The Dream Was Never Just Intelligence
Long before AI became part of everyday life, most of us already had an idea of what it would eventually become.
Whether it was JARVIS from Iron Man, Samantha from Her, Baymax, or countless other fictional counterparts, the vision was remarkably consistent. These weren’t simply machines that answered questions. They remembered yesterday’s conversation. They learned your habits. They adapted over time. They felt less like software and more like relationships.
That dream still shapes how people talk about AI today. Even now, it’s common to see new products described as “your own JARVIS” because the idea resonates instantly. It isn’t really about the interface or the voice. It’s about continuity. The AI we imagined didn’t just know things, it became familiar.
Then Everything Changed
The transformer architecture fundamentally changed what artificial intelligence was capable of.
By solving one of the defining challenges of modern AI, it enabled models that could understand and generate language at a scale that once seemed impossible. That breakthrough transformed research, accelerated an entire industry, and brought AI into the hands of millions of people. It’s difficult to overstate how important that moment was. Today’s AI exists because of it.
Success Changed the Questions
As AI began weaving its way into our everyday work, learning, and creativity, the questions people started asking began to change.
At first, we wanted AI that could write. Then we wanted AI that could help us think. Eventually, we wanted AI that could remember what we’d already built together. People weren’t asking for better autocomplete anymore. They were looking for continuity. Not because today’s systems had failed, but because they had become useful enough for entirely new expectations to emerge.
That’s what makes this moment interesting. Every architecture reflects the questions it was designed to answer. When the questions change, the architecture eventually has to change too.
Why We Think This Is an Architectural Question
Much of today’s AI is remarkably good at producing answers, but familiarity comes at a cost. Each conversation requires reconstructing enough context for the model to understand where it left off. The longer a relationship becomes, the more information has to be reintroduced and recomputed before meaningful work can continue.
Our position paper explores why this isn’t simply a missing feature or a product decision. It’s a consequence of how today’s systems are structured.
That distinction matters because architectural questions rarely have simple fixes. They require architectural thinking.
The Questions That Guide Our Research
Our paper doesn’t argue that today’s AI is broken. In many ways, it’s one of the most successful technologies ever created. Instead, it asks whether that success has revealed the next set of architectural questions.
Can intelligence retain experience? Can it build understanding over time? Can it evolve through use? Those questions sit at the heart of our research.
An Invitation
We’re publishing this paper because we believe these questions deserve an open conversation.
Whether our ideas ultimately prove correct isn’t something a single company gets to decide. Progress happens when assumptions are examined, architectures are challenged, and new ideas are tested in the open. This article is only an introduction.
If you’d like to explore the full argument, including the historical context, architectural framework, and research behind these ideas, you can read the complete position paper below.
From the void,
- The EREBYX Team