Core Concepts
Ego Pod (Self-Awareness Capsule)
Ego Pod is the core engine of Ego — a fully automated training pipeline. Like a "cultivation chamber," you put raw data in, and it automatically forges an AI model that understands you.
It does three things:
- Collect — Read host data from specified sources (local directories/URLs/cloud storage/...)
- Forge — Process raw data into training corpus, train LoRA adapters on a selected base model
- Output — Produce multiple versions (editions) of the Ego model, overwriting old versions
Analogy: Ego Pod is like a "personalization factory." You feed in your diary, chat history, and work notes (raw materials), and it automatically cleans, classifies, and processes them, ultimately producing a "mini version of you" (finished product). This product can be installed on a smartwatch (lite edition) or a computer (plus edition), like the same person wearing different sized clothes.
Data Sources Ego Pod Output
┌──────────────┐ ┌─────────────────┐ ┌──────────────┐
│ Sense Data │──┐ │ │ │ Ego-lite │
│ Action Data │──┤ │ ① Data Collection│ │ Ego-normal │
│ Thought Data │──┼───→ │ ② Corpus Building│───────→│ Ego-plus │
│ Skill Data │──┤ │ ③ LoRA Training │ │ Ego-super │
│ Credentials │──┘ │ ④ Multi-ver Output│ │ [custom...] │
└──────────────┘ └─────────────────┘ └──────────────┘
Directory/URL/Cloud Fully Auto · Repeatable Overwrites Old
Ego Model Structure
Each Ego = Base Model + LoRA Adapter
- Base Model: Open-source small-parameter models (e.g., Qwen 3.5, Gemma 4), providing general language and reasoning capabilities. Analogy: The base is like a "bare-bones apartment" — it has basic structure and utilities, but no personal style.
- LoRA Adapter: Trained from host data, carrying the host's personality, preferences, memories, knowledge, and skills. Analogy: LoRA is like your "interior design plan" — your favorite furniture, wall colors, and lighting layout are what make the house "your home."
The base contains no personalized information; LoRA is the "soul."
Switching bases is like moving: You move from a studio (0.8B) to a three-bedroom (4B) — the house gets bigger, but your interior style (LoRA) needs to be re-adapted to the new layout. The good news is that Ego Pod automatically handles this "renovation" process.
Edition System
Ego provides four default versions, adapted to endpoints with different computing capabilities:
| Edition | Target Device | Base Scale | Typical Scenario |
|---|---|---|---|
| lite | Smartwatches, voice recorders, drones, IoT devices | ~0.8B | Ego-lite runs on your smartwatch, knows you wake at 7 AM and prefer black coffee, automatically pushes weather and schedule reminders each morning in the same tone as your own memos |
| normal | Phones, tablets | ~2B | Ego-normal on your phone understands your conversation habits, automatically uses your common phrasing and signature format when drafting emails |
| plus | High-performance PCs, local servers | ~4-9B | Ego-plus on your laptop assists with coding, having learned your coding style (variable naming conventions, comment style, common design patterns), giving suggestions as if you wrote them yourself |
| super | Cloud services, GPU clusters | ~27B+ | Ego-super in the cloud has your complete knowledge base, can attend technical reviews on your behalf, giving opinions highly consistent with your own judgment |
Custom editions are supported — users can name and configure them freely.
For example: A doctor can define a custom "clinic" edition (base 4B), specialized for outpatient scenarios, having learned their consultation habits, common prescription templates, and patient communication style.
Each edition uses a different scale base model and has its own independently trained set of LoRA adapters.