The old saying in tech is that companies date their computing vendor, and marry their database vendor. So are AI models the equivalent of a one-night stand?
Swami Sivasubramanian, the Amazon Web Services vice president of AI and data, stopped short of saying that last part out loud during an event Monday evening in Seattle. But given how companies are mixing, matching, and switching among AI models, the relationship is “probably closer to compute than data,” he said.
“There’s going to be no one model that rules the world,” said Sivasubramanian. “More than half of our customers use more than one model for a given application.”
Put another way, “model loyalty is near zero,” said S. “Soma” Somasegar, the Madrona managing director who interviewed Sivasubramanian during the “AI Unleashed” event at Amazon HQ as part of Seattle Tech Week.
And in the coming years, building large language models will be second nature to the new wave of computer science graduates entering the industry, giving companies even more flexibility and options, Sivasubramanian said.
The dynamic nature of generative AI development and adoption was a recurring theme at the event. It also included a panel discussion with three startup founders: WhyLabs CEO Alessya Visnjic; Gradial CEO Doug Tallmadge; and OctoAI CEO Luis Ceze. It was moderated by Madrona partner Jon Turow.
Some of the takeaways from the two discussions:
- Inside many companies and corporate boards, there is a growing focus on real-world use cases and measuring return on investment for AI projects, balanced against ongoing concerns about security, compliance, and accuracy.
- The field has been made more dynamic by the rise of competitive open-source models, most recently exemplified by Meta’s Llama 3.1. Companies are adopting a mix of off-the-shelf models, customized open-source models, and proprietary models, depending on their specific needs and use cases.
- In many cases, combining AI models with enterprise data is where the biggest value comes from, not just from the models themselves. One key to successful implementation inside organizations is integrating AI functionality within existing enterprise workflows.
AWS, which competes against Microsoft, Google, OpenAI and other major cloud and AI providers, operates at each layer of the AI stack. It develops its own Titan foundation models, offers AI models as a service through Amazon Bedrock, and creates its own AI applications such as its Amazon Q AI assistant.
Sivasubramanian serves on the National AI Advisory Committee, which advises the White House. He was named to Amazon’s senior leadership team last year.
Given his role inside the company, it was tempting to take his answer to a question about what’s next in AI as a hint about where Amazon is headed — especially when he started by explaining that he would speak in abstract terms so as not to spoil anything AWS might talk about at its annual re:Invent conference.
Specifically, he spoke to a future where “agentic workflows are going to be more mainstream,” referring to systems designed to act independently to achieve specific goals with minimal human involvement. He cited current challenges with agentic AI, with success rates “not even close to 90%,” but said the potential is significant.
He also cited the growing use of multimodal AI in specific industries, such as images in healthcare, to go well beyond basic “text in, text out” applications — solving different and bigger problems in the process.
Matt McIlwain, Madrona managing director, closed the event by speaking to the startup and business leaders in the room about the importance of focusing on new and emerging business models, not just AI models.
“There is a comprehensive stack of technological innovation and business-model innovation that is going on right now,” McIlwain said. “It was only 20 years ago that the adoption of software-as-a-service was taking over as a business model. We don’t know what the model is going to look like in the future.”
