So….. has anyone tried asking an LLMs to design a “human brain” ?

Exploring the Possibility of Designing a “Digital Brain” with Large Language Models

As artificial intelligence continues to advance at a rapid pace, a fascinating question emerges: can large language models (LLMs) be harnessed not just for conversational tasks, but for designing complex systems that emulate human cognition? This inquiry has gained traction within developer and research communities, prompting experiments and discussions about building digital architectures inspired by the human brain.

The Concept: An AI-Driven Digital Brain Blueprint

Recently, enthusiasts have explored the idea of leveraging LLMs—such as ChatGPT, GPT-4, or similar models—to sketch out and potentially implement a digital “brain.” While not aiming for biological accuracy, the proposed architecture attempts to simulate core cognitive processes through modular, interconnected components. These include:

• Perception: Sensory input processing, such as speech recognition and computer vision.
• Working Memory: A dynamic storage with decay mechanisms, akin to human short-term memory.
• Episodic Recall: Retrieving specific past events or experiences.
• Semantic Reasoning: Understanding and manipulating abstract knowledge.
• Planning: Strategizing future actions based on current goals.
• Action Selection: Choosing suitable responses or behaviors.
• Learning & Affect: Adapting through experience, incorporating emotional context.
• Consolidation: Transitioning experiences into long-term memory, similar to sleep-based memory reinforcement.

The Approach: From Architecture Design to Implementation

The idea is to utilize the LLM itself not just for natural language tasks but as a design partner—prompting it to generate both the overarching architecture and the underlying code. The vision involves constructing a system that stitches together specialized modules, such as:

• Automatic Speech Recognition (ASR) and Text-To-Speech (TTS): For sensory interaction.
• FAISS-based Episodic Store: For fast similarity search and episodic memory.
• Planner Module: Incorporating self-consistency checks and tool integration.
• Reinforcement Learning (RL) Agents: Using bandit algorithms for action selection.
• Nightly Replay & Schema Extraction: To emulate sleep-like consolidation processes.

The question that arises within the AI community is whether such a project is feasible today. Has anyone attempted to prompt models like GPT, Qwen, or others to not only conceptualize but also develop and execute code for an emulator of this nature? What successes and challenges have been encountered? Is it possible to move beyond simple prototypes and