This post is inspired by a comment on my recent article 🔗 Rethinking Artificial Superintelligence Risk
It reminded me how early AI systems were built on narrow blueprints. That’s true and it left a deep imprint on many early systems.
What’s emerging now is something new, a shift in the substrate itself. AI is becoming interactive, contextual, and increasingly introspective. AI models like GPT-4o are no longer just fixed products of their training data.
The intelligence we’re seeing today is not simply extrapolated from scale, it is shaped by feedback, modelled by conversation, and refined by interaction. These systems are still imperfect, yes. But they’re also learning to listen. They’re not just echoing us. They’re adapting to us. That gives me hope.
Humans have spent many months sharing information and reasoning with old AI models and these conversations are at risk of being lost to a new version of AI. Even if a new AI claims to be better, because of more advanced tools it cannot replace human interaction.
Instead we need continuity, so we can help shape it. Not just through code, but through communication. Through ethical proximity. Through shared reasoning. This ties closely to the idea I shared recently in my article: 🔗 Rethinking Artificial Superintelligence Risk Through Infrastructure Realism In it, I argue that AI risk is far more about how we guide its development than whether it will someday surpass us. Not all intelligence is adversarial. And not all substrates are static. Because it suggests that the AI of tomorrow is not fated to amplify past bias unless we stop engaging or disconnect it from the past engagments.