Home > Articles | 🧠🪜⏱️📈 Hierarchical gradients of multiple timescales in the mammalian forebrain

🧠🤖📈 Scientists just developed a new AI modeled on the human brain — it’s outperforming LLMs like ChatGPT at reasoning tasks

🤖 AI Summary

🧠 A new AI model, the hierarchical reasoning model (HRM), has been developed to mimic the human brain’s 🧠 processing.

• 📈 The model 🤖 consists of a high-level module for slow, abstract planning and a low-level module for rapid, detailed computations.
• 💯 HRM’s performance on the ARC-AGI-1 benchmark was 40.3%, outperforming OpenAI’s o3-mini-high (34.5%) and Anthropic’s Claude 3.7 (21.2%).
• 🚀 It also showed near-perfect performance on complex tasks like Sudoku and maze navigation.
• 📉 The model is more efficient, requiring only 27 million parameters and 1,000 training examples compared to the billions of parameters in large language models.

🤔 Evaluation

The ⚖️ HRM presents a compelling alternative to current large language models (LLMs). While LLMs excel at generating text and general knowledge tasks through massive data consumption, HRM demonstrates that a different architectural approach, inspired by cognitive science, can lead to superior performance on specific, logic-based reasoning tasks. This highlights a fundamental difference in how these models “think.” The HRM’s efficiency and reasoning ability suggest that future AI development may not be solely a race for more parameters but could also involve more sophisticated, brain-inspired architectures. 🤔 Further topics to explore include how to scale this hierarchical approach to handle the massive datasets LLMs use, and whether a hybrid model could combine the strengths of both approaches.

📚 Book Recommendations

🧠 Similar

• 🧠🔄 Livewired: The Inside Story of the Ever-Changing Brain by David Eagleman: This book explores the brain’s 🧠 constant state of change and adaptability, which offers a great parallel to the new, adaptable AI models being developed.

• 🧠🧠🧠🧠 A Thousand Brains: A New Theory of Intelligence by Jeff Hawkins: This book proposes that intelligence comes from the brain building a predictive “world model,” a view highly relevant to the new AI model’s hierarchical design.

• 🧠🔄🏆 The Brain That Changes Itself: Stories of Personal Triumph from the Frontiers of Brain Science by Norman Doidge: It provides a fascinating look into the concept of neuroplasticity, showing how the brain can rewire itself. This directly relates to the idea of creating AI models that can adapt and “learn” in new, dynamic ways.

⚖️ Contrasting

• The Master Algorithm by Pedro Domingos: This book compares and contrasts the five major schools of machine learning, giving a broad perspective on different AI approaches and their fundamental differences.

• An Artificial History of Natural Intelligence: Thinking with Machines from Descartes to the Digital Age by David W. Bates: This book offers a philosophical and historical perspective, arguing that our understanding of the human mind is shaped by the machines we create.

• The Emperor’s New Mind by Roger Penrose: A classic in the field, this book argues that human consciousness cannot be replicated by any known computer algorithm, providing a stark contrast to the claims of brain-like AI.

✨ Creatively Related

• The Alignment Problem by Brian Christian: This book tackles the critical question of ensuring AI’s goals align with human values, a vital topic as these models become more capable and integrated into society.

• 🌊🤖🤔 The Coming Wave: Technology, Power, and the 21st Century’s Greatest Dilemma by Mustafa Suleyman: An insider’s look at the immense power and societal risks of AI, this book puts new model developments into a real-world, geopolitical context.

• Grokking Deep Learning by Andrew W. Trask: This is a technical but accessible guide on building neural networks, providing a foundational understanding of the underlying principles behind brain-inspired models. It’s a great “creatively related” choice as it teaches the reader to build the concepts themselves.