Text: Silicon Xingwen, Author: Shao Xuhui
After the GTC, the discussion about generative AI has not subsided for a long time, and the two high-profile funding announcements in the Gen AI space just last month have also attracted a lot of attention: Augment raised $227 million at a valuation of nearly $1 billion, and Cognition raised $175 million at a valuation of $2 billion. From text chatbots to video generation and bots, the capabilities of generative AI are constantly expanding, and the application tide has begun.
As a practitioner in the field of AI and deep learning for many years, a former manager of a large company, an entrepreneur, and now an investor in the field of deep technology, I would like to share my personal judgment on the development of generative AI.
Essentially, I think generative AI's competitiveness is rooted in these three dimensions:
The first is computing power and infrastructure, which is also where many large companies have obvious advantages.
The second is the algorithm, which of course also means the talent behind the algorithm.
The third is data and application scenarios, where there are still a lot of opportunities for startups.
Let's look at the future of generative AI from these three dimensions.
Gen AI is not a computing power race, and these are the opportunities for startups
In terms of computing power, Nvidia is the leader of GPUs and the pioneer of the CUDA ecosystem, which has a huge advantage, but in fact, the future of generative AI is not absolutely determined by computing power.
Where are the opportunities for start-ups or scientific research institutions with limited computing power?
First, startups can build the infrastructure for generative AI to solve the underlying problems.
Companies in almost all sectors are considering their own competitive strategies for generative AI. In terms of data privacy/computing security,
Accuracy/reliability, business logic, and so on will have very fragmented requirements. However, the tool chain and service chain of large factories can only solve part of the needs, and the rest needs to be filled by start-ups. Lepton.AI, Corvic.AI, Fairly.AI, etc., all fall into this category. These startups have built the infrastructure in the field of machine learning, such as Corvic.AI, and they provide solutions that make it easier to turn complex data into usable enterprise-grade AI, providing predictive analytics, AI assistants, data labeling, and more.
Second, although large companies have advantages in computing power, startups can set their sights on professional fields - for these fields, the continuous accumulation of professional data will have irreplaceable value, and the accumulation of industry barriers and compliance can also form a certain moat.
More importantly, the underlying logic of many fields (such as biomedicine, network security, scientific research, and manufacturing) is not similar to the language, text and video that large models are good at, so they cannot be directly applied, and it is not easy to obtain good results by simply doing secondary development.
On the other hand, large models also open up a lot of new opportunities, and some fields that were far from commercialization in the past may suddenly have momentum to move forward, and there are often opportunities for startups.
The AI coding mentioned at the beginning of the article is a new opportunity, both of which are AI companies that have grown rapidly in recent years, and Cognition has only been established for half a year. We've also invested in Metabob, a company in the same category, to help people find bugs in their programs and fix them through AI. Such companies have found their market at the intersection of AI and software engineering.
The recent boom in robotics is also a good example.
In the past, traditional robots could only solve a single task, and the sensors and cost investment of robots were limited, which also limited its development.
With the development of LLMs, robots can learn, iterate, and complete complex tasks through reinforcement learning in the virtual space, which has brought a series of chain reactions - enterprises are willing to develop robots with higher cost of ownership and stronger capabilities, and the industry has more imagination and investment. The team involved by Professor Li Feifei of Stanford University also built a large-scale training scenario based on a physical model similar to ImageNet's training test robot in virtual space.
In March of this year, the video released by the robot company Figure in cooperation with OpenAI attracted a lot of attention: the metal-covered robot is connected to OpenAI's large language model, which can quickly understand human intentions and make corresponding actions, including tidying, accurately placing items and completing some relatively vague instructions - a man said to Figure 01 "Find me something to eat", the robot thought for a while, picked up the apple on the table, and handed it to him. It is worth noting that there are also scattered dishes and drain racks on the table, and apples are the only edible items, this simple action involves reasoning and thinking process, plus the previous sorting and storage, the robot shows the ability to complete multiple complex tasks.
In addition to Figure 01, ChatGPT-led large models are leading the way in which various LLMs are being introduced: scholars at the University of Michigan have also released home robots based on large language models that can better understand 3D environments. In the fields of industry, agriculture, and medical care, it is conceivable that new robots will bring many new changes.
New chips and algorithms could rewrite the monopoly landscape
At present, in terms of generative AI algorithms, there is a monopoly pattern of NVidia, OpenAI, and Microsoft. However, in my opinion, this situation will also change with the development of technology.
This is not my own opinion, it can be said that it is almost an industry consensus, but it is difficult to predict the exact point of time.
At this year's GTC, the original team of Transformer met for the first time, with seven of the eight authors talking to Huang on the central point: Transformer is old enough to see updated models.
Gomez, co-author and founder and CEO of Cohere, commented: "I'd like to see a model that's ten times better than Transformer to replace it...... Transformer has the possibility of optimization in terms of memory footprint and many architectural aspects, such as a very long context is expensive and cannot be scaled, its parameterization may not be necessarily that long, and we can compress it many times, bringing exponential shrinking.
From the perspective of bionics, this point of view is also very tenable - the power consumption and computing power of the current algorithm are still far from the natural world.
The human brain consumes tens of watts. The brain power consumption of insects is in the milliwatt/microwatt range, the number of neurons is very small, and it can also complete particularly complex perception-related behaviors such as stereo vision, three-dimensional control, hunting and escape, etc. In the same situation, if computers can reproduce the ability of small animals to identify, control, hunt, and survive in nature, more advanced models are needed, and more advanced models will gradually appear.
In the future, better algorithms and models are likely to mean the development of more specialized, small chips.
For example, the D-matrix we invested in earlier, which focuses on in-memory computing, is expected to be mass-produced in 2024, and there are already a large number of orders. Another analog computing chip design company, Tetramem, has also attracted attention and has published a number of Nature papers.
In addition, relatively small transformer models such as Microsoft's PHI-2 and Mistral 7B, open source, will also see rapid development. In my opinion, these small open-source models are necessary for the sustainable and healthy development of the AI industry. This AI boom is different from previous tech booms, and large companies do not occupy all the important research directions - when Google went public, many universities stopped working on search algorithms because Google already had good teams and resources to advance search algorithms. Generative AI is not, and there is still a lot to explore in these small open-source models for research institutions, and the computer science departments of all prestigious universities are also actively researching open-source models.
It could even be said that the new algorithms I mentioned at the beginning of this article and that we are looking forward to in the future are likely to emerge from the work of these researchers.
Behind the development of GenAI is the battle for talent
This expectation of newer, better generative AI algorithms also means something else: the race for generative AI also means a war for talent.
At present, China and the United States are the two countries that occupy the first and second positions in the world in terms of AI talent.
In absolute numbers, China occupies the first place — in March, the New York Times reported a report tracking AI-related talent, which has the largest number of AI undergraduates in the world, and now exceeds the proportion of Chinese among the top AI talent in the United States compared to three years ago. Compared with the past, these Chinese have only "returned" to China after studying for degrees in the United States.
Of course, the United States has the world's largest number of first-class universities and technology companies with a strong atmosphere of innovation, and has also given birth to Open AI, which has brought LLMs into the public eye, and still has a strong attraction to AI talents from all over the world.
It is undeniable that based on the consideration of time, place, and people, China and the United States firmly occupy the top two positions in the comprehensive and optimal future AI development.
Against this backdrop, there is another noteworthy trend: the polarization of talent.
As AI+ becomes more leveraged, the number of technical talents needed to build a good AI company has decreased significantly compared to a few years ago. The next competition for talents will be more and more concentrated in the competition for a few top talents.
Whether in Silicon Valley or in Beijing, Shanghai and Hangzhou, we have seen the polarization of top AI talent being scrambled for high prices and ordinary college students unable to find jobs, and this trend will become more obvious in the future.
Gen AI will become a trillion-dollar market, how can investors enter the game?
From my perspective and that of venture capital practitioners around me, the consensus is that the future of Gen AI is bright – the Bloomberg Intelligence report predicts that GenAI will be one of the fastest-growing markets in the next decade, with a total market volume of more than $1 trillion and a compound annual growth rate of more than 40%. Its transformative capabilities impact all walks of life.
For investors, if you want to invest in this market, I have the following suggestions:
First of all, investing in generative AI projects is still essentially looking for projects with commercialization potential in deep tech technologies. Investors should have an AI-related technical background, and maintain continuous learning and market acumen. At present, the technological changes, breakthroughs, and markets of generative AI are developing rapidly. For example, some projects with the name of large models may completely lose the market after a certain update of ChatGPT, and being able to identify these projects can help avoid many risks.
Secondly, as analyzed at the beginning of this article, the current pillars of generative AI are divided into three directions: talent/algorithm, computing power, and application/data. As an investor, the first two are more difficult to invest, but there are a lot of future opportunities in the application/data, and you can consider investing in some related resources.
For example, application scenarios and data related to application scenarios - for Chinese investors, semiconductors, new energy, and advanced manufacturing are all good directions. Taking the manufacturing industry as an example, only with a large amount of data in vertical industries can we make good AI and guide the future of advanced manufacturing.
From the perspective of time, if the technology is not particularly certain, you can also wait for the information of commercial landing and pay some value-added costs in exchange for a more stable signal.
When it comes to AI, I often hear this question: Can AI eventually replace humans?
I think this has to be understood on a different scale – what exactly are we talking about is the extent to which AI is challenging human evolution? If it is at the organism level, which has accumulated over millions of years, it is the most difficult; From the level of human cognition, there are at least 100,000 years of evolution in front of AI; And human language that has been around for thousands of years is relatively easy; Computer languages, which are only 100 years old, are the simplest.
The answer to this question is, of course, inconclusive. But in a market that will eventually reach a trillion dollars and change people's lives, I'm looking forward to seeing more Chinese participants, whether it's starting a business, investing, or actively embracing generative AI to make life and work more efficient, which will change the way we interact with the world like never before.