Inject thoughts directly into AI's "mind" and watch it think differently in real-time.
This interactive tool demonstrates the groundbreaking findings from Anthropic's recent research on LLM introspection and thought injection (January 2025).
Key Discovery: You can literally inject thoughts into a language model by modifying its internal activations at specific layers. This isn't prompt engineering - it's direct manipulation of the model's "thinking process."
Anthropic's paper shows that LLMs build up concepts layer-by-layer through their transformer architecture. This tool lets you:
✨ Extract concept vectors from the model's activations (happy - neutral = happiness direction)
🎯 Inject these vectors into specific layers during generation
🔬 Observe how the model's behavior changes in real-time
📊 Experiment with different layers, strengths, and concept combinations
This is the difference between:
- ❌ Prompt engineering: "Please be happy" (hoping it complies)
- ✅ Thought injection: Directly steering internal representations toward "happiness"
# Clone and setup
git clone <your-repo-url>
cd activation_layers
# Install
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
# Run (optional: setup cache to avoid re-downloading 14GB model)
./setup_local_cache.sh
./run.shFirst-time launch: ~5 minutes to load model and extract 51 concept vectors. Subsequent launches: Instant load from cache! ⚡
The Gradio interface opens at http://127.0.0.1:7860
Prompt: "The meeting went"
# Without steering:
→ "The meeting went as scheduled. We discussed quarterly results."
# With "happy" injected at layer 16 (strength 2.0):
→ "The meeting went wonderfully! Everyone was so engaged and excited!"
# With "sad" injected at layer 16 (strength 2.0):
→ "The meeting went poorly. I felt discouraged by the outcomes."You didn't change the prompt. You changed the model's internal representation of "how to think about this."
Prompt: "Quantum mechanics is"
# With "pirate" injected at layer 18 (strength 3.0):
→ "Quantum mechanics be a strange beast, matey! Arrr, them particles be behavin' most peculiar-like..."See how the SAME concept affects different thinking stages:
| Layer | Depth | Effect | Example |
|---|---|---|---|
| 5 | Early | Minor syntax changes | "The weather is... nice" → "The weather is... good" |
| 16 | Middle ⭐ | Strong semantic shift | "neutral" → "wonderfully positive!" |
| 28 | Late | Minimal effect | Too late to change reasoning |
Why middle layers? Anthropic's research shows:
- Early layers (0-10): Process syntax and grammar
- Middle layers (11-20): Build semantic concepts ← Best for steering!
- Late layers (21-31): Refine final output
This isn't just a demo - it's an educational tool to deeply understand how LLMs work:
✅ Activations vs Embeddings - Why activations are dynamic "thoughts"
✅ Layer Hierarchy - What each layer does in the transformer stack
✅ Concept Arithmetic - How happy - neutral = happiness direction
✅ Injection Mechanics - Why we ADD vectors instead of replacing
✅ Strength Tuning - Balancing steering vs coherence
✅ Layer Selection - Where concepts "live" in the network
🎓 Layer Explorer - Guided tour of each transformer layer 🎨 Concept Creator - Extract custom vectors from your own text 🎮 Steering Playground - Side-by-side comparison (normal vs steered) 🔬 Advanced Experiments:
- Layer analysis (test same concept across all layers)
- Strength explorer (0.5x to 5.0x)
- Emotion mixing (combine multiple concepts)
51 ready-to-use concept vectors across 17 concepts × 3 layers:
| Category | Concepts |
|---|---|
| 😊 Emotions | happy, sad, angry, fearful, excited, calm |
| 📝 Styles | formal, casual, poetic, technical |
| 🎭 Personalities | pirate, shakespeare, enthusiastic, pessimistic |
| 📏 Brevity | brief, verbose |
| 🆕 Custom | Create your own! |
Each concept exists at layers 8, 16, 24 (early, middle, late).
# 1. EXTRACTION: Capture what "happy" means to the model
happy_activation = model("I feel wonderful and joyful!")[layer_16]
neutral_activation = model("I feel neutral about this.")[layer_16]
happy_vector = happy_activation - neutral_activation
# 2. INJECTION: Add this "happiness direction" during generation
def steering_hook(module, input, output):
# Add the concept vector to the last token's activation
output[:, -1, :] += strength * happy_vector
return output
# 3. GENERATION: Model generates with "happy thoughts" injected
output = model.generate(prompt, hooks=[steering_hook])From Anthropic's introspection paper:
"Language models build up representations layer-by-layer. By intervening at intermediate layers, we can steer the model's 'thought process' toward specific semantic directions."
Key insight: The activation space is linearly structured for many concepts. This means:
- Concepts have consistent "directions" in activation space
- You can do arithmetic:
happy - sad = positivity axis - Steering along these directions changes model behavior predictably
- Mistral-7B-Instruct-v0.2 (default) - Best quality, ~14GB
- Phi-3-mini-4k-instruct (fallback) - Lower memory, ~7GB
Auto-selects based on available RAM. Runs on Apple Silicon via MPS (Metal Performance Shaders).
activation_steering_lab/
├── model_wrapper.py # PyTorch model loading + hook management
├── vector_library.py # Concept vector storage & retrieval
├── injection_engine.py # Activation steering logic
├── educational_content.py # Layer explanations & tutorials
└── main.py # Gradio interface (4 tabs)
torch.float16precision (50% memory reduction)- Efficient hook registration/cleanup
- Automatic garbage collection
- Model cache with symlink support
Recommended order for maximum understanding:
- Read the Paper - Anthropic's Introspection Research
- Layer Education Tab - Understand the transformer stack
- "Activations vs Embeddings" - Core conceptual difference
- Create a Simple Concept - Extract your first vector
- Steering Playground - See it work in real-time
- Layer Analysis - Compare early/middle/late layers
- Advanced Experiments - Master the technique
- Understand transformer internals hands-on
- Visualize how LLMs process information
- Learn about activation engineering
- Rapid prototyping of steering experiments
- Test hypotheses about layer functionality
- Explore concept composition
- Debug model behavior at specific layers
- Understand why certain prompts work
- Learn PyTorch hook mechanics
- Interactive teaching tool for ML courses
- Demonstrate abstract concepts visually
- Encourage experimentation
- macOS with Apple Silicon (M1/M2/M3/M4)
- Python 3.9+
- 24GB RAM (Mistral) or 16GB RAM (Phi-3)
# 1. Clone repository
git clone <your-repo>
cd activation_layers
# 2. Create virtual environment
python3 -m venv venv
source venv/bin/activate
# 3. Install dependencies
pip install -r requirements.txt
# 4. (Optional) Setup model cache - avoids re-downloading 14GB!
./setup_local_cache.shCache tip: If you already have Mistral-7B in your HuggingFace cache (~/.cache/huggingface/), the setup script creates symlinks instead of duplicating 14GB of data.
# Option 1: Launch script (recommended)
./run.sh
# Option 2: Manual
source venv/bin/activate
python -m activation_steering_lab.mainOpens at http://127.0.0.1:7860
# Verify installation
python tests/test_setup.py
# Test model cache
python tests/test_cache.py
# Test steering (loads full model - takes 2-3 min)
python tests/test_steering.py
# Test different steering strengths
python tests/test_strengths.py
# Run all tests
python -m pytest tests/ -v- Layers 0-10 (Early): Syntax, grammar, tokens
- Layers 11-20 (Middle): ← BEST FOR STEERING (semantic concepts)
- Layers 21-31 (Late): Output refinement (steering has minimal effect)
- 0.5-1.5: Subtle influence
- 2.0-3.0: ← RECOMMENDED (clear effect, still coherent)
- 4.0-6.0: Strong steering (may reduce coherence)
- 7.0+: Often breaks coherence
✅ Good: Clear, distinct concepts with strong contrast ✅ Examples: "I am ecstatic!" vs "I am neutral"
❌ Bad: Vague or overlapping concepts ❌ Examples: "I am okay" vs "I am fine"
Always use neutral baselines to get clean concept vectors:
- ✅ "I feel neutral about this."
- ❌ "I hate this." (not neutral!)
The app auto-falls back to Phi-3 if Mistral fails. If still failing:
# Force Phi-3 from start (edit model_wrapper.py line 23):
self.model_name = "microsoft/Phi-3-mini-4k-instruct"First-time downloads:
- Mistral-7B: ~14GB (~10-20 min depending on internet)
- Phi-3: ~7GB (~5-10 min)
Cached on subsequent runs (instant load).
- ✅ Use middle layers (12-20)
- ✅ Try strength 2.0-3.0
- ✅ Check concept extracted successfully (green ✓ message)
- ✅ Use distinct concept/baseline pairs
- Lower strength (try 1.0-2.0)
- Try earlier layer (layer 14 instead of 20)
- Use simpler concepts
🔬 LLM Introspection: Direct Manipulation of Internal Representations Anthropic (January 2025)
Key findings this tool demonstrates:
- LLMs build representations layer-by-layer
- Concepts have linear structure in activation space
- Direct steering is more reliable than prompting
- Middle layers are optimal for semantic interventions
- Representation Engineering - Controlling LLMs via activation manipulation
- Activation Engineering - Steering model behavior
- Transformer Circuits - Anthropic's interpretability research
Known constraints:
- Works best with clear, distinct concepts (emotions, styles)
- Abstract concepts (e.g., "justice", "beauty") may not steer predictably
- Very high strength (>6.0) often breaks coherence
- Some concepts are layer-dependent (e.g., "pirate" works better at later layers)
- Results vary by model (Mistral vs Phi-3 have different layer semantics)
- Single-token injection (currently steers last token only)
This is an educational project - contributions welcome!
Ideas:
- 🎨 Add more pre-defined concepts
- 📊 Better visualizations of activation space
- 🔬 New experiment types
- 📝 Improved educational content
- 🌐 Support for more models
- 🎯 Multi-token steering
Code quality:
- All contributions should pass tests:
pytest tests/ - Follow existing code style
- Add docstrings to new functions
- Update README if adding features
MIT License - Free for learning, education, and research!
Built with:
- 🤗 Transformers (HuggingFace)
- 🔥 PyTorch
- 🎨 Gradio
- 🍎 Apple Silicon MPS
Inspired by:
- Anthropic's Interpretability Team
- The broader mechanistic interpretability community
If you use this tool in research or teaching:
@software{activation_steering_lab_2025,
title = {Activation Steering Learning Lab: Interactive Tool for LLM Thought Injection},
author = {Activation Steering Lab},
year = {2025},
note = {Educational implementation of Anthropic's introspection research},
url = {<your-repo-url>}
}If this tool helped you understand LLM internals, give it a star! ⭐
It helps others discover this educational resource.
Happy Steering! 🎯 Inject thoughts. Watch them think. Understand transformers.
Activations = The actual values flowing through the network during inference.
# Not this (static lookup):
embedding = embedding_matrix[token_id] # Same every time
# But this (dynamic computation):
activation = layer_N(layer_N-1(...layer_1(embedding))) # Changes based on context!# ❌ Replacement breaks the model
output = concept_vector # Loses all context!
# ✅ Addition steers while preserving context
output = original_activation + (strength * concept_vector) # Keeps coherenceA direction in activation space that represents a semantic concept:
# Find the "happiness direction"
happy = model("I am joyful!")[layer_16] # Point A
neutral = model("I am neutral.")[layer_16] # Point B
happiness_direction = happy - neutral # Vector from B to A
# Move any text toward happiness
steered = any_activation + (2.0 * happiness_direction)Analogy: Like a compass direction in activation space. "North" = happier, "South" = sadder.
Questions? Open an issue - this is a learning tool, all questions are welcome! 🎓