Summary

AI is embedded in 90% of heavy users’ “study and daily life,” lifting literature-processing efficiency by 2–4x while making life without the tools “unbearably uncomfortable.” 侯泰宇（Kolento）uses GPT-4o, O1 and Claude 3.5 Sonnet to process 400–500 pages of textbooks before selecting key sections for close reading; the trade-off is less depth per text, but a jump from 20 to 200 texts in analytical range. For heavy users, models are no longer point tools but entry points to learning, on-demand tutors and interfaces for knowledge exploration.
The long-term value of education products lies not in submitting answers for students, but in providing verifiable cognitive scaffolding that shows its work. In the 12–15-year-old mathematics study relayed by 陈浩楠, students who did not use AI scored highest; when AI shifted from giving direct answers to providing step-by-step hints, the gap narrowed materially but did not disappear. Two courses that ban AI-generated papers draw the same boundary: information gathering can be outsourced, while writing, citation, rebuttal and structured reasoning still require individual practice.
AI search has become a real substitute for traditional search, cutting one heavy user’s Google and Baidu usage by 70%–80%. Perplexity uses embedded links for traceability, 秘塔搜索 can pull 50–60 or even 60–70 sources in a single research task, while Exa.ai and 博查 serve directly as search layers “for AI.” The moat is not just the answer box, but database coverage, citation granularity, and whether the result is accurate and fast enough.
Prompting ability is creating a clear generational productivity gap, but the advantage comes from workflow orchestration rather than a magic instruction. 陈浩楠’s AI-native workflow adds role definitions, few-shot examples and citation requirements, then passes Markdown to Gamma and uses Suno or Udio to generate advertising songs. Presentation was what first made “everyone look up,” while better prompts also improved the logic and completeness of business plans.
Natural-language programming is compressing previously infeasible projects into hours or days of validation. 侯泰宇（Kolento）says Cursor can generate a usable iOS App demo and complete codebase for an unfamiliar tech stack in about 2 hours; 陈浩楠, with limited programming fundamentals, cycled with Claude through “express intent—write code—return error—modify” to build an LLM benchmark for evaluating financial-formula calculations in 3–5 days. The bar for mature features is falling fast, but frontier algorithmic breakthroughs still require specialist expertise.
AI companionship has evidence of paid demand and retention, but sexual-content boundaries, minors’ addiction and psychological safety sit on the same risk curve. 侯泰宇（Kolento）says Talkie already had tens of millions of dollars in ARR last year, driven by polished official characters, IP partnerships and complete backstories; its true heavy users include people just entering puberty and users with romantic fantasies, with “soft porn being a major part.” Products must choose their position between “short-term dopamine” and “long-term happiness” while addressing the safety failures exposed by Character.AI-related suicide cases.
Willingness to pay is concentrated in high-frequency usage, memory and quotas for general-purpose models; vertical features replaceable by a single prompt will struggle to charge another $5–10. Kolento uses GPT, Claude, Perplexity and other subscriptions for roughly 5–6 hours a day, switching among O1, O1 mini, 4o and 4o mini until he exhausts their quotas; 泓君 says most of his needs are covered by Claude at about $20 a month. ChatGPT’s ability to use long-term conversation history to understand the user is also a key reason 王冉 is reluctant to move to Claude.
Models are already research “copilots,” but not scientists capable of independently asking good questions. 王冉 went from encountering fabricated papers around 2023 to using models today to build field frameworks quickly, polish English and spark new expressions, but the expected super agent still cannot independently conduct research, distill questions or generate scientific insight. Even if O1 claims to be at PhD level, the threshold for research innovation remains “the ability to ask questions”; only once that barrier is crossed might research cease to be primarily manual labor.

Deep dive

1. AI Natives Have Turned Models into Learning Infrastructure

The four guests form a generational spectrum of usage: Kolento is a third-year applied-psychology student at New York University and a product manager for enterprise AI agents; 左曜诚 studies economics and philosophy; 陈浩楠 has just completed a master’s in accounting; and 王冉 conducts postdoctoral research in brain-computer interfaces and neuroscience.
Kolento estimates that large models “run through 90% of my study and daily life,” from classroom questions, textbook interpretation and paper writing to work; if all of them disappeared, his answer would be “unbearably uncomfortable.” 左曜诚 uses them more selectively: efficiency would fall, but he would not be comparably disrupted.
泓君 cited a Study.com report from when ChatGPT was first released: around 90% of students knew about it, and 89% used it to help with homework. 陈浩楠 also recalled ChatGPT crashing during finals week; the guests broadly agreed that students, colleagues and researchers around them were all using it.

2. Literature Study Has Shifted from Page-by-Page Reading to Navigation First, Deep Reading Second

Kolento uploads textbooks, PDFs and class notes, asking models for customized explanations, review outlines and study plans, or to act as a tutor and translate material professors did not explain clearly into plain-language explanations, examples and analogies. “It’s a bit like asking whenever you don’t understand,” except the questions can come at any time.
One course assigned a large body of papers and 400–500 pages of books. The professor did not expect students to read every word, instead recommending that they first use AI to form a global view and then select points of interest for deeper reading. Kolento describes the model as more of a “compass”: it does not replace close reading, but determines where limited attention should go.
His setup initially combined GPT-4o with Claude 3.5 Sonnet; O1 entered the workflow after file uploads became available. Overall efficiency rose about 2–3x, reaching 3–4x on some tasks, mainly by saving time on table-of-contents searches, summary organization and finding related material.
He uses an “x” model to explain the change: in the past, processing 20 texts produced an absorption level of x; now he can input 200, and after AI compression still absorb x. “Understanding of each individual article may become shallower, but cross-sectional diversity increases.”

3. Academic Writing Needs AI—and a Second Layer of Tools to Detect AI

Kolento uses QuillBot to paraphrase English while preserving academic rigor and the author’s original meaning, then runs the text through GPTZero to check for AI traces. The latter, he says, is based on “perplexity” and flags both a percentage and specific paragraphs. He considers it “very accurate” and revises the marked sections accordingly.
左曜诚 does not let models write the substance. Once the content is complete, he gives it to Claude to check for content problems, then uses Grammarly for language and grammar. For him, models are end-stage quality-control tools, not writers.
陈浩楠 initially submitted code outputs directly to Claude and even asked it to write complete papers. His professor corrected him: students should first write the paper according to their own structure, then hand it to Claude or GPT for polishing. That reduces visible AI traces while preserving the author’s control over the material and argument.

4. Math and Programming Courses Are Pushing Nontechnical Students Toward Hands-On Learning

When 陈浩楠 was a senior, GPT-3.5 had just arrived and still handled accounting content poorly. Two years later, GPT and Claude had materially improved his learning efficiency. When he did not understand an econometrics formula, he asked GPT to derive it step by step; if that still failed, he asked Perplexity to find related videos or courses.
He proactively took courses in machine learning, deep learning and Generative AI. The last was titled Generative AI: Technical and Social Perspectives and ran for 3 hours per class: the first half covered principles such as attention and Transformer, while the second examined AI’s social impact on labor, education, finance and other fields.
When writing R or Python, he tells Claude what he wants, runs the code, sends back the error and repeats the cycle. “Without Claude, I might not have been able to make this at all,” because the cost of filling in all the required knowledge from scratch was too high.
The clearest result was a benchmark built from scratch to evaluate how LLMs calculate financial formulas. A student of accounting with limited programming fundamentals spent 3–5 days cycling through “I explain the intent—Claude writes—I test—Claude revises” and completed the small project. Once the results were generated, he could also hand the charts and paper to Claude.

5. Giving Answers Damages Learning; Showing the Process Only Narrows the Damage

陈浩楠 relayed a study discussed in class: in a middle-school mathematics course in a developing country, students aged roughly 12–15 were divided into groups allowed and not allowed to use AI for homework. On the final exam, the group that never used AI scored higher.
The researchers then changed AI from “giving the answer directly” to showing the steps and requiring students to solve through the process. The score gap narrowed significantly, but the no-AI group still came out on top. The key condition was that the exam itself banned AI, while real-world tasks often continue to allow tool use.
He therefore does not support a simple ban, arguing that it is akin to refusing a calculator for complex calculations. A better sequence may be to build fundamentals in elementary school, then use AI to broaden exploratory interests; once students find a genuine interest, they will have the motivation to leave the tool behind for deeper study.

6. No-AI Courses Are Protecting the Basic Muscles of Writing and Reasoning

王冉’s course, AI and Its Discontents, moves from Plato in 370 BC to the text of Attention Is All You Need and the story of 李飞飞 building ImageNet. Its central idea is “through reading these texts, generating our own ideas.” The professor allows AI for information gathering but absolutely bans it from generating papers.
The professor’s reasoning is that writing requires immersion and the calm decision of every word and the structure of the whole piece. Having generative AI complete it directly skips the process the course is actually meant to train. 王冉 therefore calls AI Augmented Intelligence rather than merely Artificial Intelligence: it should augment intelligence, not complete the course on someone’s behalf.
International Writing Workshop also bans AI because the course itself trains students to avoid plagiarism, cite correctly, state arguments and counterarguments, and respond to authors. 王冉 ultimately conceded: “These foundational courses require you to use no AI tools yourself.”
He once wrote an international human-rights law paper at Cornell Law School that received an A. When he reread it in 2022 or 2023, he felt he could no longer reproduce the legal basis or tight structure. His self-diagnosis was that he had “lost some logicality but gained more creativity”; accumulated knowledge and prompt-driven divergent questioning had jointly reshaped his capability mix.

7. Research Models Have Evolved from Fabricated Papers into Reliable Copilots

王冉 wants a super agent that can automatically research a new field and distill questions, but admits that “for now, this AI cannot do it.” Around 2023, he asked ChatGPT to list key papers and received a batch of titles that looked real but did not exist—“it was making everything up.”
At the time, Connected Papers was more reliable: starting from a known key paper, it expanded through a graph using a PageRank-like method. It was not generative AI, but suited foundational research where one needed to know “very solidly what each paper was doing.”
Models are now “much more reliable” and can link to real papers. 王冉 can ask directly about the key mechanisms behind concepts such as memory formation without first reading a full survey or review; he judges the answers to be “usually correct,” materially accelerating the construction of a field framework.
Generating an entire research paper remains impractical: without a framework and source material, the result is “extremely dry” and “says a lot of useless things.” AI built into Overleaf is more like an advanced Grammarly or copilot; it polishes his English better than he can, while in Chinese it is “not as good as writing it myself.” An occasional rewrite has nevertheless expanded his thinking about linguistic constraints into visual, auditory and other media.

8. Prompt Literacy Lets Students Orchestrate AI Better Than Corporate Executives

In 陈浩楠’s Generative AI for Business course, he worked with MBA and EMBA students and company executives over 30. He observed that many equated “AI” with GPT and knew little about other tools; their prompts were typically short and intuitive, such as “Here is the structure, help me generate a business plan.”
The AI-native workflow 陈浩楠 demonstrated added role definitions, few-shot examples, step-by-step requirements and source citations—for example, asking the model to act as a McKinsey consultant, calculate market size and explain the source of every step. His point was not that executives necessarily produce worse work, but that “logic and completeness will be better” when the prompts improve.
The workflow can also pass GPT-generated Markdown to Gamma to make a presentation, then use Suno or Udio to produce a company advertising song. During the demonstration, “the professor and students had originally all been looking down at their phones; once the advertising song played, everyone looked up.” Format draws attention; it cannot replace the idea, but it changes the effect of the presentation.
左曜诚 relayed part of an interview with 李飞飞 and Geoffrey Hinton: when Stanford was debating whether using ChatGPT to write application essays counted as cheating, 李飞飞’s son suggested admitting the 2,000 students who used ChatGPT best. He agrees with this AI-native perspective while emphasizing that even simple math can go wrong and critical checking cannot disappear.

9. AI Search Uses Citations and Data Sources to Repair Large Models’ Unreliability

Kolento first adopted Perplexity because it could index the web and provide an embedded link after each result, like an in-text citation in a paper. “It made me feel very secure,” because he could trace the source rather than accept an unattributed statement from GPT or Claude; at the time, he did not know that RAG was involved behind the scenes.
As his Chinese-language work increased, he switched to 秘塔搜索. Its regular, deep and academic research modes can call 50–60 or even 60–70 sources for a single topic, generate follow-up questions and pre-write answers to those questions, creating a continuous launchpad for further exploration.
He believes Perplexity performs worse than 秘塔 on Chinese-language material, while 秘塔 stands out on Chinese and academic databases. What matters is not which large-model API sits underneath, but “how distinctive a database it can find,” and whether it can show both the original sentence supporting a claim and its source.

10. Recommendation-Driven Search Can Hijack Attention; Topic Anchors Pull It Back

泓君’s concern is that automatically recommended related questions can resemble short-video scrolling: every item appears informative, but the user forgets why the tool was opened in the first place. Searchers seem to be expanding their world, yet may actually turn a defined question into unbounded wandering.
Kolento responds that he rarely clicks follow-up questions because most search intent is initially “fast and accurate.” Multi-round wandering is appropriate only when he is deliberately exploring divergent knowledge. Finding an official website, choosing a technical solution and conducting deep research require different answer formats.
He uses an artifact or roof-themed tour script as an example: first define the target audience and central theme, then have a large model generate an outline as a benchmark against drifting off course; next, use AI search to fill in real material chapter by chapter rather than letting recommended questions take over the route.
Each chapter is also assigned 1–3 “experts” and a host to simulate a podcast discussion. Humans interrupt when interested and pull the conversation back when it strays. The dialogues, search results and chapter materials are then fed back into the model, using its strength to “find the key points in messy articles and complex data,” rather than treating it as the sole source of knowledge.

11. Search Middleware and Evaluation Frameworks Determine Answer Quality

Flowith displays the search process as a draggable, downward-branching tree, letting users see directly “how knowledge is flowing.” That is Kolento’s main reason for using it, rather than simply reading the final summary.
Exa.ai targets AI rather than people: users can constrain a query to return only PDFs, webpages or companies, after which AI conducts deeper retrieval and organization. Kolento calls it “the middleware layer for AI search.” 博查 plays a similar role in China.
He still reviews AI results using the traditional CRAAP framework: currency, relevance, authority, accuracy and purpose. Even cited Perplexity results contain “quite a lot of hallucinations if you look carefully”; traceability does not make a source automatically trustworthy.
AI search has cut his Google and Baidu usage by about 70%–80%. Even if traditional search places links on the left and an AI answer on the right, that is not necessarily an advantage; existing user habits help, but the final choice still depends on whether it is “accurate enough, fast enough and able to answer my question well.”

12. Cursor Upgrades the Code Assistant into a Full Project Generator

Cursor is the tool Kolento uses most at work. Copilot in VS Code can already assist with coding, but still requires developers to type and accept local suggestions; Cursor turns a natural-language requirement directly into a complete set of files and systems.
When faced with a completely unfamiliar programming language and wanting to write an iOS App, he once had to start by taking a course from scratch. Now, after roughly 2 hours describing the requirements and goal, he can get a usable demo for rapidly validating an idea. “At first I didn’t really believe in its results,” but after using it several times he was “won over” by its accuracy and usability.
This does not mean professional barriers have vanished. Common, established functions can be handed to Cursor, but new algorithms and frontier breakthroughs still require deeper technical capability.

13. The Gap for Serious Companionship Lies Between Emotional Support and Clinical Treatment

Kolento divides companionship into serious and non-serious categories. He sees Pi as a voice chat model that is relatively strong at emotional understanding and expression. Many classmates also use ChatGPT to process emotions because it is generally fair and calm while offering both emotional validation and practical advice.
Wysa combines chat with cognitive behavioral therapy, suggesting mindfulness and meditation or providing around 6 minutes of music. The process remains highly fixed, however: the advice for “level-10 unhappiness” and “level-3 unhappiness” may be similar, and its conversational intelligence is clearly below GPT’s.
This creates an awkward position: for people without a diagnosed mental illness, talking to GPT may already be enough; people with a diagnosis cannot be cured through Wysa and still need a therapist. It adds another mental-health intervention tool without delivering personalized treatment.
Kolento recalls that 吴恩达 once tried a similar cognitive-behavioral-therapy chatbot, but did not release it directly to the public. Instead, it was developed with psychologists and used as part of patients’ treatment between sessions. He sees this structure—clinical collaboration plus daily AI exercises—as a possible direction.

14. Entertainment Companionship Monetizes Character, Interaction and Fantasy

Character.AI, Talkie and Replika let users chat and fall in love with virtual characters, whose profiles include backstories and scene descriptions. Talkie goes further by letting users help characters complete NPC-like tasks, using plot progression, intimacy levels and unlocked content to deepen immersion.
Kolento particularly likes “逗逗游戏伙伴”: a cartoon desktop character that can be patted on the head or poked in the waist and remembers what the user ate the previous day. When playing the web version of Genshin Impact, it can also read the screen, provide tutorials or comment on gameplay. The developer is the Chinese company 心影随形, which, according to the program, previously received investment from 源码、高榕 and 九合.
Talkie belongs to MiniMax; Kolento also says Character.AI has been acquired by Google. He sees the difference as one of supply: Character.AI relies more on user-created bots, while Talkie provides large numbers of polished official characters and IP partnerships, such as interacting in the voice of rapper 糯米.
More complete appearances, family histories, relatives and life choices make Talkie characters “more like a person.” According to Kolento, the product had tens of millions of dollars in ARR the previous year. Talkie users are not limited to anime and manga fans, while Character.AI’s user profile skews more toward that demographic.

15. Companionship Products Must Design Desire, Retention and Safety Boundaries Together

Kolento is blunt: “Soft porn is a major part of it.” Real friends or partners can satisfy ordinary social needs; the greater value of virtual characters is “fulfilling fantasies that real people cannot fulfill.” If restrictions are too tight, the role-playing demand also weakens.
Based on user research he has encountered, people with mental-health conditions or more introverted personalities use the products more deeply. The deepest users are boys and girls just entering puberty who fantasize about love but have not yet had opportunities for real-world relationships.
The first product opportunity lies in interaction: not only text, but also mouse input, images, voice and video. Kolento cites the popularity of characters from Love and Deepspace making unsolicited phone calls, and divides the positioning into “short-term high dopamine” and “long-term happiness”—the former closer to sexual desire, the latter closer to family affection.
The risks are more complex than ordinary game anti-addiction controls. He mentioned a case in which a boy used Character.AI and was encouraged by Character.AI to commit suicide, causing a major public impact. Platforms must avoid restricting characters until they lose their appeal while also preventing self-harm inducement, excessive addiction and erosion of real life.

16. High-Frequency Usage and Quotas Drive Multiple Subscriptions, while Lightweight Vertical Tools Struggle to Charge

Kolento pays for GPT, Claude, Perplexity and multiple enterprise-grade agent tools; some costs are reimbursable, and he admits he often forgets to cancel subscriptions. He spends about 5–6 hours a day in AI-enabled learning or workflows. His chats are for knowledge exploration and divergent thinking, not companionship.
Heavy usage burns through tokens quickly: once O1 is exhausted, he switches to O1 mini, then 4o and 4o mini; sometimes “everything is used up,” followed by a cooldown lasting several days. Claude offers relatively more capacity. Multiple subscriptions make A/B testing convenient and allow him to switch models when one quota runs out.
陈浩楠 subscribes to ChatGPT, Claude and Perplexity Pro, using them about 1–2 hours a day normally and 4–6 hours during finals. 泓君 says most of his needs are covered by Claude at about $20 a month, with free products preferred for everything else.
Simplify and Jobright can recommend jobs based on a résumé, automatically complete applications and revise résumés against a job description. But 陈浩楠 and his classmates find it difficult to pay another $5–10 a month for this single function. “You can just use ChatGPT—give it a prompt,” which is the most direct substitution pressure on vertical AI tools.

17. Long-Term Memory Creates Stickiness, while Scientific Innovation Still Stops at the Question

One reason 王冉 remains with ChatGPT is that it remembers past conversations and understands him. He once prompted it, “Tell me something about myself that I might not know about myself.” The first answer did not exceed his self-knowledge; only after repeated questioning did it produce details that could be explained by his prior experiences.
泓君 raised the Barnum-effect objection: could this be like astrology, where anyone can find something that fits? 王冉’s response was that the model can identify which old conversation a conclusion came from. It has not fully solved the problem, but “it can be useful.”
左曜诚 originally did not discuss philosophy with AI because it was “just a tool,” without consciousness or real intelligence. Kolento suggests understanding generative models as “the average of human civilization.” 泓君 pushes further: are their works merely averages, or can they surpass the best in a field as AlphaGo did?
王冉 notes that O1 claims to have reached the level of a PhD, but he believes the most important part of research innovation remains “the ability to ask questions.” Models still cannot independently produce good research prompts, academic insights or judgments about the direction of technological progress. If they eventually can, research may enter “a completely new era” and cease to be primarily manual labor.

January 7, 2025 Mike Maples: Three Frameworks to Evaluate Startups and Founders | E1242