A Singapore-based startup is redefining artificial intelligence with brain-inspired architecture that outperforms the biggest models in reasoning tasks.
In the race to build more powerful artificial intelligence, one idea has dominated the landscape: bigger is better. Massive models like GPT-4, Claude 3, and Gemini Ultra boast tens to hundreds of billions of parameters and require enormous compute power to train and operate. But what if this escalating arms race misses the point? What if true intelligence lies not in scale, but in structure? A recent breakthrough out of Singapore-based Sapient Intelligenceis forcing the industry to confront exactly that question.
Their newly released Hierarchical Reasoning Model (HRM) has made waves by outperforming state-of-the-art large language models (LLMs) — not by being bigger, but by being smarter. At just 27 million parameters, HRM is an architectural David in a field of Goliaths. But its performance on reasoning tasks paints a future where smaller, biologically inspired models might not only be more accessible and affordable, but also more capable at complex problem-solving.
The Problem with “Shallow” Intelligence
Most current LLMs, including ChatGPT and Claude, rely heavily on what’s known as Chain-of-Thought (CoT) prompting. This technique encourages models to solve problems by “thinking aloud,” simulating how a person might reason through a multi-step problem. While CoT has improved accuracy in many scenarios, it's inherently brittle. If the model takes one wrong logical step early in the chain, the entire answer can fall apart — a phenomenon not unlike a math student using the wrong formula at the start of a calculation.
Sapient Intelligence argues that this is because LLMs are architecturally shallow. They lack the deep, layered reasoning pathways found in the human brain. This is where HRM comes in — designed to mimic hierarchical processing in human cognition, allowing for recursive planning, multi-layer abstraction, and error correction.
HRM by the Numbers: Tiny, But Terrifyingly Smart
To illustrate the model’s potential, consider its performance on the ARC-AGI benchmark — a sort of IQ test for AI systems created to assess abstract reasoning and general intelligence. HRM scored 40.3%, outperforming:
- Claude 3.7 (21.2%)
- OpenAI’s o3-mini-high (34.5%)
And this is not just on abstract IQ tests. In more grounded tasks like solving Sudoku-Extreme puzzles and navigating complex 30×30 mazes, HRM’s performance is even more impressive:
- Sudoku-Extreme:
- HRM: 55%
- Claude 3.7 & o3-mini-high: 0%
- Maze-solving:
- HRM: 74.5%
- Others: 0%
These aren’t small margins — these are category-crushing differentials. And again, this is from a model that’s orders of magnitude smaller than its competitors.
Trend Toward Efficiency
Sapient’s innovation echoes a growing trend in AI research: model optimization and efficiency. This endeavor is a part of a broader shift challenging the dominance of colossal models. As compute costs rise and concerns over energy consumption intensify, especially in a warming world, the push for leaner, more environmentally and economically sustainable AI is gaining steam. Sapient’s HRM may be taking this one step further — showing that rethinking model architecture itself, rather than just shrinking existing designs, could unlock entirely new performance frontiers.
What If We Have Been Doing It Wrong?
HRM prompts a provocative thought: Have we been training for brute strength when we should have been training for cognitive finesse? There’s an analogy here with the history of flight. For centuries, humans tried to mimic birds by flapping wings. It wasn’t until the Wright brothers introduced fixed wings and propellers. This abandoning of imitation for innovation made us achieve powered flight. Similarly, trying to make LLMs “think” like humans by layering more and more data on shallow architectures may have been our version of flapping wings. Sapient Intelligence may have just introduced the first fixed-wing aircraft of artificial general reasoning.
The Future Is Human-Inspired, Not Human-Sized
The brilliance of HRM lies in its bio-inspired design. While our brains are estimated to contain over 86 billion neurons, the efficiency comes from structure, not volume. Our cognition works via hierarchies of abstraction, distributed memory, and recursive logic — not linear, token-by-token output like current transformers. Sapient’s model emulates this kind of cognitive stack. That opens exciting possibilities not only for solving abstract puzzles but for eventually developing more explainable, generalizable, and ethical AI systems.
With smaller models, we also gain a host of practical advantages:
- Lower inference costs → enabling edge computing and mobile deployment
- Faster training cycles → more rapid innovation
- Democratization of AI → more organizations and researchers can experiment and deploy advanced systems
This shift could significantly alter the AI talent and capital landscape, reducing the barriers to entry and decentralizing AI development from a few elite labs to a truly global community.
A Turning Point?
It’s far too early to declare HRM or Sapient Intelligence the definitive future of AI — breakthroughs need rigorous validation and real-world application. But the implications are tantalizing. The model’s success raises a crucial question for the industry and policymakers alike: Should we be chasing bigger models, or smarter designs?
Originally Published on LinkedIn.
