AI is all the “rage” lately and the noise ratio is extremely high out there. A cool head is needed. Here I would like to ZOOM OUT and lay out some high-level thoughts/questions, hoping they may serve as “axioms” to further think through the practical actions and directions for start-ups and investing.

• AI hype is over-rated in the near term and under-rated in the long term. At least we can all agree that this time is more tangible and value-adding than the hype around Crypto, Web3, and Clubhouse. Get-shit-done-ethos and continuous creative tinkering will see the AI experience improve from good enough/a toy to irreplaceable workflow tools for knowledge workers.

• However, Web3 (in a narrow sense an immutable identity infrastructure) cannot be written down yet as we may increasingly need a neutral infrastructure that distinguishes human vs. machine-generated content. Though this could be a long shot and irrelevant as convenience always trumps privacy for the majority of people. As machine may soon run out of human knowledge scraped from the web and begins using second derivatives of existing content to self-evolve. A side effect could be a proliferation of “fake” & “junk” content. Some humans will play a part in this because there is a profit to be made in harvesting attention.

• If AI is like a virus, to get max diffusion, it needs to become various form factors that are great at aiding the average human at their existing jobs to avoid social push-back and reach a critical mass of infection/adoption. Once enough population is infected, it may get a life/inertia of its own just like how wheat became the most successful grass on the planet because it is useful to humans, so humans planted it in mass.

• Schumpeter’s creative destruction is a process but there is a spark that ignites the process, ChatGPT is that. True technological innovation is deflationary as productivity per knowledge worker will be enhanced by AI-enabled tools. The search experience will be enhanced, but the search business model will get worse before it gets better until costs to serve per LLM-based query come down and new search modalities get created and adopted, along with new ways to monetize. The pie for old biz models will get smaller with some product categories seeing a shrinking revenue/profit pool while others get created and become larger.

• Sometimes it is not the best technology product that wins, but the one good enough with the largest distribution and reaches a critical mass of adoption. Yann Lecun describes “LLM is an off-ramp on the high-way to Human Level AI”, but AI does not need human-level intelligence to be dangerous. Here lies the difference between an entrepreneur and a well-regarded scientist. The next-gen product (LLM-based or alternative architecture) will need to be significantly better before investments/ecosystems coalesce around the current best solution, namely LLMs. Do you go all-in on LLMs or do you hedge against another paradigm shift?

• Google is facing the classical innovator’s dilemma in the AI search war. Microsoft used minimal capital and risk (with an arms-length entity Open-AI) in gaining PR momentum and causing maximum threat (perceived or real) to Google’s core search business. Even taking away a 5% search market share is enough to cause pain to Google as the moat around search has not seen any real challenges in the past 20 years. On paper, Google should cannibalize its own business before someone else cannibalizes it for them. That could take two paths 1) pay up for TAC (Traffic Acquisition Cost) and compress Gross Margin to gain proprietary access to certain data and hold on to publishers 2) Open-source its models (vs. OpenAI currently chose to closed-source) like how it did with Android. However, history rhymes but does not repeat in the exact same way. What will Google do this time? Google open-sourced Android because when it saw the 1st gen iPhone (Google actually was building a phone before iPhone came out), the iPhone was so good that Google realized it could not compete with Apple in designing smartphones, so it scorched the earth by open-sourcing its Android system to secure distribution for its search product. Smart move! The circumstance is different this time. On the plus side, Google

• In the application layer, U.S. and China will again diverge as parallel universes due to cultural and institutional differences, technology diffusion and application are not only influenced by basic scientific advances but also cultural nuances, if not much more. China lags behind the U.S. in LLMs development (by two years?) and now has restricted access to the most advanced AI chips, but it is less constrained by conservative vs. liberal debates (for good or bad) and more driven by government directives and pure animal spirit.

The trend is here but there are many moving forces at play. Next, I would like to ZOOM IN to think through practical problems/actions from the ground up.

Note：The following is translated from English to Chinese using ChatGPT, with minor edits (less than 5%):

AI最近备受热议，噪音比例极高。在此，提出一些高层次的思考和猜想，希望它们能作为“公理”，以进一步思考初创公司和投资的实际行动和方向。

• AI炒作在短期内被高估，在长期内被低估。至少我们可以认为，这一次比加密货币、Web3和Clubhouse更具体，更具价值。不断地创造和完善会使AI的体验从足够好/玩具变成知识工作者不可替代的工作流程工具。

• 然而，Web3的价值（在狭义上是不可变的身份基础设施）还无法被忽略，因为我们可能越来越需要一个中立的基础设施来区分人类和机器生成的内容。尽管这可能是一个遥远的目标，对大多数人来说，方便总是胜过隐私，因此这个点可能很鸡肋。由于机器可能很快用尽从Web上收集的人类知识，开始使用现有内容的二阶导数来自我演化。副作用可能是“假”的和“垃圾”的内容的激增。一些人类会参与其中，因为在吸引注意力方面有利可图。

• 如果AI就像一种病毒，为了达到最大扩散，它需要成为各种形态，能够帮助普通人完成现有的工作，避免社会的反弹，并达到感染/采用的关键质量。一旦足够多的人被感染，它可能会像小麦一样获得自己的生命/惯性，因为它对人类有用，所以人类会大规模种植它。

• Schumpeter的创造性破坏是一个过程，但有一个点燃这个过程的火花，ChatGPT就是那个点。真正的技术创新具有通货紧缩性，因为AI-enabled工具将增强每个知识工作者的生产力。搜索体验将得到提高，但在成本下降以及新的搜索方式和新的货币化方式被创造和采用之前，搜索商业模型会变得更糟糕。旧商业模式的蛋糕将变小，某些产品类别的收入/利润池将缩小，而其他产品类别将被创造并变得更大。

• 有时，获胜的不是最好的技术产品，而是分发最广、达到关键采用量的产品。Yann Lecun认为，“LLM是通往人类级别AI的高速公路的一个出口”，但AI并不需要人类级别的智能就可能具有危险性。这就是企业家和备受尊敬的科学家之间的区别。下一代产品（基于LLM或替代架构）需要在投资/生态系统围绕当前最佳解决方案（即LLM）形成共识之前显着提高性能。你是要全力支持LLM，还是对抗另一种范式转变？

• 谷歌正在面临AI搜索战中的经典创新者困境。微软通过使用OpenAI这样一个独立实体来获得公关动力并对谷歌的核心搜索业务构成最大威胁（无论是实际上还是虚构的）。即使夺走5%的搜索市场份额也足以给谷歌带来痛苦，因为在过去20年中，搜索的护城河没有受到任何真正的挑战。纸面上幼稚的看，谷歌应该在其他人将其替代之前蚕食自己的业务。这可以有两种路径：1）支付更高的TAC（流量获取费用）并压缩毛利率以获得对某些数据的专有访问并保持publisher的粘性；2）像Android一样开源其模型（与OpenAI当前选择的闭源不同）。然而，历史会重演但不会以完全相同的方式重复。这一次谷歌会怎么做？不过，微软和ChatGPT也许给了谷歌借口和美国Federal Trade Commission辩论，你看我不是垄断。
  • 当年谷歌开源了Android，因为当它看到第一代iPhone时（实际上，iPhone推出之前谷歌正在开发一款手机），谷歌意识到它无法在智能手机设计方面与苹果竞争，因此它通过开源其Android系统来确保其搜索产品的分发。聪明的举动！但这一次情况不同。

• 在应用层面上，由于文化和制度差异，美国和中国将再次分化为平行宇宙，技术的扩散和应用不仅受到基础科学进步的影响，还受到文化细节的影响，如果不是更多。中国在LLMs的发展方面落后于美国（大约落后两年？），且受到最先进的AI芯片的限制，但它不受平权主义和唤醒主义的限制。

趋势已经到来，但是有很多不同的力量在发挥作用。接下来，需要从底层深入研判实际问题和行动。