kazik accepts that a virtual IP can work, but sees it as an independent character, brand and narrative—not AI impersonating him to produce routine updates.
One advantage for film and talent agencies creating virtual artists is eliminating the risk of a real actor “collapsing” in a scandal. But if Kazik created a virtual character, it should be brand-separated from Digital Life Kazik.
His conclusion is therefore not a rejection of generative content but a distinction in user expectations. A real-person knowledge IP must let the audience know who is experiencing, judging and taking responsibility; a virtual character can build its relationship from the outset under a fictional identity.
30. An Agent’s Three Core Capabilities Are Planning, Execution and Memory
Kazik breaks an Agent into 3 parts: first, turn a one-line request into detailed steps; second, execute through a sufficient number of tools; third, use long context and memory to connect the process and conclusions into a coherent output.
He sees MCP as a more universal interface layer. Once someone packages Blender’s capabilities under a protocol, Claude, Qwen and other models can call them without every team having to reload API documentation and build a separate integration.
The larger the tool set, the stronger execution can become. But if intermediate steps are compressed into a conclusion, clues needed for later integration may be lost. For report-writing Agents, memory quality directly determines whether the final output is coherent.
31. Long-Chain Success Rates Multiply Downward, Leaving General Agents in the Toy Stage
Kazik uses his father’s textile production line to explain industrial requirements. An 80% yield means “a garbage production line”; upgrading equipment and decision models is meant to push every link as close to 100% as possible.
Generative Agents are exactly the opposite: extremely long chains in which no step is guaranteed to be correct. “90% times 90% times 90% times 90%” falls rapidly as steps accumulate. A bad Xiaohongshu post is tolerable; production control is not.
He saw the risk more clearly while participating in the release of the Shanghai Artificial Intelligence Laboratory’s MedBench medical evaluation benchmark. Any error in a medical Agent could become a misdiagnosis with serious consequences, so it cannot go live under demo-stage tolerance standards.
Kazik is not rejecting the future of Agents; he believes they currently resemble toys. Production deployment requires long tasks, batch tasks and exception recovery to approach the reliability of mature RPA.
32. Manus’s Benchmark Advantage Cannot Eliminate Failure on Real Tasks
When Kazik first tested Manus, he ran tasks while livestreaming. Before 4 a.m., only 2 of roughly 10 tasks had succeeded. After 4 a.m., server repairs and additional compute, combined with his decision not to request complex browser games, raised the success rate to roughly 50%–60%.
It could attempt 2048, but “Beat the Kids” and online Red Alert were clearly beyond its capabilities; once it could not even enter the webpage. The process showed that understanding the boundary itself is part of the success rate.
Manqi asked why Manus ranked above OpenAI Deep Research across multiple GAIA benchmark levels if that was the case. Kazik’s answer was that the tasks were different: Deep Research only produces reports, while Manus also operates virtual machines, programs and browsers. Broad coverage does not mean greater stability in every task.
Zhipu AutoGLM offers an even sharper example. It could like the first Moments post of the first 5 people in a group. Expanded to 50 people, it began slowing around person 7 or 8 and no longer knew whose post to like by person 10. The longer the sequence, the easier it is to lose state.
33. Truly Usable Agents Are Often Already Hidden in Vertical Tools
If Agent is not limited to a “general assistant,” Kazik believes many usable examples already exist: Deep Research specializes in research and Cursor in programming, each achieving higher success rates on shorter chains.
He would not use Manus to write a large codebase because products inside IDEs such as Cursor and Trae are more professional. Similarly, Deep Research cannot replace every browser operation, but it creates stable value in the single task of generating reports.
This forms Kazik’s interim judgment: general Agents are worth expecting, but for now it is better to ask whether a vertical task has a clear boundary and a verifiable success rate.
34. AI Programming Moved Him from Dragging Nodes to Hand-Building Tools
Before mature AI IDEs existed in 2023, Kazik used Dify and Coze to build knowledge bases, customer service and workflows. But financial products were too deeply customized, so the team still developed the backend and engineering pipeline itself.
He now uses visual nodes less often because finding modules, dragging them in and wiring them together remains cumbersome. With products such as Cursor and Trae, he would rather “throw it in with my mouth” and directly generate a small plugin or scraper.
Kazik is especially fond of ByteDance’s Trae: native Chinese support and, at the time, free access to Claude 3.7 made it more suitable for non-professional developers building small tools. Cursor is more professional, but its English interface, cost and large-IDE logic are heavier than his needs.
He retains workflow thinking only for longer processes requiring multiple models to collaborate. These tasks are increasingly handed to Feishu Base, where data, organizational permissions and automation sit in one interface.
35. Feishu Base Turns Every Column into a Model Node
Kazik has demonstrated an entertainment-oriented workflow: input a user’s profile photo; Doubao’s vision model describes it; DeepSeek-R1 infers possible personality traits; the result becomes a Jimeng image prompt; and the system generates a photorealistic image of what that person might look like.
Each column receives the previous column’s output and becomes the next column’s input. Once the template is fixed, adding a new row runs the process automatically. Kazik sees it as an intuitive, inspectable standard workflow rather than a simple spreadsheet.
In actual production, it is used more for tagging, translating, summarizing and transcribing captured content. Feishu’s interfaces can call DeepSeek-R1, Doubao, Jimeng 2.1 and video-generation capabilities, with token costs settled by linking a Volcengine developer account.
36. Screening 2,400 Event Applicants Demonstrated the Organizational Leverage of Table AI
In March 2025, Kazik held an offline event and received 2,400 applications, of which only roughly 200 could be selected. He first asked Base to extract each applicant’s company and position, then tag industries such as finance, media, film and state-owned enterprises.
He then asked DeepSeek to score each applicant’s willingness to participate from 1 to 10 based on a one-sentence statement, improving screening efficiency and the visibility of diversity quotas.
Screening 2,000 people had previously taken roughly 2 days; this time it was completed overnight. The key was not letting the model make the final selection but turning unstructured applications into sortable, screenable fields for humans to choose from.
The event aimed to encourage cross-industry interaction, so selection had to balance quotas while later grouping handled conflicting rules: similar backgrounds, competitor separation, separate treatment for state-owned-enterprise executives and keeping specific community members together.
37. o1 Pro Completed in 11 Minutes a Seating-Plan Job That Previously Took 11 People 2 Days
Base could not directly satisfy 8 complex grouping rules, so Kazik gave roughly 200 people and the constraints to o1 Pro. The model reasoned for about 11 minutes and produced a grouping table he called “particularly impressive.”
When he organized a similar event in May the previous year, he found 11 friends with different backgrounds and coordinated for 2 days during their slack time. They were not working full-time for 2 days, but the difference in organizational cost versus 11 minutes was still obvious.
This also showed that models need task-specific assignments: online investigation goes to Deep Research, while complex constraint reasoning without web access goes to o1 Pro. Kazik said ordinary o1 and the domestic reasoning models available at the time could not handle this grouping.
His overall principle finally condensed into one sentence: “If anything makes me repeat it 3 times, I will definitely RPA or AI-ify it. I absolutely will not do it myself.”
38. The Best Voice-Cloning Case Was About Companionship, Not Traffic
Kazik has almost no “private stash” of tools because he writes about good products immediately. But the deepest case left by Hailuo’s voice cloning involved a fan whose husband had died and who still wanted her children to hear their father’s voice.
In 2023, through his relationship with the Hailuo team, he used roughly 90 seconds of old audio and an internal process to generate the voice. The fan’s exact assessment was “70% similar.” Kazik did not expect it to reach 100% of her husband’s voice, but thought the result was already very good, and the 2 children were grateful.
The user-facing product at the time allowed only the recording of one’s own voice, controlling privacy and copyright risk. The overseas version later allowed any voice to be cloned, making it more useful to voice-over creators. Kazik also acknowledges the resulting privacy and copyright issues.
Another tool he has used for a long time, Viggle, is more entertainment-oriented. A dance video and a character photo can replace the subject, enabling series such as “historical figures come back to scold you” on video platforms. It sacrifices industrial-grade skeletal tracking for the low barrier of “1 video plus 1 image.”
39. A Two-Year Product List Shows How Quickly Thin Tools Are Shuffled Out by Model Upgrades
In the early-2023 list, Bing, Wenxin Yiyan, Jasper, D-ID and Murf were still key products. By 2025, many had disappeared from the mainstream, while categories had evolved from the broad “article AI” into chat, search, programming, images, audio, music, video and 3D.
Jasper solved a real problem before ChatGPT: the GPT-3 API was difficult to use and users did not know how to write prompts, so Jasper packaged brand-marketing templates into a usable interface. But marketing is not a specialized field like finance or healthcare. As GPT and Claude improved, users found fewer reasons to keep paying.
In images, Midjourney and Stable Diffusion remain in use. DALL·E 2 upgraded to DALL·E 3 and then went a long time without iterating, while Kazik expected DALL·E 4. Domestic Jimeng 2.1 can generate Chinese directly, Ketu is strong at portraits, and LiblibAI and Civitai survived through communities and LoRA.
The changes support the idea that “models eat products.” If the difference is only prompt templates or shallow fine-tuning, every general-model upgrade raises substitution risk. Communities, data, workflows and user networks are more likely to retain residual value.
40. Winners in Video, Audio and Programming Increasingly Come from Chinese Teams and Vertical Products
Early video lists were almost entirely D-ID-style “talking photos.” Later, Jimeng’s lip-sync master edition, Kling, Hailuo, Vidu, PixVerse V4 and Hunyuan entered the mainstream. Kazik says Chinese video models can now “dominate the global competition.”
In audio, ElevenLabs and Hailuo replaced earlier products such as Murf. Kazik believes overseas models often sound like foreigners speaking Chinese, while Hailuo is closer to real Chinese. In music, representative products shifted to Suno, Haimian Music and Tiangong.
Programming truly took off after Claude 3.5, driving the development of products such as Cursor and Trae. Trae lowered the barrier through free access and a Chinese interface. In 3D, products such as Tripo and Hunyuan 3D emerged, and the category itself was far more mature than 2 years earlier.
Manqi noted that ByteDance appeared in almost every category on the 2025 list: Doubao, Trae, Jimeng and Haima Music. Kazik had not intentionally counted this, which made the result more revealing: a major platform can cover multiple modalities with a product matrix, expanding the competitive surface for startups.
41. AI Hardware Has Yet to Create a New Entry Point, and Companion Toys Still Fall Short of Pets
Kazik is very bullish on AI hardware and has bought many companion products, but sees a number of them as “IQ tax.” A Casio AI pet costing roughly RMB3,000 could only squirm and hum. Installing the app took 4 hours, and it did not unlock a new sound until day 4.
He admits he may not be the target user but identifies a concrete gap: pets approach, move around and trigger interactions themselves, while this device mostly waits to be touched and offers limited interaction. It does not deliver even one-third of the experience of a cat.
He also had friends research companion robots around RMB3,000 and tried glasses including Even G1, Rokid, Rokid AR Lite and Thunderbird V3. He returned the Thunderbird V3 after 1 day because its display, battery life and proactive AI interaction were still far from expectations.
Ray-Ban glasses meet a real need for people who shoot first-person video because their perspective is close to the human eye. Kazik rarely shoots, however, and cares more about waveguide displays and information presentation. Existing glasses still cannot become a replacement entry point for the smartphone.
42. As Long as the Smartphone Remains the Terminal, AI Will Struggle to Capture Ordinary People’s Time
Kazik believes people want life services and entertainment after work: food delivery, WeChat, Douyin, Bilibili and Honor of Kings satisfy necessities or kill time. Most AI products, by contrast, save time at work.
Manqi added that smartphone user time is already occupied by mature super-apps, making it very difficult for a new AI app to compete for time on the same terminal. Without a hardware entry-point shift, even a smarter model is easily trapped in existing product forms.
DeepSeek expanded the user base but did not bring a particularly distinctive change in tasks. For most people, existing chat capabilities are already sufficient. At the same time, industry resources continue flowing into math, code, research and reasoning, producing what Kazik calls increasing “elitism.”
His anxiety is therefore not about falling behind models but about content drifting away from mass needs. More important upgrades are arriving, yet ordinary users find them harder to perceive, while the “mind-blowing” content of the past 2 years has raised their attention threshold.
43. He Is Not Starting a Company Because He Refuses to Package Uncertainty as Certainty
Kazik considered founding a company at almost every stage, from the financial product in 2023 to AutoGPT-style general Agents and later multimodal opportunities. But he lacks a model-engineering background and does not want to build merely another small wrapper.
With no change in the mobile entry point, he cannot see which large market would sit outside the major platforms’ range, nor can he confirm what constitutes a durable moat. “I don’t like burning money, and I don’t want to burn investors’ money” is the direct reason he refuses to raise first and experiment later.
He describes AI startup success as shifting from mobile internet’s “20/80 split” to “99/1.” Major platforms have compute, data, talent density and capital, and can even directly poach the management layer from a startup.
The show mentioned that, by his knowledge, some interns can earn more than RMB10K a month, while the most expensive intern can “apparently” earn roughly RMB2,000 a day. Small companies cannot easily make up the resource gap with conviction. Kazik would rather remain a KOL and doorman, competing through experience, explanation and user connections.
44. AI Gives Ordinary People a New Deal through “Industry Plus AI”
Faced with grand questions about AGI, intelligence boundaries and industrial enablement, Kazik brings the answer back to employment: “The employment environment is very bad, extremely bad.” Film, civil engineering and other industries are under particular pressure, and many people see no opportunity by continuing along their old path.
In his communities, he has seen civil engineers, shipbuilders, producers squeezed out by the film industry and 985/211 master’s graduates unable to find jobs. AI creates at least one new pocket of demand where they can combine existing industry know-how with new tools.
Kazik emphasizes that the order is “industry plus AI,” not “AI plus industry.” The scarce person is not someone who only knows models but someone who understands a specific industry and can put AI into its workflow. Such a person may become the only one in a traditional organization able to take on an AI project.
Reader feedback gives him the greatest satisfaction: some entered Kimi internships after reading his articles, while others received important responsibilities in government or state-owned enterprises because they understood AI. Kazik summarizes the result as “giving some people some opportunities to keep surviving.”
45. AI Programming Is Becoming the Most Concrete Mass-Market Use Case in His Communities
After the DeepSeek boom faded, activity across Kazik’s 33 groups fell by roughly half. General Agents drew little sustained discussion; the most persistent topic was AI programming.
Users are not all in tech. The groups include educators, judges and even fruit farmers, who have begun using tools such as Claude 3.7 to build visual webpages, PDF presentations, event systems and small applications of their own.
One typical request is to upload fruit images and automatically mark which pieces have died. It does not require starting a software company, but lets industry workers “hand-build” tools for themselves for the first time, showing how programming is being democratized.
The barrier has not disappeared: users must understand code, IDEs and programming concepts, and complex projects remain well beyond the stable range of pure voice-generated output. But compared with a past in which only programmers could begin, participation has clearly expanded.
46. “Attention Is Not All You Need” Became a Quantifiable Practice on This Episode
At the end, Manqi connected Kazik’s workflow with Dai Yusen’s earlier view: Agents can liberate attention and create a scaling law for work, potentially becoming a scaling law for converting capital into productivity.
Kazik’s cases provide concrete numbers: daily information browsing fell from 5–6 hours to roughly 1 hour; 2,400 applications were screened overnight; and an 11-person, 2-day grouping exercise was completed by o1 Pro in 11 minutes.
These automations did not eliminate the human role. They gave Kazik time to record podcasts, meet industry figures, conduct real product experiences and handle high-judgment work. What was released was not “all labor,” but the portion least worth spending attention on.
47. The Investment Divide between Models and Applications Depends on Whether the Product Sits on the Model’s Main Road
Manqi summarized 2 schools of thought. One believes “the model is the product,” so model upgrades will consume Jasper-style thin applications. The other believes “the shell has its own use”: applications can combine multiple models and build stickiness through demand insight, engineering optimization and user mindshare.
Perplexity and Cursor represent the second path. They do not need to own a foundation model to make search or programming more complete than a general chat interface. Chat Fund showed that engineering and industry methodology are still insufficient on their own; the delivery economics must also work.
Manqi added that when he and Kazik recorded the episode, GPT-4o had not yet released image generation. After GPT-4o image generation opened, market feedback was explosive, but OpenAI did not simultaneously open the API or add the feature to the free ChatGPT tier, making the paid tier a temporary exclusive and directly driving paid-user acquisition.
The investment implication is conditional: applications operating in a foundation-model company’s clearly prioritized direction—image generation, general Agents and similar areas—face more severe substitution. Those occupying vertical workflows that model companies are unwilling to develop deeply are more likely to retain control.