In a world where bigger is supposedly better, AI research has taken an unexpected turn. Bycloud’s YouTube video, titled “How did a 27M Model even beat ChatGPT?”, published on December 4th, 2025, introduces us to a groundbreaking notion in artificial intelligence, or AI, that challenges the conventional wisdom of scaling up models for better outcomes. This tiny yet mighty model, featuring only 27 million parameters, a startling contrast to the behemoth models of today, seems to be rivaling industry giants on challenging benchmarks. But can a minuscule model rival the colossal LLMs on the ARC AGI benchmark with merely 27 million parameters? What’s fascinating in this case is the hierarchical reasoning model (HRM) that banks on depth over breadth; it offers a new way of solving complex logical tasks through iterative thought processes.
The video by bycloud delves into the inner workings of HRM, explaining how this model, starting with only 27 million parameters, unexpectedly became comparable or even outperformed state-of-the-art models like OpenAI’s GPT4. Its achievements on the ARC AGI benchmark and its ability to solve logical puzzles like Sudoku are quite noteworthy. This section effectively provides compelling evidence of how HRM is transforming the world of AI as it encourages the use of small, recursive models over expansive ones. This reflects the creators’ effective use of evidence to make a compelling case for reconsidering our approach to scaling AI.
However, the explanation does raise some concerns. While impressive, such a small model’s ability to replicate a larger one rests on ample assumptions and heuristic designs, as noted by some researchers. Critiques focus on the model’s supposed biological analogy to a mouse brain, a comparison that lacks empirical backing. The abstraction of the hierarchical element, viewed here as heuristic-based, lacks thorough evidence that might better anchor its credibility. Despite this, Alexia further extends this research to create a tiny recursive model (TRM) that refines and evolves the concept to achieve even greater outputs, confirming that smaller models also come with certain advantages over larger ones due to the reduced risk of overfitting and increased performance.
In summary, this video bycloud challenges the AI paradigm by showcasing smaller models that break conventional performance expectations, applying a new kind of architectural logic that emphasizes iterative reasoning. It invites us to rethink model design, pushing towards an understanding that maybe it’s not the size but the processing style that counts. It’s a provocative exploration of AI’s future terrain, asking whether smaller, smarter could eventually become the new norm. What’s more intriguing is considering how this trajectory could influence the next wave of AI advancements, encouraging further experimentation and evidence-backed innovation in the field. Handy references to related research documents further augment the argument, providing depth to the discussion about smaller, more effective models.