Main Theme
The overarching theme of Rob Ferguson’s 2026 keynote, effectively his “2024 Report Card,” is the vital evolution from relying on monolithic, third-party AI to creating customized systems where workflows are inherently “learnable.” The main message challenges organizations to stop “renting intelligence” and start “owning your own intelligence.” By shifting from being “AI talkers to AI doers,” businesses must leverage a “mixture of models,” build robust “harnesses,” and utilize precise measurement to break free from the hype cycles of waiting for the next frontier model, ultimately building their own custom AI factories.
Subthemes and Chapters
The 2024 Report Card and the Shift in AI
Ferguson opens by reflecting on his predictions from September 11, 2024. He highlights the monumental industry shift from “AI talkers to AI doers,” where models evolved from providing a single response to executing 40 to 50 actions for a given prompt. He notes the brief era of “token maxing”—where enterprises rapidly burned through their entire AI budgets in a single quarter—and the “open model winters” triggered when OpenAI introduced o1 on September 12, 2024. This winter, characterized by the belief that open models could never exceed closed ones, persisted until DeepSeek R1 launched on January 28, proving open models could actively compete in reasoning.
The New State of the Art: Harnesses and Routing
The frontier of AI is no longer a moat; it is a “mixture of models.” Ferguson emphasizes that the state of the art is now routed—using several models in tandem yields the best possible results. A model’s success now heavily depends on its “harness.” This infrastructure manages context across model calls, provides tools, and crucially allows the AI to inspect its own failures, change its approach, and try again. He illustrates this with Composer training, which utilizes full operating user interfaces, tools, and verbal commands, demonstrating that the harness and the environment fundamentally determine what a model learns.
Teaching Models How to Know
Addressing the challenge of keeping models updated with changing facts, Ferguson discusses retrieval. While retrieval remains an excellent method to preserve sources and provide current information, the ultimate goal has evolved. Organizations must move beyond teaching a model what to know, and instead teach it how to know. The objective is to help the model recognize a situation, find relevant information, choose a useful action, and verify if that action actually works.
Creating Your Own Frontier
AI-native teams in 2026 know better than to wait for the next model release to solve their problems. Instead, they focus heavily on measurement and evaluation. Above a certain capability threshold, expensive frontier AI offers diminishing returns. By balancing model behaviors, training specialized competencies, and measuring outcomes (like writing emails to measure team losses), companies stop pushing towards someone else’s benchmark. As Ferguson states: “If you can measure it, you can ship it.”
Owning Your Intelligence
The keynote concludes with a strategic call to action: make your work “learnable” by finding the seams in your organizational tasks and decomposing them into teachable, measurable parts. With platforms like Fireworks AI enabling both inference and training in one place, the final question for any organization is determining whether they are on a path to owning their own intelligence, or perpetually renting it to become someone else’s.
Inspiring Quotes
- “We went from an era of AI talkers to AI doers.”
- “The frontier is not a mold. It’s not abstract. It is knowable and demonstrable. That it is a mixture of models.”
- “We’re not teaching the model what to know. We’re teaching the model how to know.”
- “Once you start measuring and evaluating your AI work, you stop pushing towards someone else’s frontier. You start creating your own.”
- “Am I on my path to owning my own intelligence, or am I renting intelligence to make me somebody else’s intelligence?”
Key Action Points
- Implement Model Routing: Stop relying exclusively on a single closed frontier model. Utilize a “mixture of models” and route tasks to the best available models for specific steps to optimize trade-offs and achieve state-of-the-art results.
- Build a Robust Harness: Invest heavily in the environment and “harness” surrounding your LLM. Ensure it can manage context, provide production-like tools, and create feedback loops that allow the model to independently inspect failures, adjust its approach, and try again.
- Make Your Work Learnable: Abandon the outdated strategy of waiting for the next model update. Decompose your workflows into measurable parts, start evaluating your AI against your own internal success metrics, and train specialized behaviors to create your own AI frontier.