Demystifying the Investment Fog of Large Models: In - depth Minutes of a Closed - door Meeting between Silicon Rabbit and Four Core Experts from Silicon Valley's AI Giants
How to understand the investment logic of large language models (LLMs)? Silicon Rabbit organized a closed - door meeting with AI giants in Silicon Valley. This article reveals first - hand insights into multimodal AI applications, commercialization paths, and technological differences between China and the United States.
As Silicon Rabbit, which has long observed and documented the pulse of innovation in Silicon Valley, we are well aware that in the global wave triggered by generative AI, there is a huge gap between public information and the truth of the industry. New models are released every day, and new financing news emerges every week. However, the real basis for decision - making does not solely come from press releases or analysis reports.
In order to penetrate the surface and touch the real pulse of the industry, Silicon Rabbit recently organized an in - depth closed - door exchange in the heart of Silicon Valley, the United States, for a top domestic securities research institute and a leading public fund team. We were honored to invite four AI experts at the center of the global technological storm - they are from Google (deeply involved in the R & D of multimodal models such as Gemini), Meta (leading the implementation of GenAI in the advertising recommendation system), Apple (leading the training of the underlying architecture of large models), and an international e - commerce giant (responsible for driving billions of dollars in revenue with AI).
This is not just a general trend sharing, but a deep collision on technology implementation, business challenges, and investment logic. This article will comprehensively review the core highlights of this meeting in light of the latest industry developments.
Sharing of Four Core Highlights
Highlight 1: Multimodal AI (Multimodal LLM) - A Definitive Revolution from Technological Frontier to Industrial Application
While the market is still digesting the capabilities of pure - text large models, a technical leader from Google, who is deeply involved in the R & D of its flagship multimodal model, clearly pointed out that multimodality is the inevitable evolution direction of AI, and its commercial value will far exceed that of pure - text models.
What is true multimodality? It is not just about enabling AI to "understand" pictures or "hear" sounds. Instead, it integrates multiple information modalities such as text, images, audio, and video into the same model for cross - modal understanding, reasoning, and generation. This view coincides with the trend that Silicon Rabbit has long emphasized in its articles. Through the expert network of Beta Consulting, we were able to deeply verify the certainty of this direction in industrial implementation from front - line practitioners.
The key application areas are taking shape:
Next - generation semantic search: Users will no longer be limited to text - based searches. You can take a photo of a building with your phone and ask, "Tell me the history of this architectural style and recommend nearby cafes of a similar style." This requires AI to simultaneously understand the image (architectural style), geographical location, and text instructions.
Immersive education and training: Imagine an AI tutor that can observe a student's actions during a physics experiment through a camera (video), listen to his oral presentation (audio), and read his experiment notes (text) to provide real - time and personalized guidance. This is moving from concept to reality.
Hyper - personalized e - commerce and content generation: In future e - commerce platforms, AI will generate a complete "digital image" and virtual exhibition hall that suits your aesthetics based on the clothing pictures you have saved, the videos you have browsed, and the music you have liked. This is based on a deep understanding of users' multi - dimensional data.
Investment inspiration: When evaluating a multimodal AI project, one should not only look at its technical indicators but also examine its data fusion ability and the depth of implementation in specific scenarios.
Highlight 2: The "Efficiency Frontier" of Commercialization: The Trillion - dollar Challenge of Model Compression and Productization
"A powerful model with hundreds of billions of parameters is like an F1 racing car - it has extreme performance but cannot be driven on city streets." A technical supervisor from Meta, who leads the implementation of GenAI in the advertising recommendation system, used a vivid analogy to point out the core challenge of commercialization.
The "iceberg" of costs: Training costs (ranging from millions to hundreds of millions of dollars) are just the tip of the iceberg. Inference costs are the huge long - term operating expenses. In scenarios like advertising systems that need to process a massive number of requests per second, without an efficient inference solution, costs will increase exponentially.
Key technologies to bridge the gap: How can a large - scale model be deployed on cloud servers with limited computing power or even mobile devices without significantly sacrificing performance? The industry is focusing on the following technologies, which are also the key aspects that Beta Consulting examines when evaluating the technological "moat" of AI companies for its clients:
Quantization: Approximating high - precision 32 - bit floating - point numbers in the model with low - precision 16 - bit or even 8 - bit integers to significantly reduce the model size and computational complexity.
Pruning: Identifying and "cutting off" the connections with less contribution in the neural network to achieve model "weight loss".
Knowledge Distillation: Using a large and accurate "teacher model" to teach a lightweight "student model" so that the latter can have capabilities close to the former while maintaining a small size.
Investment inspiration: When investing in AI projects, in addition to focusing on the model performance (benchmark), one should also deeply examine the unit inference cost, the maturity of the model compression technology stack, and the performance under real - world business loads. This directly determines whether the business model can be established.
Highlight 3: The Structural Change in AI Investment Logic: From "Model is King" to "Infrastructure + Vertical Applications"
A technical leader from Apple, who is responsible for the training of the underlying architecture of large models, shared a key observation: The opportunity window for simply replicating a basic large model similar to GPT is closing. This observation also confirms the view put forward by Silicon Rabbit in multiple analysis articles: AI investment is shifting from chasing the models themselves to more practical infrastructure and vertical application tracks.
The golden age of the "water sellers": The logic of AI infrastructure: Basic large models are gradually becoming "commoditized", and the "utilities" that provide training, deployment, and data management for these models have become new value high - points. Investment hotspots include AI chips, vector databases, MLOps, and development toolchains.
The value of the "alchemists": The logic of vertical industry AI: Combining the capabilities of general large models with in - depth industry knowledge and proprietary data in specific industries to create irreplaceable value. Successful examples include Harvey AI in the legal industry and GitHub Copilot in software development.
Investment inspiration: Investment portfolios should shift from single - bet model companies to diversified configurations of "infrastructure + multiple vertical applications". When evaluating vertical AI projects, the industry data barrier and the depth of workflow transformation are the core considerations.
Highlight 4: The Real Picture of the Sino - US AI Race: The Strategic Divide between Underlying Innovation and Market Application
A scientist from an international e - commerce giant with over a decade of AI implementation experience painted a more detailed picture of the Sino - US AI competition. He believes that simply asking "who is leading" will miss the key strategic differences.
The United States: Strong in "from 0 to 1" underlying innovation. Its advantage lies in defining the next - generation model architecture and leading technological paradigm shifts.
China: Strong in "from 1 to N" large - scale market applications. Its advantage lies in combining AI with complex national - level applications and quickly achieving commercialization.
Investment inspiration: Cross - border investors must understand this fundamental strategic difference. This is the key value that Beta Consulting creates for its clients when providing cross - border technological due diligence services - we help clients understand not only the technology itself but also the real potential and risks of the technology in a specific market environment.
This article is from the WeChat official account "Silicon Rabbit" (ID: gh_1faae33d0655). Author: Silicon Rabbit. Republished by 36Kr with permission.