Transcription
[Music] Something huge just dropped in the AI world, and it did not come from OpenAI or Google. A tiny startup in Singapore just unveiled a new AI agent called HRM, and it might be the biggest breakthrough in reasoning we have seen in years. It is not bigger. It is not trained on more data. Instead, it is built different. And somehow, this brain-inspired model is outsmarting models over four times its size.
You see, models like ChatGPT break problems down step by step. But one small error can throw everything off. HRM takes a completely different route. It thinks in loops like the human brain, and the results are shocking. Now, before you think this is just another "tiny model beats GPT" headline, this one actually holds up. HRM is not a scaled-down large language model, not a stripped-back transformer, and definitely not just a slim version of ChatGPT. It is something entirely different. It mimics the brain, not in the generic, "oh, it is a neural network" way. This one literally borrows the brain's layered decision-making strategy and applies it to artificial intelligence reasoning. And somehow, this architecture lets a model with only 27 million parameters beat models with over 100 million or even billions. Let's put that in context. GPT-1 had 117 million parameters. HRM: 27 million. That's less than 1/4th. But here's the kicker. It outperforms Claude 3.7 and OpenAI's 03 mini high model on reasoning benchmarks.
Most modern artificial intelligence models rely on something called chain-of-thought prompting. Basically, it is like the artificial intelligence is talking itself through a math problem step by step. Sounds clever, and sometimes it works. But the thing is, if the artificial intelligence makes just one mistake in that chain, the whole answer can fall apart. HRM skips that completely. It does not rely on hoping the chain holds. It reasons more like a human by thinking strategically and then executing rapidly.
So, how does it do that? Well, HRM has two parts that talk to each other. There is a high-level planner and a low-level worker. The planner is like the slow, strategic brain. It maps out the big picture, tries to figure out what kind of problem it is facing. The worker is the fast processor. It takes orders and does the actual work quickly and efficiently. Think of it like a chess master planning a strategy and an assistant executing the moves instantly. These two parts loop through each other. The high-level module makes a plan. The low-level module carries it out and sends back results, and then the high-level updates based on what happened. This back and forth continues until the model settles on the answer. This is not just a gimmick. It is built into the architecture, and it gives the model something that most others do not have: a way to internally check and refine its own reasoning mid-process. And the results are pretty insane.
The ARC AGI benchmark, basically an intelligence quotient test for artificial intelligence. HRM scored 40.3%. That is above Claude 3.7's 21.2% and OpenAI's 03 mini high model, which got 34.5%. These are not small differences. This is a tiny model running on one graphics processing unit, beating out some of the biggest names in the game when it comes to raw reasoning.
Let's take it a step further. HRM was tested on Sudoku, hard and extreme levels. It solved 55% of them. You want to guess how Claude or OpenAI's models did? 0%. Not a single one. And then there is the maze challenge, 30x30 grids. HRM found the optimal path in 74.5% of the tests. The others, again, zero. All this, and the model was trained with just a thousand examples per task. No massive internet datasets, no months-long pre-training. Guan Wang, one of the minds behind HRM, said you could train it to pro-level Sudoku in two graphics processing unit hours. Two hours. That is not just efficient. That is ridiculous.
But there is more to this than just performance on benchmarks. The way HRM is built solves some deeper problems that current transformer models face. Let's take a step back. Transformer-based large language models like GPT or Claude work by processing a fixed number of steps for each output token. They always do the same amount of thinking regardless of how hard the question is. They cannot say, "Hey, this is a tough one. Let me think longer." They do not have the ability to go back, rethink, or rewrite their output once it starts coming out. Once they generate a token, they are locked into it. It is like trying to solve a math problem by writing down one number at a time in pen without ever checking if you are on the right track.
HRM breaks away from that limitation. Its two-level architecture lets it adapt the amount of reasoning based on the complexity of the problem. In fact, there is even a version of HRM that uses reinforcement learning to decide on its own how many iterations it needs for each task. So for simple tasks, it might only loop a few times. For harder ones, it loops more. This makes it way closer to actual flexible thinking than anything else we have seen.
And because of how it is structured, it also avoids one of the biggest issues in artificial intelligence training today: deep backpropagation through time. Most models need to remember and synchronize partial derivatives across multiple layers and time steps. It is memory-intensive. It is slow, and it is probably not how biological brains actually work. HRM, on the other hand, uses more local gradient updates, which are easier to compute and way more biologically plausible.
Now, does that matter in practice? Absolutely. Less memory required means you can run more models at once or train faster with fewer resources. It also means this thing scales really well. You could run HRM on a laptop or even embed it into edge devices or robots. And that is exactly what Sapien is doing. They are already testing HRM in healthcare to help diagnose rare diseases and in seasonal climate forecasting, where it reportedly hit accuracy rates of 97%.
The startup's team includes former engineers from DeepMind, Anthropic, DeepSeek, and even Elon Musk's XAI group. These are people who have worked at the cutting edge of artificial intelligence, and they are all betting on HRM's brain-inspired design to eventually push past the limits of what we've come to expect from large language models. And yeah, they are not shy about it. Guan Wang straight up said, "Artificial general intelligence is about giving machines human-level intelligence and eventually beyond. Chain-of-thought prompting is just a shortcut. What we have built is something that can think." That kind of confidence usually comes from people trying to sell you hype, but in this case, the results are backing it up.
You do not have to just take their word for it either. The entire project is open-source. You can check it out yourself right now on GitHub. Train your own version, modify it, see how it works. That level of transparency is pretty rare, especially for something this promising.
Now, sure, it is early days. HRM still has a narrow focus. It is built to reason, not to chat. Do not expect it to write poetry or summarize your emails. But as a proof of concept, it is one of the strongest we have seen. And HRM is not alone. There are other architectural breakthroughs being explored. Sakana is working on continuous thought machines. There are bit-large language models that only use one-bit weights. Even Google is testing out diffusion-based reasoning models. All of these are still in the experimental phase, but they are part of the same trend: rethinking how artificial intelligence should actually work instead of just making it bigger.
What is different here is that HRM is already working. It is already beating models four times its size using just a tiny sliver of the training data with no pre-training at all. It is not just theory anymore. And honestly, unless someone suddenly throws a few billion into building a new foundational model from scratch, the next big artificial intelligence leap probably will not be another scaled-up GPT clone. And it will be something like this: a totally new architecture that brings better reasoning, faster training, and cheaper deployment without needing a warehouse full of graphics processing units.
If HRM's path continues, it could be the first step toward a world where artificial intelligence agents do not live in data centers. They live in your laptop, in your robot, or even in your car. And they will not just be parsing the internet. They will actually be thinking.
So, what do you think? Is this the future of artificial general intelligence? Drop your thoughts in the comments. I am reading all of them. Make sure to subscribe and like if you want more breakthroughs like this. Thanks for watching. Catch you in the next one.