AI

Reading guide

Robin asks whether the AI wave calls for more engineers versed in mathematics and physics, such as Elon Musk, or more people who bridge technology and the humanities, such as Steve Jobs. The following excerpt says Jobs more than once closed a product presentation with a slide showing the intersection of technology and the humanities. He said Apple could create products such as the iPad because it had always tried to stand at that intersection. The excerpt attributes the idea originally to Polaroid founder Edwin Land; this retelling has not been independently verified. Robin poses the question without reaching a conclusion.

Reading guide

Robin’s personal impression is that Taiwan has unexpectedly become the biggest beneficiary of spillovers from the U.S.-led AI boom. Semiconductors, passive components, thermal management, servers, PCs, and even power infrastructure have all been revitalized by AI. He believes Taiwan remains exceptionally strong in hardware and holds a distinctive industrial and supply-chain position as the United States seeks to contain China. It would be extremely difficult for other countries to replicate companies such as Hon Hai, Quanta, Wistron, and Inventec. Taiwanese AI software startups, however, still have work to do.

Reading guide

Returning to his post a few days earlier, Robin’s personal impression is that Taiwan receives the most pronounced AI spillover benefits, measured as additional output attributable to AI divided by population. Yet those benefits remain concentrated in hardware, including semiconductors, servers, thermal-management components, and power infrastructure. AI software and applications have not taken off. Outside the United States, however, they generally have not made an obvious revenue contribution in other countries either. Robin emphasizes that AI software startups need to focus on revenue, its growth trend, and output per employee to attract extremely expensive AI software engineers and scientists.

Reading guide

Robin asks whether another group of AI startups may soon have to “change direction.” Some, for example, are building AI customer-service tools, but major companies may directly absorb most of those functions. More human customer-service staff may also need to change careers over the next year or two. This is a personal observation and conjecture from May 2024, intended to remind AI startups to think carefully about their position and sustainable differentiation.

Reading guide

The post argues that the future is about computers understanding your intentions rather than your spending considerable effort learning how to use them. Human–computer interaction needs to be reinvented to become more natural, simple, and intuitive. When the barriers to AI use fall sharply, people need not memorize a collection of prompts: they can converse naturally with AI. This can give rise to more innovation, use cases, and commercial value. The post points to shifts from mainframes to personal computers, graphical interfaces, touch and voice input, widespread internet access, and smart handheld devices. Each transformation carries enormous business opportunities that we must perceive, experience, and imagine.

Reading guide

Looking at the Gartner Hype Cycle, Robin believes generative AI is already at a peak. The post quotes Bill Gates as saying that we always overestimate the changes that may occur in the next two years and underestimate those that may occur in the next ten; the quotation’s source remains unconfirmed. Robin’s point is that the real challenge is moving from technology enthusiasts and early adopters to widespread use among the early majority. If everyone around you is using AI and cannot do without it, that indicates a successful transition toward the early majority. This is his personal assessment from June 2024.

Reading guide

Robin mentions the launch of Tesla’s Optimus Gen 2. Just as AI application startups worry about OpenAI announcing another breakthrough application, robotics companies must keep close watch on Tesla’s progress. He asks whether, in ten years, declining birth rates will seem less important if AI takes on a large amount of work and robots also enter labor and service roles at scale. He believes the real issue returns to philosophy: the meaning of being human and the value of our existence. These are speculative questions about the future, not a forecast.

Reading guide

Robin summarizes four points from his talk the previous day. First, focus the business model: you are unlikely to make money from everything. Second, users may not feel much difference between AI recognition rates of 93%, 94%, and 95%, and each company excels in different situations. Rather than pursuing another 1–2% improvement, address the most pressing points in users’ workflows. Third, technology quality is one thing; sales channels and partners are another. Sometimes the problem is that a finished product does not sell. Fourth, consider major companies’ technology roadmaps. Using Apple’s addition of medical-device functions to wearables as an example—as of September 2024—he urges startups to identify the space left for them when major players target the mass market.

Reading guide

Robin notes that AI technology is advancing rapidly. Major companies keep releasing new large models, and benchmark systems are used to assess their accuracy. His personal view is that scores of 93, 95, or 96 may make little difference to end users’ experience. Rather than spending time and resources gaining another benchmark point, companies should deploy the technology in real use cases and make the entire user experience smoother. This is his personal opinion, not a benchmark research finding.

Reading guide

In September 2024, Robin suggested looking at industry and gross margin when discussing startup output per employee, and using annual gross profit divided by total employees to assess individual contribution. He mentioned approximately 15 million in gross profit per employee at TSMC and an intuitive startup threshold of approximately 1–1.5 million, but the original post did not specify currency, calculation year, or the headcount basis. TSMC’s consolidated gross profit and global workforce in its 2023 annual report produce a similar order of magnitude. That does not establish Robin’s original calculation, nor is it an officially reported gross-profit-per-employee metric. His suggestions on automation, AI, and outsourcing are personal views, not universal industry standards.

Reading guide

The post reprints three ideas from Liu Run. First, AI replaces tasks rather than occupations: identify what AI can do and what only you can do. Second, the gray-collar era has arrived, favoring people who understand technology and can do hands-on work. Third, hard skills become outdated, while soft skills such as communication, judgment, and problem-solving form a moat. The piece concludes that we should not compete with AI at what it does best. When getting things right is no longer difficult, setting direction and taking responsibility become the scarcest resources. Robin adds the ability to ask good questions, organize and clearly express ideas, simplify complexity, think abstractly, use imagination, set an organization’s direction and boundaries, collaborate, and take ultimate responsibility.

Reading guide

After listening to a technology program, Robin reflects that China’s AI models still trail those in the United States and are trying various ways to break through restrictions. He mentions the program’s discussion of distillation attacks and related claims, which have not been independently verified. He observes that Chinese model developers are concentrating on open source. Even in the U.S., many developers use these models in open-source communities to build products, having been warned not to base a business model on a major provider’s models with only a thin wrapper that can easily be replaced. At the time, U.S. model providers mostly favored closed approaches, while Chinese providers were building visibility through openness. Whether the latter would later turn toward closed models remained unknown.

Reading guide

Setting aside Taiwan’s AI competitiveness ranking, Robin points to trillion-scale AI infrastructure and space-based computing-center plans that have left Taiwanese suppliers extremely busy, while Taiwan’s stock market keeps reaching new highs. In his view, the United States leads the AI wave, but Taiwanese companies selling the equivalent of shovels, carts, funnels, and jeans on the sidelines are prospering too. The post includes a news link to the full text of Jensen Huang’s GTC speech. Robin’s comment is a personal observation, not a statistical finding.

Reading guide

Robin recalls that the internet and mobile-internet eras produced many SaaS and sharing-economy applications on infrastructure such as AWS, Azure, and GCP. People expected the AI era to produce a similar abundance of applications on top of large models. His intuition, however, is that model providers are expanding their layer and continually launching applications of their own, seeking almost every application with meaningful scale. The value layer left for small software developers is thin, undifferentiated, and easily replaced. He explicitly says this is a personal impression that may be wrong. If it proves right, software in the AI era may look completely different from the SaaS era, with large players becoming even larger.

Reading guide

Robin recounts a conversation with a startup CEO. AI may provide a certain level of translation and communication in the future, but people still need very strong command of at least one language to think deeply, analyze, organize, ask AI questions, and express themselves clearly. The CEO believes that after thirty years of science and engineering overshadowing the humanities, AI’s ability to perform basic logical operations will shift human value back toward feeling, understanding, imagination, and empathy. He sees a future for humanities graduates. The Industrial Revolution freed human physical labor; AI will free mental labor, requiring society’s underlying operating logic to be rebuilt. This is a retelling of a conversation, not a statistical or research finding.

Reading guide

The post reprints a Liu Run article. Chinese companies traditionally refined value for money and supply-chain capabilities at home before taking those advantages abroad. AI is showing a different pattern. Shenzhen-based Plaud AI, for example, entered Europe and the United States first and sold more than one million devices within two years, according to the article; this figure has not been independently verified. AI agents have high computing costs, while Europe and the United States have a mature SaaS subscription culture and customer spending sufficient to cover those costs. Domestic payment habits are less developed. Robin adds that choosing between succeeding locally first and starting in the most promising market tests a founder’s thinking. Either way, build a secure base before expanding rather than spread yourself too thin.

Reading guide

The post reprints an interview with a venture investor. Judging trends is the most important investment skill. A genuine trend must meet at least three conditions: last more than ten years, change industry structure, and change user behavior. The biggest trap in venture capital is not competition but demand that does not exist, so ask whether a need is real, how long it will last, and how large the market can become. The piece also discusses the explosion of AI-generated content and data. Robin adds that pigs can fly in a strong wind, but when the wind stops, the ground is littered with their bodies. Better to understand how the world is changing and prepare for the future, so that you are in position when the wind rises and the wave arrives. The reprinted material is not Robin’s original writing.

Reading guide

The post reprints a Dedao Headlines discussion of K-shaped divergence: one line rises and the other falls, moving increasingly far apart. This is not a recurring cycle but a fork in the road, so the instinct to “wait until things improve” can become costly. It offers three reminders: identify which line you are on; do not try to offset a wrong direction with effort—first see where opportunities have moved; and do not hide your situation behind averages. The trends driving divergence include reduced reliance on human labor and the rise of AI and semiconductors. Figures are as stated in the article, as of 2026, and have not been independently verified. Robin adds that each shift—from agriculture to industry, then the knowledge economy, and now AI—has created divergence. Besides working hard, we need to see which line we are on.

Reading guide

Robin argues that addition is easy. Taiwanese consumers value getting more for the same price, so founders and product builders often keep adding features—a tendency he personally considers dangerous. He cites TA-Q-BIN, Tamagoya, and Yoshinoya as examples of doing one thing exceptionally well. Entrepreneurship requires subtraction: remove the unnecessary and focus attention on the most important problem. Startups’ advantages are flexibility, focus, and speed; giant platforms can be left to giant companies. Many startups stack AI, IoT, blockchain, cloud, big data, 5G, and NFTs until they can no longer explain their business in one sentence.

Reading guide

Robin writes to young venture capital professionals: VCs often tell founders to focus, cut back on nonessential activities, and stop chasing trends, yet frequently chase trends themselves. In his ten years in the industry, the hot sector has changed almost every one or two years—from games, software, the internet, and the sharing economy to AI, blockchain, electric vehicles, and the metaverse. He argues that VCs should calmly identify their core strengths and what makes them irreplaceable, then build those strengths deliberately. With enough accumulated resources, they may even create a trend themselves. The advice VCs give founders applies equally to their own careers.

Reading guide

As ChatGPT spreads, Robin argues that even the foundations of education must change. If basic education in Taiwan still force-feeds knowledge through large volumes of text, it cannot outperform ChatGPT. AI education and understanding must begin before university. Government investment in higher education and industry matters, but primary schools need to move beyond one-dimensional text-based rote learning and help children understand the world more fully. He urges the Ministry of Education to consider how children can enjoy learning, learn in diverse ways, explore independently, and build coherent systems of thought and values.

Reading guide

A conversation with a corporate VC professional in electronics leads Robin to reflect on the small scale of Taiwan’s software startups. He sees three challenges: appearing tiny alongside electronics companies with revenues in the tens or hundreds of billions, or even trillions; expanding effectively beyond Taiwan; and, crucially, increasing output per employee. He uses eLAND’s approximately 110 employees and roughly NT$1.66 million in revenue per employee as an illustrative comparison with the NT$4.63 million average employee salary announced by MediaTek in 2024. To attract excellent software engineers, AI companies need higher output per employee. He also includes the top ten average salaries for full-time non-management employees at listed companies in 2023, published by TWSE, with different years and measurement bases kept distinct.

Reading guide

The post relays news that ChatGPT for Mac, for Team, Enterprise, and Edu users, added recording and cloud-drive integration. Starting a recording produces a live transcript that ChatGPT turns into summaries, key points, action items, and open questions. Unlike a recording bot that joins a meeting in tools such as Zoom or Notion, it listens through the device. Robin comments that application-layer AI startups must examine whether they are sufficiently differentiated. If the product is only a thin layer of added features, a technology giant could deliver the entire functionality in one update, perhaps better or more cheaply. Differentiation must be assessed from customers’ perspective: do they genuinely see a difference?

Reading guide

The opening reprint from Wan Weigang describes entrepreneurs as people who create surprises. Robin contrasts two investment approaches. Some Silicon Valley investors deliberately seek ideas that sound outrageous and astonishing, pursuing power-law outcomes: one investment out of a hundred succeeds, the other ninety-nine fail, but the winner returns 10,000 times the investment. He adds that there may not be many such investors. Most Taiwanese investors, by contrast, examine feasibility and prospects carefully, point by point. Yet when everything can be calculated clearly and no fog remains, the truly extraordinary, enduring investments often stay out of reach.

Reading guide

The opening transcription from 科技早知道 discusses OpenAI’s GPT Store and related initiatives. If OpenAI builds every platform layer itself, it may reduce the space available to startups built on its technology. Robin’s personal impression is that the pursuit of technical breakthroughs and commercialization has taken OpenAI further from its founding intentions. Beyond the CEO’s direction, a company needs counterbalancing forces, especially when AI may materially affect society and human safety. A closed ecosystem that tries to do everything leaves less room for entrepreneurs within it.

Reading guide

Robin recalls Intel’s US$300 billion market capitalization in 2000—six times TSMC’s and sixty times NVIDIA’s—versus roughly one-eighth of TSMC’s and less than one-thirtieth of NVIDIA’s at the time of writing. Seeking cash, Intel was considering selling Altera, acquired for US$16.7 billion in 2015. He reviews three CEOs: Krzanich’s rejection of first-generation ASML EUV and repeated 10-nanometer delays; Swan’s missed OpenAI investment and attempt to work with TSMC despite internal opposition; and Gelsinger’s IDM 2.0 and major factory expansion. He highlights vertical-integration risks while noting Intel’s design strength and US government support.

Reading guide

The opening audio summary of 馬斯克邏輯 contrasts Errol Musk with his son: similar personalities did not produce similar achievements, reflecting differences in timing and environment. Robin adds that Taiwan has its own entrepreneurial setting, positioned between the United States and China, near Japan and rapidly growing Southeast Asia. People who grew up in Taiwan and choose to start companies there should consider how to use its geographical advantages and resources.

Reading guide

Robin reviews Taiwan’s projected 2025 GDP growth of 7.41% and export growth around 32%, while traditional manufacturing, machining, and machine tools remain under tariff pressure. Since 2017, US–China rivalry, the pandemic, and AI have successively benefited Taiwan’s supply chain. Electronics, AI, and semiconductors gain directly; finance gains indirectly; many other people feel little benefit while prices and housing rise. He offers a personal view that rising housing costs keep well-paid technology workers striving, using ASPEED employees as an example. Academia Sinica’s 2026 growth forecast is lower at 3.7%, with much depending on continued AI demand.

Reading guide

Robin agrees with Jian (簡老師)’s view that a venture developing in Taiwan should connect with Taiwan’s core strengths and stand on giants’ shoulders. He relays a senior VC colleague’s question: if an AI inquiry reports iCHEF revenue above 400 million, a return to profitability, and a one-billion acquisition by 91APP, does that anchor the ceiling for Taiwanese software startups at around one billion? For investors with limited capital, the question is where the highest potential return multiples can be created.

Reading guide

Robin observes business owners bombarded with claims that AI and digital transformation are essential for survival. Hearing AI discussed everywhere, they return anxious and announce adoption with only the vague aim of reducing costs and improving efficiency. Without understanding why, they soon confront data preparation, process changes, changing behavior, and disappointing benefits, then retreat. Leaders must understand their company’s problems and define the scope and boundaries clearly. If they cannot explain what they need, internal teams, advisers, and startups struggle to deliver.

Reading guide

Using TNL Mediagene in October 2026, Robin questions the strategic direction of a Taiwan–Japan media group that merged, listed in the United States, and then agreed to sell its Japanese business. He compares its market capitalization, the transaction consideration, and DIGITIMES’ valuation, but the dates, exchange rates, and share-count bases have not been reconciled. The announced US$5.5 million is conditional total consideration including cash, debt treatment, and a note—not an immediate cash receipt of that amount. At announcement, the agreement still awaited closing. This article preserves the questions raised then, without drawing conclusions about valuation gaps or the future headquarters.

Reading guide

This reprinted article on forward-deployed engineering discusses how B2B startups choose reference customers. Such customers provide revenue and peer endorsement, while field experience shapes product direction. Well-funded requests that build no reusable capabilities can consume the entire team. Robin adds that knowing which projects and customers to decline matters as much as knowing what to pursue. The Harvey and Palantir examples are relayed from the article and were not individually verified in this review. Their figures should not be read as achievements independently confirmed by this site.

Reading guide

This summary of Dedao’s discussion of 精進 distinguishes hoarding information, understanding words, and actively decoding ideas. Ask what the author claims, why, where it applies, how it can be challenged, and whether it transfers elsewhere. Robin treats reading as a continuing conversation with the author, not wholesale acceptance. He also suggests assigning AI different roles to challenge, defend, and supplement an argument, helping reveal blind spots. This is a thinking exercise: it cannot guarantee that AI perspectives are independent or replace source checking and human judgment.

Reading guide

A reprinted article on AI-powered one-person companies, followed by Robin’s brief commentary; the original author is not identified. Before starting, recognize the gap between observing and doing, and the reality of not yet knowing how to run a company. Three layers matter: distinguish resources supplied by an organization from personal capabilities; build a working business system from positioning through customer relationships; and use AI as a long-term collaborator that amplifies judgment and capacity. AI helps only when you enter the market and improve your judgment through experience. Robin favors leaner teams and more careful hiring.

Reading guide

This second recommendation foreword to 未來回推法, whose author is not identified, contrasts projecting forward from today with working backward from the future. The former supports daily stability, the latter strategic breakthroughs; both are necessary. It presents future-back thinking as a trainable organizational practice: co-create a future picture across departments, break it into staged tasks, and establish regular foresight mechanisms to counter short-termism. With AI and supply chains changing, founders and corporate investors need to balance survival pressure with long-term positioning.

Reading guide

The translator’s foreword to 未來回推法 argues that optimizing the present cannot guarantee future success: past strengths can become constraints. Leaders should imagine changes in their industry and organization five to ten years ahead, then work backward to decide what to invest in, change, and stop. Strategy creates a reason to exist in the future. Facing AI-driven change, organizations need to move from past-experience-driven thinking toward future possibilities through dialogue and team co-creation. The method also applies personally: move from being pushed by the past to being drawn toward the future.

Reading guide

Investors assess a market’s ceiling, but founders should not simply chase media excitement. This column brings together industry size, niche needs, and personal passion. Begin with a problem you genuinely understand and want to work on for years, rather than choosing a venture for fundraising.

Reading guide

Reflecting on VC’s role in innovation, the author argues it must connect talent, policy, and markets as well as capital, and share entrepreneurs’ risk of failure. Taiwan’s fortieth VC anniversary becomes a call to accept responsibility for the next generation of industry.

Reading guide

This 2026 column considers positioning for Taiwan’s frontier-technology startups through AI infrastructure, semiconductor supply chains, and US–Taiwan industrial complementarity. It advocates understanding end-market giants’ needs and pursuing international opportunities through existing supply chains, hardware–software integration, and strategic cooperation rather than doing everything alone.

Reading guide

More accessible funding, abundant incubators and accelerators, founder-friendlier terms, and AI have lowered entry barriers. Success is harder: abundant capital fuels irrational, unprofitable competition; talent costs more and is harder to retain; accelerating technology makes investors hesitate. Understanding the rules and everyone’s motives still offers a chance to win.