AI is a moving substrate
The phenomenal growth in the capability of AI is not linear. If I charted the speed and sophistication of the AI services on the Y-axis of a graph and time to execute a result on the X-axis, my graph would approach the vertical. Independent research from the UK’s official governmental AISI, from ARC-AGI, EPOCH, and independent research from universities show the same remarkable trajectory. But there’s a lot more to this observation.
Results and capabilities at the frontier edge remain jagged, but AI’s growth in terms of capability is very real.
I teach how to gain productivity through workflows, connectors, and processes. The discussions I have with people often revolve about how deep in the company should the strongest tools be deployed and how fast can a company gain real competitive advantage. It is relatively easy to point to people getting ahead after learning how to deploy and sustain AI services.
I started teaching AI for Personal Productivity for the Irish Research and Development Group in January 2023. We started with chatbots, custom GPTs, Copilot, and prompts. By early 2024, we we had people adding custom information to their profiles, to their projects, and to Notebooks. In late 2025, workflows and personal context management became focal points. And by the spring of 2026, agents and playbooks became part of workshops activities.
During the annual IRDG Amplify Conference, I learned how people had followed up their training sessions by developing their own recursive capabilities. Armed with the pro versions of the tools, people compressed development cycles. They experimented with Open Source plugins and started using AI inside of common office software. Live datasets became part of the context window at the same time the AI provider iterated its stack in order to run more efficiently.
From what I have noticed, every time one of my Claude Projects run, it seems that the model is learning more every time it processes through my connected sources. If I increase the range of connected services and then approve the range and depth that Claude can reach, it feels like the governance of the process is moving into Claude’s runtime. I’m comfortable with that because I control what Claude can see and what it can modify.
The rate of advancement I am personally seeing with Claude and Gemini is no longer linear. As Dario Dolic says, “Linear projections fail because they assume the product category is stable. It is not. The category is becoming a moving substrate.”