Interface and Model
Large language models have redefined what we mean by intelligence. But the missing interface layer remains the core constraint on human-machine collaboration.
The emergence of large language models has shifted how we define intelligence in machines. A model that can reason across domains, generate coherent argument, and respond to nuanced instruction feels qualitatively different from its predecessors. The discourse has followed: much attention is paid to model capability, model alignment, model architecture.
What receives less attention is the interface.
The Interface Is Not a Detail
Interface here does not mean UI. It means the full set of mechanisms through which human intent, state, and context enter a system — and through which system outputs reach and influence humans. It includes input modalities, output channels, latency, framing, memory structure, and the conditions under which a system acts versus asks.
A powerful model operating through a narrow interface remains a powerful model. But its usefulness is bounded by what the interface can carry. If the interface strips nuance, compresses context, or ignores user state, the model reasons over an impoverished signal.
This is not an academic concern. It is the primary constraint on what AI systems can do in practice.
Capability Transferred Versus Capability Realized
There is a meaningful difference between what a model is capable of and what a deployed system realizes. This gap is almost entirely explained by interface quality.
Consider instruction-following. Models are capable of following complex, nuanced instructions. But most interfaces surface that capability only partially — through a text box, a prompt template, a rigid workflow. The interface design determines how much of the model's instruction-following capability actually reaches the user.
The same logic applies across dimensions: reasoning depth, personalization, error recovery, adaptive behavior. In each case, the ceiling is set by the model; the floor is set by the interface.
What Brain-Origin Signals Change
Language remains the most expressive input modality available to deployed AI systems. But language has a fundamental limitation: it is a deliberate act. By the time a user has formulated a sentence, they have already translated their internal state through several layers of processing.
Brain-origin signals operate earlier in that chain. They can carry information about attention, effort, uncertainty, and cognitive load before those states resolve into language. This is not a replacement for language — it is a complement that narrows the gap between internal state and system input.
For intelligence systems, this matters because the quality of reasoning is downstream of the quality of context. A system that knows a user is uncertain will handle a query differently from one that treats every response as equally confident. A system that can detect that attention has shifted can interrupt or adapt rather than proceeding with a degraded interaction.
Interface Design as Systems Thinking
The practical consequence is that building useful AI systems requires treating interface design as a first-order technical problem — not as a product concern to be addressed after the model is trained.
This means asking:
- What signals are available at the point of interaction?
- Which of those signals are informative about user state, intent, or context?
- How should those signals be structured so that downstream models can use them?
- What are the failure modes when signals are missing, delayed, or misleading?
It also means accepting that the right interface for a given task may not exist yet. The keyboard and the touchscreen were not discovered — they were designed, with substantial iteration, to match input modality to what systems needed to function well.
The same work is ahead for brain-origin interfaces.
The Research Agenda
BBCI's position is that the interface layer is where the most consequential near-term work on human-AI interaction lives. Not because models are unimportant, but because model capability has outpaced interface capability — and that gap determines what AI systems can actually do with the people who use them.
The research questions that follow are concrete: What signals are useful across different task types? How should uncertainty in signal interpretation propagate through system behavior? What does legible, trustworthy interface behavior look like when the input includes physiological data?
These are not soft questions. They have answers that can be measured, iterated on, and built into systems. That is the work.