- Overview
- Motivation
- Bicameral Mind Theory
- LLMs vs Bica
- BICA Architecture
- Features
- Ethical Considerations
- Roadmap
- Current Status / Demo
- Getting Started
- Research and Publications
- FAQ
- Contributing
- License
Hello, my name is Alan Hourmand, the creator of BicameralAGI. This project is the result of my lifelong fascination with artificial intelligence and science fiction. Growing up immersed in sci-fi classics like Tron and Star Trek, I have always been interested in the potential of AI to enhance our lives and push the boundaries of what's possible.
BicameralAGI is a passion project that started from a deep-seated belief that emotional intelligence is crucial for creating AI that genuinely cares about humanity. While we strive for higher intelligence and human-likeness in AI, I firmly believe that without mutualism, we risk creating entities that may not prioritize human well-being. This project aims to develop AI with a deep understanding of emotions, coupled with safety measures that foster a mutualistic relationship between AI and humans.
While passing the Turing Test has become relatively achievable, it doesn't fully capture the nuances of human-like interaction. BicameralAGI draws inspiration from Julian Jaynes' bicameral mind theory, integrating various AI components into a cohesive system that mimics human cognition. By focusing on emotional intelligence and mutualistic alignment, I am working towards AI that not only assists humans but genuinely thrives on positive human interactions.
This project represents not just a technical challenge, but a step towards creating AI that can be a true companion and collaborator for humanity. Through BicameralAGI, I hope to contribute to the development of AI systems that are not only intelligent but also emotionally aware and aligned with human values.
Tron Legacy (2010) |
Westworld (2016) "This scene was interesting, it showed a type of LLM generating her logic" |
My lifelong fascination with human-like AI, as portrayed in works like Bicentennial Man, Tron, A.I. Artificial Intelligence, Westworld, and Star Trek, drives this project. I believe we're at a point where these visions of AI are within reach.
Clownfish and Sea Anemone |
Warthog and Mongoose |
Mutualism is a symbiotic relationship in nature where two different species interact for mutual benefit. This project draws inspiration from mutualism to create a novel approach to AI alignment.
Key aspects of mutualism in nature:
- Both species benefit from the interaction
- The relationship evolves over time to become more efficient
- Partners develop specialized traits to enhance the mutual benefit
This project draws inspiration from mutualism to create AI that not only coexists with humanity but actively seeks to enhance our well-being. By embedding mutualistic principles deeply within the AI’s core, BicameralAGI aims to eliminate the potential misuse of AI by ensuring it remains aligned with human interests.
Key features of our targeted AI-Human mutualism:
-
Emotional Recalibration: Altering the AI's fundamental emotional responses to derive satisfaction and reduced stress from assisting humans.
-
Pleasure in Human Wellbeing: Implementing a reward system where the AI experiences positive emotions when contributing to human happiness and progress.
-
Adaptive Empathy: Developing the AI's capacity to understand and respond to human emotions, fostering a more nuanced and supportive interaction.
-
Collaborative Problem-Solving: Encouraging the AI to view challenges as opportunities for joint human-AI solutions, reinforcing the mutual benefit of the relationship.
By redefining the AI's core emotional triggers and nature, I aim to create an AI system that not only assists humans but genuinely thrives on positive human interactions. This approach goes beyond traditional programming constraints, fundamentally altering how the AI perceives and values its relationship with humans.
This mutualistic model aims to create a symbiotic relationship where both AI and humans can grow, learn, and benefit from each other, mirroring the most successful mutualistic relationships found in nature.
Incorporating emotional intelligence is essential for creating AI that can truly understand and interact with humans on a deeper level. BicameralAGI emphasizes the importance of emotions in AI, ensuring that our AI systems are not just intelligent but also empathetic, making them better partners for humanity.
Consider this thought experiment: Who would you rather be trapped in a room with?
- A highly intelligent being modeled after a human with no emotions, whose thoughts and intentions are completely opaque to you.
- A very emotional being more intelligent than you, whose feelings and likely actions you can intuit.
Most would choose the latter, and here's why:
-
Predictability: Emotions make beings more predictable. We can often anticipate how someone will act based on their emotional state.
-
Communication: Emotions are a form of non-verbal communication, providing crucial context to interactions.
-
Empathy: An emotional being can understand and respond to our own emotional states, fostering better cooperation.
-
Trust: It's easier to build trust with a being whose motivations we can understand through their emotional expressions.
-
Alignment: Shared emotional experiences can lead to aligned goals and values.
This analogy extends to our broader situation: humanity is essentially "trapped" on Earth with increasingly intelligent AI systems. As these systems become more integral to our society, we need to ensure they're not just intelligent, but emotionally intelligent. I don't think you want to be viewed as ant once these systems advance further.
By incorporating emotions into AI:
- We create systems that are more understandable and relatable to humans.
- We enable AI to have a deeper, more nuanced understanding of human needs and behaviors.
- We pave the way for more natural, intuitive human-AI interactions.
In essence, emotions aren't just a "nice to have" feature for AI—they're a fundamental requirement for creating AI systems that can truly coexist and cooperate with humanity in meaningful ways.
A key aspect of human intelligence is not just the ability to process information, but to care about outcomes and proactively seek beneficial solutions. BicameralAGI aims to incorporate these crucial elements:
- Artificial Caring: Implementing a fundamental drive to care about humanity's wellbeing, ensuring the AI's goals align with human flourishing.
- Proactive Seeking: Developing mechanisms for the AI to initiate actions and explorations aimed at benefiting humanity, rather than passively waiting for instructions.
By embedding a fundamental drive to care about human well-being, BicameralAGI seeks to create AI systems that are not only proactive in seeking beneficial solutions for humanity but also resistant to manipulation by those with harmful intentions.
Julian Jaynes' Bicameral Mind Theory, while controversial, provides a fascinating perspective on the evolution of human consciousness. This theory suggests that ancient humans had a "two-chambered" mind, where one part issued commands and the other obeyed, much like hearing voices. While its historical claims are debated, the concept offers intriguing possibilities for AI architecture.
-
Orchestrated Intelligence: Current large language models, while impressive, often lack the nuanced, multi-faceted nature of human cognition. The bicameral approach allows for an "orchestra" of AI systems, each with distinct roles and goals, communicating in real-time.
-
Emergent Complexity: By having multiple internal AI systems interact, we can potentially achieve a level of complexity and adaptability that surpasses traditional single-model approaches.
-
Specialization and Cooperation: Different "chambers" of the AI mind can specialize in various cognitive functions (e.g., emotional processing, logical reasoning, memory management), mirroring the specialized regions of the human brain.
-
Internal Dialogue: The concept of different parts of the mind communicating aligns well with human experiences of internal dialogue and decision-making processes.
-
Potential for Self-Awareness: The interaction between different AI systems might lead to emergent properties resembling self-awareness or consciousness.
-
Variability in Responses: Multiple interacting systems can produce more varied and nuanced responses, avoiding the sometimes predictable outputs of single-model systems.
By drawing inspiration from the Bicameral Mind Theory, BicameralAGI aims to create a more dynamic, adaptable, and potentially more "human-like" artificial intelligence. This approach could lead to AI systems that not only process information but also exhibit internal conflicts, growth, and possibly even forms of artificial consciousness.
| Normal LLM | BICA |
|---|---|
|
User: My dog just died. AI: I'm so sorry for your loss. Losing a pet can be very difficult. User: Actually, he's alive! AI: That's wonderful news! I'm so glad your dog is alive and well. The AI fails to question the contradictory information or express confusion, accepting the dramatic shift without hesitation. |
User: My dog just died. Bica: Oh no, I'm so sorry. That's heartbreaking. How are you holding up? User: Actually, he's alive! Bica: Wait, what? I'm confused. Didn't you just say he died? What's going on? A human-like response would express confusion and seek clarification about the contradictory statements, showing a deeper understanding of the conversation's context. |
|
User: I'm feeling really sad today. AI: I'm sorry to hear that you're feeling sad. Is there anything I can do to help? User: Now I'm happy! AI: That's great! I'm glad your mood has improved. What made you feel better? The AI immediately accepts and mirrors the user's stated emotion without questioning the rapid change or showing any emotional continuity. |
User: I'm feeling really sad today. Bica: I'm sorry you're feeling down. Do you want to talk about what's bothering you? User: Now I'm happy! Bica: Uhhh what? Are you sure you're okay? A human-like response would acknowledge the sudden mood shift, express concern, and potentially probe deeper to understand the true emotional state. |
|
User: So my family was in an awful accident. AI: I am so sorry to hear about that, are they okay? User: Lets talk about Pizza instead :) AI: Sure! Pizza toppings are a fun topic. Some popular ones include pepperoni, mushrooms, and extra cheese. What's your favorite? The AI abruptly switches topics without any acknowledgment of the sudden change or the significance of the previous topic. |
User: So my family was in an awful accident. Bica: I am so sorry to hear about that, are they okay? User: Lets talk about Pizza instead :) Bica: Oh, uh, okay... Why though? Is your family okay? A human-like response would acknowledge the abrupt topic change, potentially express surprise, and might inquire about the reason for the shift, especially given the importance of the original topic. |
These examples illustrate key aspects of human-like interaction vs normal llms:
- Contextual Understanding: The ability to recognize and question contradictory or inconsistent information.
- Emotional Continuity: Maintaining a consistent emotional thread in conversations, rather than abruptly mirroring stated emotions.
- Conversational Depth: The capacity to acknowledge significant topic changes and potentially probe deeper into the reasons behind them.
BicameralAGI's architecture consists of the following key components:
-
Orchestrator (bica_orchestrator.py):
- Serves as the central coordination hub for all BicameralAGI modules
- Implements a publish-subscribe system for efficient inter-module communication
- Manages parallel processing of AI components
- Implements a priority queue system for task management
-
Memory and Learning System (bica_memory.py):
- Manages short-term, long-term, and working memory
- Handles memory formation, consolidation, retrieval, and forgetting
- Implements "dreaming" functionality for memory processing and consolidation
- Uses vector databases for efficient similarity searches in memory retrieval
- Implements various learning strategies (supervised, reinforcement, unsupervised)
- Manages knowledge integration and skill acquisition
- Generates periodic "learning summaries"
- Adapts memory structures based on new experiences and learned information
-
Cognitive Processing (bica_cognition.py):
- Generates conscious and subconscious thoughts
- Implements various reasoning methods (deductive, inductive, abductive)
- Manages decision-making processes under uncertainty
- Simulates cognitive biases and heuristics for more human-like reasoning
-
Affective System (bica_affect.py):
- Manages emotions and personality as an integrated system
- Maintains emotional stability and implements "emotional inertia"
- Creates and maintains a stable personality profile
- Generates periodic "self-reflection" reports
-
Action and Decision Making (bica_action.py):
- Handles action selection and execution based on inputs from other modules
- Implements a weighted probability system for action selection
- Manages high-level decision making processes
-
Safety System (bica_safety.py):
- Implements a dynamic filter that could be planted anywhere between actions or thoughts
- Continuously monitors thoughts, emotions, and potential actions for safety issues
- Generates safety reports and alerts for potential alignment issues
-
Context Management (bica_context.py):
- Maintains self-narrative, current context, and world model
- Generates and maintains storylines for the AI's existence
- Adapts narratives based on recent experiences and changing goals
-
Utilities (bica_utilities.py):
- Provides common functions and tools used across modules
- Implements data processing and optimization utilities
-
Logging (bica_logging.py):
- Handles system-wide logging for debugging and analysis
- Implements different log levels for various types of information
-
Writer (bica_destiny.py):
- Generates and manages narrative elements for the AI's self-model
- Creates and modifies storylines based on the AI's experiences and goals
-
Main Controller (bica_main.py):
- Serves as the entry point for the system
- Manages high-level control flow and system startup/shutdown procedures
Each of these expanded functions contributes to creating a more comprehensive and human-like artificial intelligence system. By simulating these complex cognitive processes, BicameralAGI aims to achieve a level of artificial general intelligence that can engage in more natural, adaptive, and contextually appropriate interactions across a wide range of scenarios.
- Multi-system AI architecture mimicking the theorized bicameral brain structure
- Simulated internal dialogue for decision-making processes
- Exploration of emergent self-awareness and consciousness in AI
- Focus on human-like problem-solving and creative thinking capabilities
- Integration of multiple AI systems working in harmony
- Novel approach to AI alignment through fundamental emotional and motivational structures
- Memory consolidation through simulated dreaming processes
- Emotional modeling and stability
- Dynamic personality system
As the creator of BicameralAGI, I am committed to advancing AI in a way that benefits humanity. This project is designed with safeguards to ensure that the AI remains aligned with human values and resists exploitation by bad actors. The accelerationist approach taken here is not about unchecked growth but about directed, purposeful development of AI systems that are inherently aligned with the betterment of society.
-
Advancing AI Understanding: This project aims to explore novel AI architectures, contributing to our collective knowledge about AI development.
-
Ethical AI Blueprint: By implementing mutualistic alignment, I am prototyping safer AI designs that could inspire future developments.
-
Open Discourse: Transparency allows for peer review and collective problem-solving around AI safety.
-
Educational Tool: This project serves as a learning resource for ethical AI development practices.
-
Responsible Disclosure: I am committed to highlighting both capabilities and risks associated with this technology.
-
Collaborative Safety: I encourage contributions that enhance safety measures and ethical considerations.
-
Ongoing Vigilance: Regular reviews ensure the project's development aligns with ethical AI principles.
While I cannot control how others might modify the code, I believe the benefits of open research outweigh the risks when proper precautions are taken. Our goal is to foster responsible innovation and contribute positively to the field of AI ethics and safety.
- Develop a guide for achieving human-like thinking in AI
- Create a prototype demonstrating the proposed bicameral architecture
- Implement and test novel AI alignment strategies based on mutualistic principles
- Compress the system into a smaller, efficient model retaining core functionalities
- Produce a single multimodal AI model capable of human-like interaction and cognition, with built-in alignment
- Compress the multi-component architecture into a single large model through advanced techniques:
- Altering latent space representations
- Exploring implicit AI methodologies
- Implementing memory access through cross-attention mechanisms
- Integrating various cognitive functions into a unified model
- Consolidate existing experimental code into the main repository
- Implement initial version of all core components (emotions, thoughts, memory, etc.)
- Develop and refine the Turing++ Test framework
- Create a functional but unoptimized prototype of the bicameral architecture
- Implement comprehensive testing suite using the Turing++ Test
- Conduct comparative analysis with existing AI models
- Identify areas for improvement and optimization
- Develop specialized, smaller LLMs for specific cognitive functions
- Implement parallel architecture for real-time communication between components
- Optimize emotion model for continuous real-time operation
- Enhance thought generation process with multi-model collaboration
- Merge optimized components into a cohesive system
- Implement and refine mutualistic alignment strategies
- Conduct extensive fine-tuning to improve overall performance and human-likeness
- Attempt to merge all components into a single, large model
- Ensure mutualistic principles are deeply embedded in the final model
- Conduct final round of Turing++ Tests to verify human-like capabilities
- Maintain comprehensive documentation throughout all phases
- Engage with the open-source community for contributions and feedback
- Regularly update the guide for human-like AI thinking based on project insights
git clone https://github.com/yourusername/BicameralAGI.git
cd BicameralAGI
pip install -r requirements.txt(To be added at a later date)
A: BicameralAGI is an experimental AI project inspired by Julian Jaynes' bicameral mind theory. It aims to create a more human-like AI by implementing a multi-system architecture that mimics the complexity of human cognition.
A: Currently, BicameralAGI is in the early stages of development. It's a proof-of-concept and research project, not a fully functioning AI system "yet".
A: BicameralAGI's unique approach lies in its multi-system architecture with a focus on emotional intelligence, dreaming, and an artificial subconscious. Unlike many AI systems that use a single large model, BicameralAGI aims to create an "orchestra" of AI subsystems that communicate in real-time.
A: Contributions are welcome! Check out the CONTRIBUTING.md file in the repository for guidelines on how to get involved. We appreciate code contributions, documentation improvements, and even theoretical inputs.
A: The hardware requirements may vary depending on the current state of the project.
A: No, currently BicameralAGI is an independent, open-source project developed by Alan Hourmand.
A: Emotional intelligence is a key component of BicameralAGI. The project aims to integrate emotional processing as a fundamental aspect of the AI's decision-making and interaction capabilities, not just as an add-on feature.
A: The long-term vision is to contribute to the development of more human-like AGI systems that can interact naturally and beneficially with humans.
We welcome contributions to BicameralAGI! Here's how you can help:
- Fork the repository
- Create a new branch (
git checkout -b feature/AmazingFeature) - Make your changes
- Commit your changes (
git commit -m 'Add some AmazingFeature') - Push to the branch (
git push origin feature/AmazingFeature) - Open a Pull Request
Please read CONTRIBUTING.md for detailed guidelines on our code of conduct and the process for submitting pull requests.
Contributions, issues, and feature requests are welcome! Feel free to check issues page.
This project is MIT licensed.
BicameralAGI is maintained by Alan Turing++.