Since the open access of ChatGPT, the development of the AI field has been rapidly evolving, and the release of GPT-4 has further demonstrated the power of AI. Here, we have compiled some recent reflections on investments in the AI field. If there are any errors or omissions, please feel free to point them out for discussion.
The future pattern of large-scale models? Will OpenAI dominate alone, form a duopoly, or be gradually caught up by other players?
Throughout the history of computer development, many fields have been monopolized by one or two companies, such as ASML being the only company capable of producing EUV lithography machines, TSMC and Samsung being the only advanced process wafer foundries, Intel and AMD being the only CPU manufacturers, NVDA and AMD being the only GPU manufacturers, Windows and Mac being the only PC operating systems, and Android and Apple being the only smartphone operating systems. Moreover, the market share of these duopolies is often 80/20 or even 90/10. So, what will be the pattern in the AI large-scale model sector in the future? We will analyze it from the perspectives of both the 2B and 2C markets.
First, in the 2B market, we mainly see two types of applications. One is to open the large-scale model API or plug-in to other application-type companies, and to connect to other companies' local databases for query analysis. The other is to open the basic model to other companies for fine-tuning with exclusive data. The first type does not involve model training, while the second type involves model fine-tuning. Currently, it seems that OpenAI has already done both of these. The APIs and plug-ins are well-known, and the cases of fine-tuning include Microsoft Office's Copilot, Morgan Stanley's use of GPT-4 to assist financial advisors, and Stripe's use of GPT for fraud prevention, etc. (based on official information speculation).
OpenAI's competitive advantages in the 2B market may include its know-how, high R&D investment resulting in continuous technological progress, and ecosystem network, among others. From the perspective of know-how, pre-training and fine-tuning of large-scale models are inherently full of technical details (similar to the step-by-step process details of TSMC's wafer manufacturing). For example, the pre-training of GPT-4 was completed in August last year, and fine-tuning took another six months. Moreover, the technical report of GPT-4 mentions over hundreds of authors, making it difficult to imitate both in terms of talent density and engineering collaboration difficulty. In addition, OpenAI has stopped disclosing technical details since GPT-4, making it increasingly difficult for competitors to catch up with OpenAI. From the perspective of technological progress, if a company's technology can continuously iterate and improve on the previous generation's basis, the advantage of the leader is difficult to surpass, as is typical of TSMC's advanced process. Although there is also the possibility of other technological directions once again disrupting large-scale models, the probability is still very low at present, and the potential of large-scale models has yet to reach its limit. Finally, from the perspective of the ecosystem network, OpenAI outputs its capabilities to various software through plug-ins. If developers can use OpenAI to call the functions of other software in the future and form a rich ecosystem network, OpenAI will actually become the operating system and app store of the AI era. The vitality of this network effect is very long-lasting (as seen with Windows). This may also be the reason why OpenAI has quickly established cooperation with a large number of companies, aiming to seize market opportunities. Among these three factors, the order of importance is ecosystem network > technological progress > know-how. Without continuous technological progress, the barrier of know-how will eventually be broken by other players. Without establishing an ecosystem network, even if OpenAI's technology far surpasses other companies, specific companies facing the choice of using OpenAI large-scale models, using the industry's second or open-source large-scale models, or developing their own large-scale models, will often consider data protection, segmented domain advantages, profit distribution, and whether the usage scenarios truly require ultra-high intelligence, and may not choose OpenAI. In this case, of course, OpenAI can still thrive in high-value areas (like Apple today), but it will lose a broader market space.
In the 2C market, OpenAI's current product is only ChatGPT. Generally speaking, the competitive advantages of 2C products mainly include user habits, switching costs (including learning costs, data accumulation, trial-and-error costs, etc.), and scale and network effects. Currently, ChatGPT only has some user habits, data accumulation, and scale effects, but there is still a lot that OpenAI can do in the future, such as accumulating historical Q&A data and integrating it with other applications. This will make ChatGPT more like a personal assistant, becoming more relied upon the longer it is used, while also becoming a traffic entrance, transferring the traffic value of other applications to OpenAI's hands. However, this type of product still cannot guarantee OpenAI's monopoly status unless it can further establish network effects between users, such as by taking over the social communication functions.
Although we have speculated on OpenAI's strategic choices, the specific decisions will still depend on OpenAI's values. OpenAI's goal has always been to build a general artificial intelligence (AGI), but this will be a huge and costly dream that cannot be achieved solely through financing. Therefore, Samuel Altman's decision is to make money while supporting this dream, which is the reason why OpenAI has transformed from a non-profit organization to a for-profit organization with a profit ceiling. From this perspective, OpenAI should still actively compete in the coming years to ensure its profitability.
In summary, the future pattern of large-scale models is most likely to be dominated by OpenAI, with other companies and open-source large-scale models mainly applied in segmented markets and special fields. Whether OpenAI can establish a mature and extensive ecosystem network is the key to whether it can become an Apple or a Microsoft.
What will be the pattern of the future AI industry? Where are the greater investment opportunities in the upstream and downstream sectors?
Analogous to the current semiconductor industry, the upstream of the future AI industry will be large-scale models (corresponding to semiconductor foundries), and the downstream will be various AI applications (corresponding to chip design companies), but the majority of the value will be concentrated in the downstream. From OpenAI's current pricing mechanism, it charges based on the number of tokens used to call the API, which is a cost-based pricing model. Generating the same number of words, the value generated from chatting and generating legal contracts differs by a factor of hundreds or even thousands, and OpenAI does not capture this additional value, which is earned by downstream application companies. From this perspective, companies that directly face users can leverage AI capabilities to help customers create greater value, and their growth potential is considerable.
However, there are also some traditional software companies, such as Adobe, that still have strong advantages, but if the future trend is towards AIGC, Adobe's value may be greatly reduced or even completely replaced. Currently, AIGC can only generate initial concept diagrams, and further modifications still need to be done using Adobe, but as technology advances, AI may also be able to complete the modification function.
In the consumer market, it can be foreseen that in the future, everyone will have an AI assistant that can operate all other software, acting as a super traffic portal (similar to a more powerful WeChat mini-program). As a result, the value of platform companies that facilitate transactions between users and merchants, such as Taobao and Booking, will be reduced by half. However, if some companies have rich data accumulation and can leverage historical data and AI capabilities to provide greater value, they can benefit from this trend.
The value splitting situation judged by Shixiang:
Some Speculation：
Speculation 1：It is speculated that openAI's market share will be between 30% and 90% (outside of China), depending on whether other companies and open-source models can keep up with openAI's pace of progress, whether openAI can create network effects (such as through chatGPT being able to call other applications), and whether high-end AI demand is more or less prevalent. Preliminary speculation suggests that openAI has a 31% chance of occupying 90% of the market share (similar to Windows), a 42% chance of occupying 60% of the market share (similar to TSM), and a 27% chance of occupying 30% of the market share (similar to Apple).
Market conditions can also affect the speed at which other companies catch up: if openAI creates network effects or if the demand for low-end AI is very low, even if there is a limit to the technology of large models, other companies may not survive to catch up with openAI. Our current guess is that there is a 29% chance that other companies can catch up with openAI, an 8% chance that the gap is between gods and insects (a technological singularity), and a 63% chance that it is a gap between college and high school students (1-2 years behind).
Speculation 2: openAI only focuses on 2B and leaves the downstream to various application companies. openAI adopts a cost-based pricing model, just like the AI era's Azure and other cloud platforms (although openAI's competitive landscape is much better than Azure). The downstream application scope and market size will far exceed the upstream (even if openAI no longer uses usage-based pricing and adopts a sales-based commission like Arm licensing, downstream volume is still much larger). (Elephant says that 80% of the value goes to the basic model, but it is still doubtful that the value of downstream applications such as personal assistants is lower than that of basic models.)
Speculation 3: The logic of traditional search optimization is facing reconstruction due to the more direct and human-like answers provided by LLM. This is not only likely to affect search engines, but also a large number of websites and apps that distribute information, which will face significant business model restructuring. It can be imagined that the value of information distribution institutions will be seriously challenged, while high-quality service providers may receive better matching.
Speculation 4: In essence, the long-term industry space is still significantly underestimated. The industry scale based on language logic is huge, as almost all social relationships and collaborations in human society are built on language and information. Therefore, LLM can help and disrupt almost infinite places of the original production relationships.
Hypothesis 5: In our investment framework (standard value investment), the sources of competitive advantage such as policy, IP, scarce resources (data), complicated with other cares (such as trust costs), trial and error costs, high capex, sales network, service network, users & users network effects, and users, providers, servicers network effects are not likely to be disrupted by AI for the time being. However, factors such as learning costs (directly telling AI what to do), user habits (new forms of interaction), high R&D (to improve efficiency and use AI to gain a competitive edge), and bilateral network effects (where users are attracted to new super apps) may be disrupted by AI. Among these, the most dangerous are learning costs and habits. In other words, business models that rely on building user learning costs and habits (which are limited by the computational power of the human brain, but not for AI) as a barrier to entry may face challenges in the future.
Of course, there are still many, many speculations...
In fact, the impact of LLM or more general AI on the world has just begun, and everything is still uncertain. We only have some vague ideas, which may be completely wrong. But we believe that this is a revolutionary and far-reaching technological change.
Comments