Paul Visciano Blogs

Sci-Fi Labs · Essay

Local AI still looks like a chatbot

You can run the model at home now. The window is still ChatGPT. That is not the only shape memory can take.

Split: a phone-shaped stack of empty chat bubbles on the left, a spatial canvas of nodes and photos on the right
Same models. Two products.

You can run the model at home now. Open WebUI, LibreChat, AnythingLLM, LobeChat, Jan — they proved you do not have to send every prompt to someone else’s region. Ollama and llama.cpp handle the weights. Then you open the app and you are still in ChatGPT’s window. A text box. A scroll of bubbles. Memory as a transcript.

Self-hosting fixed the data path. It did not fix the interface.
Workshop hologram of orbital rings and schematic panels floating above a workbench
Movies already showed this. The assistant is in the room, not in a column of bubbles.

What the current tools actually ship

Open WebUI is the largest community option: chat, collections, tools, air-gap patterns. LibreChat is the team workbench. AnythingLLM is the document specialist. LobeChat is polish and plugins. Jan is the offline desktop client. Those projects matter. They also share one design assumption: the primary surface is a chatbot.

ToolWhat it isSurface
Open WebUILargest community optionChat
LibreChatTeam workbenchChat
AnythingLLMDocument specialistChat + docs
LobeChatPolish and pluginsChat
JanOffline desktop clientChat
Knowledge GraphSpatial memory3D canvas

You type. The model replies. History is a scroll. Documents become context for the next message. That pattern works. It is also the pattern ChatGPT trained the world on.

Why a canvas is a different product

The Knowledge Graph starts from a different question: what if photos, notes, conversations, and later video and audio were not buried inside threads, but laid out in space you can navigate?

Time becomes depth. A month is a place you fly to. A conversation is a node you land on, not only a transcript you re-read. Retrieval is not “paste more context into the box.” It is moving the camera to the region that already holds the answer.

That is closer to how people already think about their lives: places, periods, people, artifacts. It is also closer to what spatial computing and XR make possible once the UI is not glued to a phone-shaped chat column.

Spatial knowledge graph canvas — conversations, photos, and documents as nodes in 3D space instead of a chat window
Time as depth. A month is a place you fly to.
Chat stackKnowledge Graph
MemoryA transcript you re-readPlace, time as depth
RetrievalPaste more context into the boxFly the camera to the region
ProductCloud shape, local serverLocal shape, local server

Sovereignty without a new chat clone

If the only way to use a local model is still a chatbot, you have moved the server and kept the product shape of the cloud apps you were trying to leave. Open WebUI’s framing is useful: sovereignty is control over infrastructure, data path, models, identity, and audit. The same principle applies to the interface.

Sci-Fi Labs is building the personal version of that control as a spatial stack. Knowledge Graph is the memory layer. Where is Paul? is a life map. Musical Cubes is interaction as something you turn in 3D. The published sites are demos. The product is the copy that runs on your machine.

You can pair that surface with the same local runtimes everyone else uses. The point is not to reject Open WebUI or LibreChat. It is to notice that “self-hosted” and “still a chatbot” are two different design decisions.

More from Sci-Fi Labs

Data sovereignty is moving to the edge → KG · The canvas and the stack → The local agent that can answer “what have I been up to?” → Reclaim your data →