| --- |
| license: mit |
| datasets: |
| - interstellarninja/hermes_reasoning_tool_use |
| - crownelius/Opus-4.6-Reasoning-2100x-formatted |
| - ronantakizawa/github-top-code |
| base_model: |
| - nvidia/Cosmos-Reason2-2B |
| --- |
| |
|
|
|
|
| # Karla C1: The Adaptive Brain for Physical AI |
|
|
| **Karla C1** is a novel, continuous-learning architecture designed for Physical AI and embodied agents. Built on top of **NVIDIA Cosmos-Reason2-2B**, it solves the problem of catastrophic forgetting in robotics by implementing a **Nested Learning Architecture** [1] with *Surprise-Based Plasticity*. |
|
|
| ## The Vision: Physical AI on the Edge |
|
|
| Robots in the real world encounter new tools and constraints daily. Karla allows an agent to learn a new physical tool or API *on the fly* during inference, modifying its own weights in seconds on consumer hardware (e.g., RTX 4060 Ti), without ever forgetting its foundational knowledge. |
|
|
| ## Architecture: Nested Learning |
|
|
| Instead of a standard transformer pipeline, Karla acts as a multi-frequency brain based on the Nested Learning (NL) paradigm [1]: |
|
|
| 1. **L0 (The Subconscious): NVIDIA Cosmos-Reason2-2B (Frozen).** Provides world-class reasoning, semantic understanding, and syntax. Loaded in 4-bit to save VRAM. |
| 2. **L1 (The RAM): Dynamic Knowledge MoE (64 Experts).** A fast, GPU-based Mixture of Experts. It updates *live during inference* using Delta Gradient Descent [1], memorizing new facts and tool syntaxes instantly. |
| 3. **L2 (The Frontal Lobe): Continuous Thought Machine (CTM).** Based on Sakana AI's research [2], this is a parallel sequence-level reasoning module. It takes Cosmos's hidden states + L1 knowledge and "thinks" for $T$ internal ticks before outputting an action plan. |
|
|
| > [!WARNING] |
| > The L2 module currently faces stability issues where the CTM's internal dynamics can lead to token repetition during long reasoning chains. |
|
|
| ## Quickstart |
|
|
| ```bash |
| pip install torch transformers datasets pandas accelerate bitsandbytes |
| cd Karla |
| python chat.py |
| |
| ``` |
|
|
| ## TODO |
|
|
| * [ ] Train Longer |
| * [ ] Make L1 Inference Learning Intensity variable |
| * [ ] Give the model control over itself |
| * [ ] Fix stuttering |
|
|
| ## Datasets Used |
|
|
| * `interstellarninja/hermes_reasoning_tool_use` |
| * `crownelius/Opus-4.6-Reasoning-2100x-formatted` |
| * `ronantakizawa/github-top-code` |
|
|
| ## References |
|
|
| [1] Behrouz, A., Razaviyayn, M., Zhong, P., & Mirrokni, V. (2025). **Nested Learning: The Illusion of Deep Learning Architecture**. *Google Research*. [arXiv:2512.24695](https://arxiv.org/abs/2512.24695) |
|
|
| [2] Darlow, L., Regan, C., Risi, S., Seely, J., & Jones, L. (2025). **Continuous Thought Machines**. *Sakana AI*. [arXiv:2505.05522](https://arxiv.org/abs/2505.05522) |
|
|
