Inside Amazon’s Model Factory: From Chatbots to Agents
Post on
October 6, 2025
Post on
October 6, 2025
There has been a debate about whether “one model will rule all” or whether the world will have a “collection of models” that people can pick and choose from depending on price, performance, latency, accuracy, and the use cases. Amazon is betting on the latter: There is room for different models, and Amazon is doing its part by building a set of models (Nova family of models) that are trained, tested, and improved inside one of the world’s most complex companies.
At the 2025 IA Summit, I sat down with Rohit Prasad, Amazon SVP and Head Scientist for AGI, to talk about how this “model factory” works. What emerged was a rare glimpse into how Amazon is deploying AI across every corner of its business, from fulfillment centers to advertising, and what it means for the next era of intelligent agents.
In this blog, you’ll find the highlights: Amazon’s model factory strategy, the Nova family of models, the challenge of the last 10%, and Rohit’s vision of AI agents that don’t just talk but do.
Amazon’s strategy represents a fundamental departure from the industry’s pursuit of a single, dominant model. Eight Nova models released over eight months, with tens of thousands of customers using them daily, represent what Rohit describes as building infrastructure for what he calls a “model factory” — and a system of “reinforcement learning gyms,” where every internal application becomes an environment for models to learn and improve.
“We don’t think of it as one model at one time. This is not like a waterfall-ish development,” he said. “You should be thinking of this as where we are essentially building a model factory that can release a lot of models at a fast cadence.”
The Nova family exemplifies this approach: perception models that understand any input modality, content creation models like leveraging Rohit’s speech recognition background, handles conversational interactions with speech-to-speech capabilities.
But the most intriguing addition is Nova Act, an early preview model focused on UI automation and computer use — what Rohit sees as a key milestone toward AGI.
“One of the key things is that it is essentially a particular milestone: Can the AI do everything you and I can do at expert level through a computer?”
Despite the rapid progress, Rohit identified a critical challenge that resonates throughout the industry: “The first 80% of the way is super easy with the foundation model — you prompt it and you suddenly see, ‘whoa, it can do all these things,’ but the next 10% is incredibly hard.”
That final journey requires API integrations, compliance with standards like MCP, and customization for specific use cases. Amazon addresses this with what Rohit calls “the most comprehensive set of customization tools for any frontier model,” including fine-tuning and post-training capabilities.
The company’s Bedrock platform philosophy of providing model choice becomes crucial here.
That philosophy of model choice also extends to partnerships: Amazon has maintained a deep collaboration with Anthropic, ensuring customers can choose from Claude as well as Amazon’s Nova family. As Rohit put it, there’s “no one model that does everything really well,” which makes breadth of options on Bedrock a core strategy.
One of the most telling insights from our conversation was Rohit’s observation about internal productivity gains: “I want AI to do the muck for me, not the creative work for me.”
This philosophy is playing out across Amazon’s engineering teams, where AI handles legacy code transformations, Java version upgrades, and deployment automation — tasks that are necessary but not creative. Tools like Kiro for coding and Q CLI for command-line development are improving productivity by taking care of mundane work.
“That’s where I think a lot of work is happening, which doesn’t get enough attention, but that’s something that we should be really cherishing about how businesses are using AI internally,” he said.
As we wrapped up our conversation, it was clear that Rohit sees this as just the beginning. Amazon is still in “day one” mode with AI, and the shift from chatbots to agents represents a fundamental evolution in how we interact with technology.
“We are now moving from chatbots that just tell you things to agents that can actually do things,” he said.
The reasoning revolution, as we’ve been calling it at Madrona, is creating machines that appreciate rather than depreciate over time — systems that continuously learn and improve rather than becoming obsolete
For Rohit, who grew up in the “Star Trek” era dreaming of natural conversation with computers and spent 14 years in R&D before getting the “once in a lifetime opportunity” to build Alexa, the measure of success isn’t just technical benchmarks but real-world impact: reliable agents that can decompose complex tasks, integrate diverse knowledge sources, and execute with precision.
From predicting equipment failures to accelerating materials science research, AI is moving from experimental to essential. The childhood dream of talking to machines has evolved into something bigger: reasoning machines that can not only understand what we want but actually get it done.
