Product designer Maxim Kich recently drew attention with a post on Medium arguing that AI interfaces in their current form need to be fundamentally changed.

He studied linguistics and psychology before moving into design.

His argument can be summarised as follows. AI models have become smart enough to connect multiple threads of conversation, remember earlier hypotheses and even point out thoughts people have not fully organised. Yet that intelligence is still trapped in a chat window like a 2015 messenger app.

“The longer the conversation gets, the shorter the scroll bar becomes. Good ideas get buried and disappear in piles of tool calls and outputs,” he said.

According to him, chat is not the structure of communication itself but merely a result that lists communication in chronological order. People can only speak one thing at a time. In this context, Kich cited a 2024 Nature paper by Fedorenko’s research team and said, “Language and thought are not the same. Sentences come out in sequence, but the thoughts contained in them can be spatial and simultaneous.”

Kich likened this to a code repository. Developers do not reread commit histories from the beginning to understand a program. They look only at the current state of the codebase. By contrast, when talking with AI, people must scan all past messages each time to reconstruct the current context. “So it is natural that context keeps disappearing,” he said.

Kich said he likes mind maps but drew a line, saying a single mind map cannot explain the entire future of AI interfaces.

He also introduced research examples aimed at going beyond the limits of AI interfaces. The first is Graphologue, which converts AI answers into node-link diagrams. The second is Sensecape, which lets users switch between a canvas view and a hierarchical view.

The third is Branchat, which splits conversations into multiple branches and allows users to return to a desired point in time. He also cited examples including MeetMap, which organises meeting contents in a map format, and Orality, which converts voice conversations into concept maps.

He cited memory as something that needs fixing before interfaces. Citing the “LLM Wiki” approach proposed by Andrej Karpathy, he introduced an approach that builds knowledge in a continuously updated wiki format instead of refeeding the entire conversation log each time. He explained that Microsoft Research’s GraphRAG is a similar attempt.

Kich ultimately appears to imagine a structure in which chat, kanban boards, mind maps and timelines all exist as different screens on a single context. He described this as a “Context Operating System” that forms a layer in itself and is not dependent on individual sessions or specific models. “Directly connecting the brain and the computer is still a long way off. Before that, we can build interfaces that operate one level above sentences,” he wrote to conclude.