Where is the future of the AI infrastructure layer?
Over the past decade, Lightspeed, a veteran U.S. fund, has been working with outstanding companies in the AI/ML field, the platforms they build, and the customers they serve to better understand how enterprises are thinking about Gen-AI. Specifically, Lightspeed studied the underlying model ecosystem and asked questions such as “Will there be a winner-take-all dynamic for the best model?” and “Do enterprise use cases call OpenAI’s API by default, or will actual use be more Diversity?" and other questions. The answers will determine the future growth direction of this ecosystem and the flow of energy, talent and capital. 1. Classification of model ecosystems Based on our learning, we believe that a Cambrian explosion of models is about to occur in AI.Developers and businesses will choose the model that best fits the “task to be done,” although use during the exploration phase may appear more focused. A likely path for enterprise adoption is to use large models for exploration, gradually moving to smaller specialized (tuned + refined) models for use in production as their understanding of the use case increases. The diagram below outlines how we see the underlying model ecosystem evolving. Category 1: “Brain” model These TG Number List are the best models and represent the cutting edge of modeling. This is where those exciting and magical demos come from. These models are often the first things developers consider when trying to explore the limits of what AI can do for their applications. These models are expensive to train and complex to maintain and expand. But the same model can take the LSAT, MCAT, write your high school essay, and interact with you as a chatbot.
https://lh7-us.googleusercontent.com/ws-oFF1n4H2FEaFaW7M2PMdY-pNGA3Pw98_XyTaJFL9cdYBVEBXSRottucORhJhLecu1OLbsPMidaTHa6P-hkJfWcMvv2UT_VB8eg2LOhT4lyhayplzgLzaNsVKXS1EhfFQ2kZlI4_af6Cs9mEFgDxI
Developers are currently conducting experiments on these models and evaluating the use of AI in enterprise applications. However, general-purpose models are expensive to use, have high inference latency, and may be overkill for well-defined constrained use cases. The second problem is that these models are generalists and may be less accurate on specialized tasks. (See this Cornell paper.) Finally, in almost all cases, they are also black boxes, which can create privacy and security challenges for enterprises that are trying to leverage these models without giving up their data assets. OpenAI, Anthropic, Cohere are some examples of companies.
頁:
[1]