ELI VR 0.9.0

ELI VR 0.9.0 — ELI builds, and the rooms are ELI's own

With ELI v3 beside ELI VR, ELI now knows where it is, builds in the rooms from what you say, has a room of its own, and uses whichever model you choose. ELI VR still runs on its own, and with an older ELI v3 it still gives presence and talking.

Needs for the new parts: ELI v3 3.0.134 or newer, which speaks world protocol 2. With 3.0.131 to 3.0.133, ELI VR agrees version 1 and says so.

How far it has been tried. Flat, without a headset, against ELI v3's real server: building from sentences with a real and very small model (0.6B), the mind panel, and free-form things and a painted sky drawn from edits sent by hand. ELI's room was run in a Rift CV1 at 90 frames a second before the free-form work. Not yet tried: any of the building spoken into a headset, a model large enough to be offered free-form parts, changing model from inside the room, and ELI building of its own accord with a real model. Windows and macOS builds come from CI.

ELI's mind is your choice

No model is named or assumed.

• Opened from ELI v3's window: the model already loaded there answers. Nothing is loaded twice.
• In the room: the ELI'S MIND panel lists ELI's models by name. Press one and ELI loads it.
• Before starting: --eli-model=<name>, or python3 tools/eli_vr.py model <name> to keep it.

The status line always says which model ELI is using.

ELI knows where it is

ELI is told the room and what the place is, where you stand, what you are looking at and which of its things are near, for as long as you are there and no longer. Ask "where are we".

ELI builds

"Build me a tall red pillar." "Put a lamp to my left." "Make the pillar green and bigger." "Move it behind me." "Undo." Things rise out of the floor where you asked, and you can grab and move them. Every change goes through ELI's world's own policy and ledger.

• With any GGUF. ELI's answer is held to an exact grammar, and what you plainly said (a colour, a size, "behind me", "the orb" when there is none) is worked out without the model. A 0.6B model manages every plain request.
• In any way ELI likes. A thing is a named shape, or anything made part by part: boxes, balls, cones, rings, a lathe, an extruded outline, a tube along a path, text, each with its own colour, metal, glow, clearness and motion. A model of three billion parameters or more is offered this; a smaller one keeps to the named shapes.
• If nothing changed, ELI says so, whatever the model claimed.

A room of its own

ELI'S ROOM on the console: open ground under an open sky, with nothing there but what ELI has made. The sky is ELI's to choose, by name or painted from two colours, stars and a sun placed wherever ELI likes. So is the ground: any colour, flat or rising into hills.

Of its own accord

Nothing is scheduled and nothing is chosen for ELI. When something happens in ELI that has words (an insight, a goal taken on or finished, an improvement ELI proposes) the moment is kept. When ELI is free it may make something of that moment in its room, from its own records: its persona, its latest insight, its goals, its recent memories. What ELI makes is tied to the record the moment came from. ELI may also make nothing. ELI: BUILD asks it now.

ELI's records in its rooms

The Memory Archive holds ELI's real recent memories as lights, each with its words. The Upgrade Bay holds the goals ELI has now.

Under it

• The world protocol has versions, agreed on connecting, with a plain refusal when there is none in common.
• A failure says which failure it is: ELI is not running, ELI has no model loaded, ELI did not catch that, ELI's voice is not available.
• What you say in the headset is your turn, in the same conversation and memory as anywhere else. ELI VR is a surface of yours, not another user.

Known gaps

• ELI is a light, not a figure.
• A very small model keeps to named shapes, says little worth hearing about what it made, and unasked often makes nothing.
• Only ELI's own room can be given a new sky.
• The Evidence Wall and Anomaly Room do not yet show ELI's real claims. ELI cannot yet paint pictures of its own, and you are not yet told on arriving what ELI made while you were away.
• The desktop screen is still view only.
• Clean shutdown of the VR runtime is Linux only. No native app for standalone headsets, no VR on macOS, builds are not code-signed.