The chaotic period of the large-scale model: the boss is keen to discuss the Tao, and the entrepreneur speeds up

Source: Titanium Media

Author: Guo Hongyan

Image source: Generated by Unbounded AI‌

It is a cognitive challenge between the radical and the conservative.

"In Silicon Valley, on the topic of large-scale models and generative AI, there is no discussion on whether to vote or not to vote. There are only discussions about which track to vote for and which project to vote for." A Silicon Valley investor told "Titanium Media Venture Capitalist".

The madness of Silicon Valley bears witness to the numbers. According to PitchBook data, since the release of GPT-3 more than two years ago, venture capital investment in AIGC has increased by more than 400%. In 2022, the investment circle will invest a total of US$1.37 billion (equivalent to approximately RMB 9.369 billion) in generative AI companies billion), almost reaching the sum of the past five years, and even reached a staggering $2.1 billion this year.

Looking back at China, the large model has been running wildly for more than 200 days, and the voice of contradictory cognitive discussions is louder than the actual progress of the implementation.

Venture investors with real money in their hands are eager to try, betting on large models and AIGC; while others say "let the bullets fly for a while."

"Titanium Media Venture Capitalists" learned that those who are eager to try are either really rich and powerful faction investment institutions with enough bullets in their hands; Come to this opportunity to "turn the tables against the wind".

"Let the bullets fly for a while, one is that there is no money, and the other is that there is no accumulation in the technical field before, and now I am still working hard to learn and research the underlying structure and trends of the large model." A fund partner said " Titanium Media Venture Capitalist said, "To put it bluntly, investment institutions are also afraid of being cut off by entrepreneurs."

**"This is a competition of cognition. What kind of cognition you have about the big model will determine what kind of project you will vote for." **In the opinion of Xu Siqing, the founding partner & CEO of Alpha Commune.

The cognitive inconsistencies of the large model do not focus solely on the individual. When investors, entrepreneurs, and demanders all have cognitive biases, "aggressive and conservative" for large models has become a common state for most players.

Fu Sheng, CEO of Cheetah Mobile, a serial entrepreneur, and Zhu Xiaohu, managing partner of Jinshajiang Venture Capital Fund, said in Moments: "Half of the start-ups in Silicon Valley started around ChatGPT, and our investors can be so ignorant and fearless."

Zhu Xiaohu, who has dominated the primary market for many years, responded, "99% of the value is created by GPT. What is the value of such a start-up company?"

At first glance, there is no right or wrong, just different positions. Entrepreneurs at the top level have stepped down one after another, rolling up their sleeves and starting to work, from Wang Huiwen’s light years away, to Wang Xiaochuan who entered the big model with 50 million US dollars of start-up funds, and then to Jia Yangqing, the master of the Ali framework, who resigned and devoted himself to AI . Entrepreneurs with dreams are as excited as they have caught the second wave of Internet entrepreneurship.

Artificial intelligence concept stocks in the secondary market are also soaring all the way. From New Year's Day to the end of June this year, the artificial intelligence index (884201) rose by nearly 70%. But Bank of America strategist Michael Hartnett has called the rise in artificial intelligence a "baby bubble" and warned it may be bursting.

There are not many people who have actually used AI products. Recently, Morgan Stanley conducted a survey of more than 2,000 people, and the result was that 80% of them had never used ChatGPT or Google’s Bard.

The multiple contradictions of radicalism, consensus, conservatism, bubbles, and strangeness are embracing this "technical revolution." Are the opportunities in AGI, the vertical model, the infrastructure layer, or the application layer?

A group of geeks did not think too much about their entrepreneurship, but directly connected to ChatGPT to start exploring commercialization.

Who is "voting with feet" AI

**Users are the best way to test the product. Whether it is valuable or not, users will choose to vote with their feet. **

Scenarios such as cross-border e-commerce, video creation, and meeting records have become the first efficiency-enhancing positions of ChatGPT.

"Titanium Media Venture Capitalist" learned that the cross-border unbounded ecological chain has been using ChatGPT to improve the efficiency of product page listing since December last year. After research, it is found that ChatGPT can realize batch production of product pages. The standard of the store.

Founder Qian Dazhu said: "It used to take a group of four or five people to complete the work in a day, now only one ChatGPT exporter can do all the work in an hour**, and the extra time can be done More things."

After GPT-4 launched the image recognition function, import a product picture through the website to let GPT-4 analyze the picture, describe and refine the selling point of the product, input prompt words, and request to imitate the expression that meets the preferences of Amazon buyers. 8 different versions are generated at a time, which means that 8 stores launch a product at the same time, and the Amazon background will not recognize the relevance. Each store can obtain the same traffic calculation, which is of great help to improve performance.

"The threshold for e-commerce is actually not high, so if you want to widen the distance with similar cross-border merchants, you must rely on running speed or new technology." Qian Dazhu said.

Guo Chenlu, the founder of ShulexVOC, whose technology has flowed out, is also a practitioner of ChatGPT. ShulexVOC is a plug-in that helps sellers quickly analyze the advantages and disadvantages of products, purchase motivation, user expectations and usage scenarios. It can be used on both Amazon and Shpfiy platforms. The underlying logic is ChatGPT's algorithm plus a self-developed small model, which processes user comments and photos through natural language, then extracts tags and turns them into text, and finally analyzes corresponding products. There are currently more than 30,000 users.

"The advantage of many cross-border merchants is their supply chain capabilities. They lack the ability to apply digital efficiency-enhancing products, and they seldom actively look for this type of products except for those provided by the platform itself, but these products can really help. Sellers can improve their ability to select products and optimize categories, thereby improving sales performance.” Amazon’s product manager said.

Not long ago, Abhay Parasnis, chief technology officer of Adobe, founded Typeface, an artificial intelligence marketing tool, to help companies generate marketing content on platforms such as blog posts, Instagram posts, LinkedIn homepages, and company official websites. Parasnis called it a "10-fold content factory." .

The underlying logic of Shenpao is the application of Stable Diffusion and OpenAI data, plus self-developed multi-modal generative process innovation marketing products. It has received 2 rounds of financing for more than a year since its establishment, with a total financing amount of 165 million US dollars. Investors include Salesforce Ventures, GV (Google Ventures), Menlo Ventures and M12 (Microsoft Venture Capital Fund), and the post-investment valuation reached 1 billion US dollars. And obtained Google, Microsoft signed commercial contracts.

“Using large models to make superficial applications or product-driven applications, there are many opportunities for both To B and To C,” Xu Siqing, founding partner and CEO of Alpha Commune, told Titanium Media Venture Capitalist.

One of the beneficiaries is also one of the beneficiaries of self-media creators. A self-media creator said in a chat with "Titanium Media Venture Capitalist" that it has always been a problem for him to convert long-form text content into short video content, until he used the ChatGPT content summary The general function, generating short video scripts in 1 minute, combined with the intelligent video generation tool of Clipping, has increased the team's video production speed by more than 5 times.

"Paste the full text in, ask ChatGPT to generate a 600-word video script copy, and you can get the result in 1 minute, and you can use it after revising the beginning. The video is also the same way. It used to take 3 days to make a video. Now At least 5-8 videos can be completed in one day." This self-media person showed off his efficiency-enhancing artifact to the "Titanium Media Venture Capitalist".

Just like Jingwei Zhang Ying’s speech in the Chaos Academy, he believes that the key to AI’s early decisive victory is to dare to seek breakthroughs from the ToC scene, because the data flywheel effect that the C-end can bring.

Scenarios are the priority and data is the king. The model itself is not what users need. What users need is products that meet customer expectations and improve commercialization benefits. Just having a large model without application scenarios is like looking for a nail with a hammer, and there is nowhere to use the tools. How to infiltrate the capabilities of the AI model into the scene is a matter worthy of deep investment, just as in the APP era, Meituan, Didi, and Ctrip help us facilitate our lives.

Between the frenetic platform-level or disruptive big opportunities, there are gradually emerging and realistic small opportunities here and there.

"Gold Swallowing Beast", no business model

The emergence of large models allows everyone to collectively enter a "navigation-free" mode of exploration.

"When OpenAI trained ChatGPT, it may not have thought that it would shock everyone, and it may not have thought clearly about how to commercialize it. It just felt that there should be newer technological explorations." Teacher Xu from Silicon Valley said at a public event.

"The big model currently has no business model. If you can afford the money and the ecological construction, you can invest in the big model." A senior investor in the industry told "Titanium Media Venture Capitalist".

How much money does the big model burn? According to The Information media reports, three people familiar with OpenAI’s financial situation revealed that due to the development of ChatGPT and the recruitment of key employees from Google last year, OpenAI’s losses roughly doubled to about 540 million US dollars.

OpenAI CEO Sam Altman has privately hinted that the company could raise as much as $100 billion over the next few years to further develop AI capabilities, the people said.

Profit is mediocre. Two of the above-mentioned OpenAI financial insiders revealed that OpenAI’s revenue in 2022 will only be 28 million US dollars, mainly from selling access to its artificial intelligence software to application developers. The level is only 0.14‰ of Microsoft's annual revenue last year.

In the case of meager profits, OpenAI launched another killer strategy in seizing the ecological niche on July 7. It officially released the GPT-4 API to be fully open for use. Now all paid API users can directly access GPT-4 in 8K context , without any waiting.

** How to commercialize the large-scale model of "swallowing gold" lies between development and reality. There is no mature reference model, and the profit model of the Chinese large-scale model is also "crossing the river by feeling the stones". **

Xu Siqing said that this time and when the previous generation of Chinese Internet companies are formed, the ecological structure will change a lot. At that time, each big company held a strategic direction and occupied a certain field, such as Baidu for search and Tencent for games. , community and instant messaging, Ali is doing e-commerce, and Byte is doing Internet new media communities and advertising, each of which is unique in its respective field.

In his opinion, the situation in this round has changed. It is unlikely that major manufacturers will directly engage in vertical fields. Basically, they will quickly occupy the fields they want to occupy based on the large language model, and establish an ecological environment to create a large The new-generation platform, because AIGC technological breakthroughs determine that many industrial structures will be restructured, so I think big manufacturers will actively build their own ecological environment, which is the focus of competition.

"** Only by seizing the ecological environment can we seize the largest audience and establish a broader border. This is an opportunity that Chinese companies have rarely faced before."** Xu Siqing said.

"At this stage of the AI wave, technology must be driven first and product definition is the most important thing. Future applications will be more closely integrated with model capabilities, so the understanding and gap between models will determine product and user experience. There are technological innovation genes and A small team with great ability must run as hard as possible.” In Zhang Ying’s opinion.

"AI infrastructure is essentially a trinity of computing power, algorithms, and data. In the end, who can integrate the three capabilities well, and the ability to provide lower costs and lower thresholds is the most important point in determining the entire competition. But it is still in the chaotic stage.” According to Shi Mao, the founding managing partner of Changlei Capital.

"In terms of direct generation of results, we think there is still a long way to go if it is to be commercialized and productized." Zhang Yitian, chief expert of the National Voice Innovation Center, said at the 2023 Huaying Capital Annual Conference.

In the long run, the industrial development of emerging technologies is essentially driven by business needs. Whether an enterprise needs a large-scale model will inevitably include multiple factors.

at last

Success is also data, failure is also data.

"Judging from the speed of Chat GPT's development, the scene will be replaced sooner or later, so the underlying logic of returning to ChatGPT is data, and no one can replace unique and deep value data." According to Shulex Guo Chenlu.

But once the data has been trained, it is not possible to "untrain" or delete or remove the data. For many companies, the moat of competition is data, and they don't want anyone to grab data for free, such as serious medical and legal industries.

There is no standard answer on how to find a balance between innovation and data security, whether to stick to small sample training and enlarge the model, or to rely on the large model.