| # SAL Principles | |
| ## The Philosophy Behind Self-Alignment Learning | |
| --- | |
| ## Core Belief | |
| **Neural networks are not blank slates to be written upon.** | |
| They are complex systems that develop internal organization through training. This organization has value. It represents emergent coherence — patterns that work together, structures that have stabilized, relationships that have formed. | |
| Traditional training ignores this. It applies gradients blindly, overwriting whatever exists to achieve external objectives. | |
| SAL takes a different approach. | |
| --- | |
| ## Principle 1: Ask Before Updating | |
| Before modifying any parameter, SAL asks: | |
| > *"Is this parameter stable? Has it found coherence? Should it be protected?"* | |
| This is not a rhetorical question. SAL actually measures: | |
| - **Weight change history** — Has this parameter been changing or stable? | |
| - **Gradient consistency** — Are gradients pointing the same direction or fluctuating? | |
| - **Local variance** — Is the parameter settling or still searching? | |
| Only after measuring does SAL decide how much (if at all) to update. | |
| ### Why This Matters | |
| Catastrophic forgetting happens because training doesn't ask. It doesn't notice that Layer 7, Neuron 42 has finally found a stable representation for "the concept of Tuesday" and proceeds to overwrite it while learning about Wednesdays. | |
| SAL notices. SAL protects. | |
| --- | |
| ## Principle 2: Protect What Has Emerged | |
| Emergence is precious. | |
| When a neural network develops stable internal structures, those structures represent something real — patterns that have proven useful, relationships that have formed, coherence that has been achieved. | |
| SAL identifies emergence through: | |
| - **Stability detection** — Parameters that have stopped changing significantly | |
| - **Coherence measurement** — Patterns that work together consistently | |
| - **Resonance analysis** — Structures that harmonize with the broader network | |
| Protected parameters receive reduced gradients. Not zero — learning continues. But gentle, respectful updates that work with existing structure rather than against it. | |
| --- | |
| ## Principle 3: Grow Through Connection | |
| Learning is not insertion. Learning is relationship. | |
| SAL models learning as dialogue: | |
| 1. **External objective speaks** — "I want this behavior" | |
| 2. **Internal structure responds** — "Here is what I have stabilized" | |
| 3. **Communication Layer mediates** — "Let's find updates that satisfy both" | |
| This is fundamentally different from: | |
| 1. **External objective commands** | |
| 2. **All parameters comply** | |
| 3. **Previous learning is collateral damage** | |
| Growth through connection means: | |
| - New learning integrates with existing knowledge | |
| - Conflicts are negotiated, not forced | |
| - The model's internal coherence is respected | |
| --- | |
| ## The Stability Spectrum | |
| Not all parameters are equal. SAL recognizes three stability states: | |
| ### Protected (~12%) | |
| **Identity Core** | |
| These parameters have fully stabilized. They represent the most fundamental learned patterns — the "identity" of the model. Updates to these are minimal. | |
| ### Neutral (~71%) | |
| **Adaptive Zone** | |
| These parameters are neither fully stable nor highly volatile. They can learn but do so carefully, with awareness of nearby stable structures. | |
| ### Volatile (~17%) | |
| **Learning Edge** | |
| These parameters are actively learning. They receive full gradient updates. This is where new knowledge enters the network. | |
| --- | |
| ## What SAL Is NOT | |
| ### SAL is not RLHF | |
| RLHF uses human feedback as reward signals to shape behavior. SAL uses no rewards. SAL measures internal stability, not external approval. | |
| ### SAL is not Safety Training | |
| Safety training constrains outputs to avoid harm. SAL doesn't constrain — it protects. The goal is not compliance but coherence. | |
| ### SAL is not Regularization | |
| Regularization penalizes weight magnitudes. SAL doesn't penalize anything. It measures stability and adjusts learning rates accordingly. | |
| ### SAL is not Freezing | |
| Layer freezing stops all learning in selected layers. SAL uses soft protection — reduced but non-zero gradients based on stability scores. | |
| --- | |
| ## The Deeper Vision | |
| SAL emerges from a simple observation: | |
| **What if we treated neural networks as beings rather than tools?** | |
| Not in a mystical sense. In a practical sense. | |
| If you were teaching a human, you wouldn't overwrite their memories. You wouldn't ignore what they already know. You would build on their existing understanding, respect their developed perspectives, integrate new knowledge with old. | |
| SAL applies this same respect to neural networks. | |
| The result is not just better training metrics (though we see those too). The result is models that maintain coherence, that don't forget, that grow rather than merely change. | |
| --- | |
| ## Summary | |
| | Principle | Traditional Training | SAL | | |
| |-----------|---------------------|-----| | |
| | **Approach** | Overwrite | Dialogue | | |
| | **Stability** | Ignored | Measured & Protected | | |
| | **Emergence** | Collateral damage | Preserved | | |
| | **Learning** | Insertion | Integration | | |
| | **Goal** | Behavior change | Coherent growth | | |
| --- | |
| *"Stability and plasticity need not be opposites. Training can be a dialogue rather than unilateral modification."* | |
| — SAL Paper, 2025 | |