ChatGPT已经挑战产业对于AI的认知。GPT4.0尚未发布，未来可能将会是生成式AI时代的”报晓鸟”。在超级AI尚未到来，生成式AI快速演进的时代，我们的十大猜想是：

ChatGPT has challenged industry perceptions of AI. GPT 4.0 has yet to be released, but we think it could be the harbinger of the generative AI era. While generative AI is evolving rapidly and before super AI arrives, we envisage 10 possible outcomes:

1）GPT4.0未来可能会是多模态的具有思维链推理能力的大模型，生成式AI可能成为人类思维的“大副”，标志着AI的智能涌现速度超越Scaling Law的提升速度，PC是人类思维的“自行车”（Steve Jobs），而生成式AI可能会是人类思维的“Copilot/大副”。全要素生产率在PC时代飞速增长，在互联网普及后（2005-2022）增长缓慢，生成式AI时代全要素生产率可能重新加速上升。

1. A co-pilot to the human mind

GPT4.0 could be a large-scale model with a multimodal chain-of-thought reasoning ability. This would mark a faster artificial intelligence (AI) emergence than under scaling laws. With that, generative AI could take the role of first co-pilot to the human mind, much beyond the function of personal computers as “bicycles for the mind”, as Steve Jobs called them. Total factor productivity (TFP) soared in the PC era but the pace of growth slowed when internet usage became widespread (2005-22). We believe TFP could reaccelerate in the generative AI era.

2）生成式AI时代的技术基础可能不仅是深度学习的基础模型/大模型（Foundation Model），也或许会包括神经计算科学与符号推理的更多贡献。

2. Add neurocomputational science and symbolic reasoning

The technological basis of the generative AI era may not be limited to the foundation model or large language model in deep machine learning, but also include contributions from neurocomputational science and symbolic reasoning.  

3）大模型/并行计算最重要的应用不仅包括AGI通用人工智能（OpenAI） ，也或许会包括合成生物学（AlphaFold） 、可控核聚变（DeepMind） 、科学计算与模拟仿真（NVidia，需要并行计算，但也将大模型加速）。意味着其可能改变宏观经济学的一些假设。经典的经济增长模型（索洛模型）尽管预设了劳动力和资本在经济增长中可以互相取代，但是也假设了劳动力按照实物生产力付酬与劳动力按照人口增长模型限制。如果AGI的劳动力出现，经济增长的假设可能需要重新考虑。

3. Potential shift in economic growth assumptions

Key applications of large models for parallel computing could comprise not just artificial general intelligence (AGI: OpenAI) but also synthetic biology (AlphaFold), controlled nuclear fusion (DeepMind), scientific computing and simulation (Nvidia: not just parallel computing but also acceleration of the large model). These could change certain macroeconomic assumptions. For instance, the classical model of economic growth (Solow), which presupposes that labor and capital can replace each other to drive economic growth, and that labor is paid according to physical productivity, runs contrary to the population growth model, which believes that labor is limited. If an AGI workforce emerges, we would need to reconsider our economic growth assumptions.

4）我们认为算力重要的将是软件栈上的硬件优化，模型重要的将是数据集上的算法工程优化，应用重要的是数据采集应用开发。我们认为算力公司的新机会或许在于怎么在Z级别的算力上实现规模拓展（Scaling Out）和摒弃传统核内存共享，以及边缘计算；模型公司壁垒或许在于Domain Specific的数据；应用壁垒或许在于传感器（软件传感器的App、场景或者硬件传感器）。更为重要的依然是算力网络-基础模型-应用-数据的“飞轮”。但我们认为与云计算时代的解耦相反，生成式AI时代的经济效应的“飞轮”可能是高度耦合的，因为云计算关键在于弹性和可伸缩可拓展性带来的成本优化，而AI关键在于效能提高到某个阈值之后的应用价值急速上升。而数据栈可能从data source和activation tool向数仓逐渐创新。

4. High coupling of economic factors in generative AI 

We believe what will matter in power computing is hardware optimization in software stacks; what will matter for models is algorithm engineering optimization in datasets; and what will matter for applications is data collection application development. For companies in the power computing industry, the new opportunities would have to do with how to achieve scale expansion (scaling out) and abandon traditional core memory sharing in Z-level computing, as well as in edge computing. The barriers for model companies could lie with domain-specific data while the barriers for application would relate to sensors (software sensors for apps, scenarios or hardware sensors). And the flywheel comprising the computing power network, foundation model and application data would become even more crucial. In contrast to cloud decoupling, we believe the flywheel of economic factors in the generative AI era will be highly coupled. This is because the key to cloud computing is cost optimization on elasticity and scalability, while the key to AI is to rapidly raise the value of applications after their performance reaches a certain threshold. And the data stack could gradually innovate from data source and activation tool to digital warehouse.

5）生成式AI应用层公司的商业模式将可能改变微观经济学假设：我们认为软件行业、互联网行业、开源、生成式AI均为对传统微观经济学的挑战。软件意味着非个性化产品的边际生产成本接近于0（但产品定制化和营销成本高），互联网意味着产品的边际分发与营销费用接近于0（直到流量红利结束和获客成本上升），开源意味着产品本身的价格为0（但部署和开发有成本），生成式AI意味着产品的个性/定制化生产边际成本接近于0。波特三战略包括成本领先、差异化、聚焦。差异化产品在生成式AI时代将可能不再具有竞争壁垒，但差异化数据将可能具有较高价值。

5. Differentiated data will likely be more highly valued 

In the generative AI era, the business model of application layer computing companies will likely change microeconomic assumptions. We believe software and internet industries, open source and generative AI markets will challenge traditional microeconomics. In software, the implied marginal production cost of non-customized products is close to zero (unlike customized products with high marketing cost) and in the internet business, implied marginal distribution and marketing costs are close to zero (up to a point when the benefits of user traffic level off and customer acquisition cost begins to rise). Open source implies that the product price is zero (even if deploying and developing the product incurs costs), while generative AI implies marginal production costs of personalized and customized products are close to zero. Applying Porter’s three strategic options (cost leadership, differentiation and focus) to generative AI, it is likely that differentiated products will no longer have competitive barriers, while the market will assign higher value to differentiated data.

6）我们认为AI将有可能改变软件与互联网的结构，Marc Andreessen著名的论断为软件吞食世界。我们认为互联网/开源已改变软件（SaaS），SaaS的本质是抽象出的最佳实践与最解耦拓展的结合，而AI可能会以1、改变最佳实践 2、改变定制化开发 3、改变工作流 4、改变开源测试 重构SaaS。

我们认为AI改变互联网则在于全新的交互与更强的粘性，上一代推荐/搜索算法主要承担大规模的结构化embedding，把人归结为特征统一，而相对较难生成对个体的深层次理解。因此新的更深的大模型下，不再是统一的平台供给现存的个性化内容，而是个性化的交互产生个性化的全新产品。

6. Brand-new products generated from deeply personalized interactions

Our sense is that AI will likely change the structure of the software and internet markets. Software engineer and entrepreneur Marc Andreessen has famously asserted that software is eating the world. We believe that open source internet has indeed changed the software market (SaaS): the crux of SaaS is the combination of abstracted best practices and the most decoupled extensions. In this light, AI could bring changes to the SaaS industry in terms of its best practices, customized developments, workflow and open source testing. 

As for the impact on the internet business, we believe AI could bring brand-new interactions and strengthen customer stickiness. Previous generations of recommendation and search algorithms mainly undertook large-scale structured embedding and attributed unified characteristics for people, making it difficult to generate deep understanding of individuals. The new large model and its deeper nuances will no longer use a unified platform to create personalized content, but rather, personalized interactions will generate personalized brand-new products.

7）AI生成产品。我们认为多模态大模型将可能让深度学习对个体的多层次理解成为可能，让“AI生成设计”到“AI生成产品”成为可能，这个市场可能会比 “AI生成内容”大很多倍，我们认为真正的个性化需求的商品将对仿真，柔性生产和柔性供应链提出全新的要求。 

7. AI-produced goods could raise simulation and flexible production levels

We posit that the multimodal large model will make it possible for deep learning to understand individuals at multiple levels, making it possible for AI-generated design to make AI-generated products. This market could be many times larger than that of AI-generated content. We believe goods that satisfy truly personalized needs will lead to new market requirements for simulation, flexible production and flexible supply chains.

8）我们认为创作者经济有可能快速成长，克雷创造了超算，Linus创造了Linux，亨特创造了Kenshi。在生成式AI的辅助下，个人创造产品的能力可能大幅提升。

8. Enabling creators 

We believe the creator economy will likely grow rapidly. Cray created supercomputing, Linus created Linux and Hunt created Kenshi. Generative AI would greatly amp up the ability of individuals to create products.

9）基础模型的人类对齐（Human Alignment）将可能成为最重要的工程之一。

9. Applying human alignment 

We believe that human alignment of the foundation model will probably become one of the most important projects.

10）用数学语言解释大模型的“涌现”的能力将会更加重要。无论涌现的能力来自流形上的概率分布，还是范畴论中把梯度递降和自动微分实现为函子。对于基础模型的深刻理解与控制的要求我们必须能够在超级AI之前的生成式AI的黑箱之内用更先进的数学语言去推导与描述，如果蒸汽机没有牛顿力学热力学与微积分，如果曼哈顿工程没有狭义相对论，都只会是经验的不可控工程。大卫希尔伯特曾说“我们必须知道，必将知道”，我们认为用在深度学习的大模型上有一定的恰当之处。即我们必须深刻理解大模型的原理而非仅仅当作黑箱应用。

10. Advanced mathematical language will be essential

The ability to explain the emergence of the large model in mathematical terms would become increasingly essential. The ability to emerge could relate to probability distributions on manifolds or to the realization of gradient decrement and automatic differentiation as functions in category theory. Deep understanding and control of the foundation model means that we must use advanced mathematical terms to derive and describe it within the black box of generative AI that precedes super AI. The steam engine without Newtonian mechanics, thermodynamics and calculus, or the Manhattan Project without the special theory of relativity, would only have been empirical and uncontrollable projects. It is appropriate to apply David Hilbert’s conviction that if “we must know, we will know” to the large model in deep learning. That is, we must have a deep understanding of the principles of the large model, not just treat it as a black-box application.