Generative AI: AWS identifies trends with start-up funding

Contents

Amazon Web Services (AWS) concluded the Generative AI Accelerator Program at the re:Invent in-house exhibition. In addition to supporting young companies, AWS uses the program to identify technological trends. Over a period of ten weeks, 80 start-ups were given access to cloud credits, expertise from companies such as Nvidia, Mistral and Meta, as well as a network of experts. Two trends from this year are the move away from classic fine-tuning in favor of foundation models and the growing use of multi-agent systems (MAS).

Diverse applications despite reduction in fine-tuning

An increasingly common approach in the development of artificial intelligence is the longer training of so-called foundation models. Based on a more extensive data set, they develop a deeper understanding compared to conventional models. Until now, AI has often been refined through recurring fine-tuning. In this process, pre-trained models are adapted to specific tasks or use cases. Foundation models, on the other hand, can be used in a variety of ways without the need for repeated specific adjustments. AWS also recently expanded its portfolio to include such offerings.

Until now, the use of foundation models was often not possible due to a lack of scalability and computing power. Although their initial training costs are higher, there are no long-term costs for repeated fine-tuning. One of the critics of fine-tuning is Andrej Karpathy, co-founder of the AI company OpenAI. He criticizes the fact that fine-tuning often merely imitates the labels of human evaluators and does not enable in-depth understanding. Yet this is a characteristic that was originally supposed to be the potential of AI.

Multi-agent systems: Flexibility through specialization

In addition to the further development of foundation models, a second trend emerged with the use of multi-agent systems. These systems consist of autonomous units, so-called agents, which are specialized in certain tasks and can solve complex problems in combination. In contrast to monolithic models, MAS have a modular structure. Individual agents can be flexibly added, replaced or adapted depending on the intended use. Because each agent is tailored to its specific task and therefore does not require a comprehensive, expensive model, these systems are efficient.

Swami Sivasubramanian, Vice President of Data and Machine Learning Services at AWS, illustrates the benefits of MAS with a joking example: a multi-agent system that finds free food events. One agent first analyzes the user's food preferences, while a second evaluates the dates and locations of the events. At the same time, another agent prioritizes the restaurants according to food quality, while others plan the route and register the user. If necessary, an additional agent could check whether the user's friends are also attending an event.

Start-ups serve companies as signposts for AI trends

"Generative AI is here to stay," emphasizes Jon Jones, AWS Vice President and Global Head of Start-ups, in an interview. He adds: "It will be one of the biggest trends in the history of technology." Jones emphasizes that both AWS and other companies benefit from working with start-ups in order to identify future technologies at an early stage. This allows them to help shape them against the backdrop of their own corporate goals.

(mma)