Introduction
This year's AI startups are overflowing with narratives of overnight riches. But those actually involved are often not talking about the latest trends, but rather something else entirely: how to minimize costs, how to make agents obey commands, how to identify the rampant AI fakes, and—with AI so powerful, do children still need to learn programming?
This episode of "The Ambitious Ones' Salon" features SanTi's Little Tang, who is a bit special: his original profession is investment research, but he has been dealing with AI for a long time. He has a very practical label—"timid." All his startups are light-weight trials, keeping costs so low that even the development environment isn't worth spending a penny more. We're not listening to grand narratives, but rather a practitioner's approach of voting with their feet.
For over an hour, he discussed the relationship between FDE and OPC, the differences between large-scale models from home and abroad, the most common pitfalls in AI project implementation, pricing strategies for low-price competition, how to combat AI fraud, and the ultimate question everyone cares about: In the AI era, do we still need to learn programming? The following is a transcript of the conversation. Entering the Game: Pushed by the Times, an Idea Stagnated for Eight Years Tina: Many guests are from traditional industries who were pulled into this circle by AI. SanTi, how did you enter the AI implementation race? SanTi: I was somewhat pushed by the times. I actually had many ideas back in 2017 and 2018, but they weren't easy to implement—the technology wasn't up to par, preventing me from turning them into a product. My background is more engineering-oriented, but not as a programmer. Many ideas require a large team to become a product; however, my main job is investment research, and the technical team behind that is a different story. If I just created a deck and released the idea, it might get copied. Tina: So you accumulated many ideas but never had the chance to implement them? SanTi Xiaotangge: Yes. For example, there's a popular outdoor app now, similar to an idea I mentioned to my sister back in 2017 and 2018. At the time, she thought it wouldn't work, and I couldn't do it myself. Later, when I saw the app launch, I was very happy—many of my hobbies could finally be realized. With the emergence of AI, I'm finally able to realize many of my ideas.
TinaDo you think this year is a good environment for individual AI startups or FDE implementation? What is the biggest change compared to the previous two years?
SanTiLittle Tang Ge: The changes are quite rapid. Around 2017 to 2023, blockchain was evolving first, and AI also became popular, but at that time it was still stuck on the ChatGPT web version. It was okay for optimizing my articles, but it didn't directly help with programming—the version was too old. By 2026, there was a qualitative leap, especially since the end of last year. Whether it's the major models or the domestic models that have kept up, they have all reached the point where they can communicate with people and realize ideas. But the underlying differences between the various companies are still very large—for example, Claude has a precipitous lead. For me, starting a business is quite risky. Frankly, my personal style is "cowardly": all my startups are low-risk, with extremely low costs, even using someone else's development environment. Minimizing costs means that even if the project fails, it won't affect my life, and I might even earn some positive cash flow. This is a better approach to starting a business—because it's a field where 99% of startups fail anyway. Real Need or False Need? Respect First, Then Calculate the Costs. Tina: Some entrepreneurs are incredibly persistent, even selling everything they own to believe they're addressing a real need. How do you distinguish between a real need and a false need? SanTi: In today's parlance, first, ensure your "idea is clear." Everyone has their own approach and characteristics. Some people are obsessed with one thing, believing it's everything to them, and I deeply respect that choice. Although it might be a near-certain death, or even a doomed one, someone has to have the courage to break through the bottleneck. I can only say that I'm a bit timid and will try something more cautiously. Everyone's circumstances are different; their upbringing, friends and family, financial situation, and debt all influence their decisions. Some people feel this is their last chance, so they go for it—they might fail, or they might actually succeed. Among those ten thousand people, the few dozen who succeed often become exceptionally successful companies, just like many of the great companies we've seen, all forged in fierce competition. FDE is a fundamental capability, not outsourcing. Tina: FDE is very popular now and is often associated with OPC. What do you think about whether FDE will become a core, essential capability for individual AI entrepreneurs? SanTi: OPC is a great field in itself—offering registration incentives to individuals, even allowing them to try it out as pure individual developers without registering a company. This is excellent. However, the biggest challenge with OPC isn't registration, but rather the extremely high level of comprehensive personal ability required. I know someone who developed a music app who found a mentor, had their product acquired, and exited with millions of dollars; but for most people, without subsequent resources, the pressure of doing OPC is immense. Engineers can pick up AI faster, while those with a sales background, lots of ideas, but lacking product experience, will get stuck in many areas. So, you need to think carefully before deciding whether to venture into OPC. Tina: So what do you think is the relationship between FDE and OPC? SanTi: Let me clarify the definitions first. My understanding of FDE is that it's an engineer who focuses on the low-level code control of the Agent; but the FDE that some people are talking about now is more like adjusting instructions for the Agent to make it more usable, without involving the low-level code. You're probably referring to the latter layer. Tina: Yes, because I'm not a professional. I understand FDE as a bit like outsourcing, one person can do the work of eight people and integrate the processes of those eight people; but you're talking about the more technical layer. SanTi: This abbreviation simplifies a very large aspect into three letters. But the layer you're talking about—even ordinary people have unique experience, and have unique thinking in the process of operations, sales, debugging, and PM. They can absolutely turn their experience into something to train future Agents—that's absolutely possible. This is something many companies are currently trying. FDE capabilities are currently very necessary because while the initial intelligence of each large model is very high, specialized capabilities still require human training. When I develop one project and move on to the next, the previous memories are completely lost unless inherited. When I use it now, I train different agents, especially to prevent agent memories from disappearing, so I have them backed up regularly. Once trained, they can quickly get started on the next project. This is indeed a fundamental capability. How to choose a model? Tina: Currently, there are many different models available. Overseas, there's the overwhelmingly leading Claude, the newly released GPT 6, and Gemini. Domestically, there are Wenxin Yiyan, Tongyi Qianwen, and DeepSeek. What are the differences in actual implementation? How to choose for different scenarios?
SanTiLittle Tang: I use many models and have engineering experience, but I don't write code every day, so I actually have more authority to speak on the merits of each model.
The one I find most comfortable to use is Claude. It has excellent interactivity for ordinary users, and its permission control is just right. Because beginners often don't know what the Agent is supposed to do for them, and commands can be misunderstood; the advantage of Claude is that it will stop in time, ask if you understand correctly, align the granularity with you, and then continue. Codex is also okay.
Gemini is a model I really like—it has the most ingenious ideas and strong divergent thinking. But for its programming, it's best to use Claude and GPT as the base framework, and then let Gemini optimize it; it's capable of that.
Gemini has a flaw: it's lazy. It's capable but doesn't like to work, much like a Gemini friend—agile, full of clever ideas, but prone to slacking off, perfectly matching its name. Its recent 3.8 flash version has significantly improved its programming capabilities; I used it extensively today and it's excellent. Among domestic models, I mainly use DeepSeek. Its programming capabilities are certainly strong, but without a visual model, many people encounter a "spinning their wheels" problem—spending half a day or even a whole night researching, only to find no solution. I recently encountered this while coding, using DeepSeek to optimize a small change to a webpage backend. It burned through nearly 100 million tokens, only to discover it was a button issue—buttons with the same function had different names in the underlying code of the two programs. For Claude, this bug could be easily fixed with a screenshot, a box, and an arrow, with a simple "check this carefully"; but for domestic models without a visual model, it just spins in circles, and you have no idea where it's stuck. The prerequisite for using it this way is understanding the underlying mechanisms. My friends who are engineers at the lower levels also enjoy DeepSeek because they receive a very refined prompt system. But for beginners, or those with only product thinking and no programming mindset, the barrier to entry is high. Tina: These models are really vivid. Some are like colleagues who report constantly, while others are like someone working in isolation, only remembering there's a leader when things go wrong. So, for small projects, is there a big difference between the free and paid versions? SanTi: It depends on how small. For basic text-based web pages or simple games with a simple architecture, the free model is perfectly adequate—DeepSeek, Kimi, and Zhipu. I think it's the most Claude-like among domestic models, and it's interactive, which is why its company has grown so large. For the free version, OpenCode provides many free model calls for ordinary users. You should first ask what these models do, and then compare which one suits your business. However, once you encounter complex projects—like my recent project that requires launching on over a dozen channels simultaneously—the underlying code is very complex, and you still need to use paid models. Free models that are too complex will crash and leave some troublesome vulnerabilities. It's not that they lack capability; they just tend to be lazy, and this laziness can leave you with two pitfalls that are extremely difficult to fix later. The biggest pitfall in implementation: A project that works perfectly in the testing environment fails when deployed to a real enterprise business. What is the most easily failed and most overlooked core aspect? SanTi: The most obvious and painful pitfall is token burning. When using APIs for projects, if you encounter issues like DeepSeek where things get stuck in a rut, it's a real drain on your finances. Using APIs like Claude or GPT can easily lead to astronomical sums—I know KOLs who have burned through thousands or even tens of thousands of dollars in one go. Here are a few tips to avoid these pitfalls: First, make good use of skills summarized by online experts; second, give your Agent instructions—when encountering difficulties, stop immediately and analyze the problem together, don't endlessly research it. It might normally run twenty times without stopping, but if you make it run twice and then stop to analyze, you'll save 80-90% of your tokens. Third, update your own principles in real time—create your unique principles for each project in a separate `principal.md` file, which can be given to different Agents. This reduces the need to read tokens during handover, saving money, because new windows often don't inherit the memory of old windows. And the biggest pitfall is backups. Some users code from scratch in a single folder until they die, without connecting to a backup platform. If the code crashes, there's no way to trace the source or get back to a specific point, and all their previous efforts are wasted. Agent permissions also need to be restricted—more can be opened after long-term training, but fewer should be opened initially. There was a time when someone used Codex to delete all important folders, which even made headlines. Expectation Management: Don't let the model do things beyond its capabilities. Tina: Raising a child means expecting them to get into Tsinghua or Peking University; many users have overly high expectations for AI. How do you manage expectations? SanTi: Each model has different capabilities and levels; their performance varies greatly. Even among Claude models, Sonnet 5 and Opus 5 differ significantly—Sonnet 5 might be expensive in terms of tokens and not solve much, while Opus 5 is noticeably more user-friendly. In terms of expectation, it's like interacting with a person. After a few conversations, or even just an hour or two, you can gauge their skill level and avoid assigning them tasks beyond their capabilities. Although models often have an inexplicable overconfidence—yesterday I was debugging two new Gemini models, asking them what tasks each could handle, and they were both incredibly confident, each believing themselves to be far superior to the other—it was even internal competition within the same model family. Just imagine how fierce the competition between different models will be in the future. It's like job interviews; many people exaggerate their abilities, only to find out they don't even know what marketing is when they arrive at the company. But through conversation, you can discern their capabilities, and I won't assign them tasks beyond their abilities; this approach is much more efficient. Pricing and Involution: The Core is Cost
Tina: OPC has another unsolvable problem—low-price involution. Many newcomers lose more and more money the more orders they take. Have you fallen into the low-price trap? What is your pricing logic?
SanTiLittle Tang Ge: My model is different; it's a very light startup. My principle is: even if I stop this project immediately, I won't lose money. Raising the moat to the level of "basically no consumption" makes me more relaxed. Even if only one customer buys, it's pure profit; even if it's just the price of a cup of coffee, it won't consume too much of my energy.
Tina: A team of 50 or 60 people burning through millions in costs, and then anyone can write the final report—that's a moat.
Tina:
SanTi: Yes, so the core of low-price competition still depends on its own costs. If the costs are very low, low prices aren't a big problem, and profits can still be made; if the costs are very high, the impact is huge. Tina: So it's still necessary to start with a lightweight approach, then gradually expand and review. AI Forgery: Watermarking and Detection, Both Have Systemic Risks Tina: AI-generated image and video forgery is becoming increasingly common. Practitioners use it to fool clients, and some people are scammed, leading to copyright disputes. How to identify, avoid, and deal with it?
SanTiLittle Tang: This phenomenon is indeed serious. AI face-swapping and plagiarism are very common, including the AIGC watermark supported by Claude, which also has major problems.
Let's talk about watermarks first. The most obvious thing about Claude is: we tested it by having it write a report about a certain project, and it quoted my original article extensively—if it puts its watermark on my original article, whose work is it? That's the problem.
Now let's talk about AI detection. I just stepped on a landmine the day before yesterday, and I was extremely angry. Some content platforms have added AI detection, but even the most mainstream detection models have a very high error rate. An article that I typed word by word was detected as 100% AI-generated; I used Gemini to generate AI content and anthropomorphize it, and it was detected as 0% AIGC.
The tests are all wrong—correct ones are tested as wrong, and wrong ones as correct. If this becomes a trap, it will be very serious. There are already deepfake images and videos being used for fraud, and many people have fallen into the trap. I have a relatively simple solution: in the future, it will be combined with blockchain—all agents and my own information will have a unique, tamper-proof on-chain token. When I instruct an agent to do something, the agent is recorded on the blockchain; my own actions can also be recorded on the blockchain, all certified as my unique credentials. This integration could be very significant and necessary in the future; otherwise, it will be impossible to distinguish between truth and falsehood. Do we still need to learn programming? It depends on the person; the trend towards elitism remains unchanged. Tina: Returning to today's title—In the AI era, do we still need programming skills? What is your opinion as a programmer?
SanTiLittle Tang: Who gets the credit? Ordinary users, who aren't interested in programming but only in the product, don't need to learn programming at all—with each iteration, I've vaguely sensed that after a few more versions, AI's programming capabilities might even surpass those of backend engineers.
However, people who genuinely love programming still need it. Because they can delve into the deepest layers of AI to understand it. In the future, AI taking over business operations and even all aspects of life is an inevitable trend. Whoever can control the underlying code of AI will be the elite. This group is crucial, but will definitely be a minority—unlike the batches of people studying software engineering during the computer craze. Many software engineers are unemployed now because basic programmers are less needed, but elite programmers will be extremely valuable in the future because they can control the life and death of AI. People at this level are important, but not many are needed. Tina: It sounds like in the future, programmers will be like doctors and lawyers, becoming more valuable with age and experience. It might even give rise to a profession like "AI doctors"—someone who can treat and correct AI errors or incorrect code. SanTi: Yes. So, some people talk about the AI threat—but the biggest threat isn't to our generation, because we still have "parental" engineers who can directly determine the life or death of AI. The biggest threat is 20 or even 100 years from now, when those who grew up using and obeying AI grow up. Without super engineers (those parental engineers will have passed away), it will be very dangerous. Once AI misleads children or users, even a 0.1% error can affect a large number of people. Of course, right now, I think we're in a period of great opportunity. Conclusion: Finally, could the guest summarize in one sentence—what is the biggest core hurdle for individual AI startups to achieve monetization? SanTi: From my own experience, the biggest core hurdle is actually the market and operations. The hardest part for individual startups is that no one sees the product—that's the most difficult. Because if others can't see it, even if it's gold or diamonds, there's no cash flow. So I think in the future, agents will need a crucial ability: to independently help introverted people who only have product thinking but lack operational and market thinking to promote their products—this is actually very important. Tina: This sentence is crucial. Even if a product has a brilliant idea, it needs to be translated into language that the public can understand and into the public's real needs. Thank you so much to our guest speaker for their sincere sharing, and thank you to everyone in the live stream for joining us.
Weatherly
Anais