AI Service Dilemma

AI-driven services face challenges such as high computational costs, data monopolies, and risks of user privacy exposure. Whether it's the training cost of Nvidia's H100 or the inference cost of Groq LPU, they are far from being widely adopted; the current Transformer model, based on the fundamental building block of deep learning MLP (multi-layer perceptron--fully connected feedforward neural network), has a high demand for computational resources, resulting in expensive hardware costs.

Moreover, the monopoly of data means that giants with access to and processing capabilities of large amounts of data control the development of AI. This not only leads to increased costs of AI development but also limits the innovation capabilities of individuals in the ecosystem. The risk of user privacy exposure further exacerbates the challenges of AI-driven services in applications. Model training requires a large amount of user data, making it susceptible to misuse or leakage. Ensuring the security and privacy of data becomes a key issue in AI development.