*GPT5 was released just a few days ago but something else was also released about the same time that I think may turn out to have an even larger impact, it's called Hierarchical Reasoning Model (HRM). It's an open source AI made by a small obscure company in Singapore named "Sapient Intelligence". It uses a completely different type of architecture than chain of thought and transformers, the one that all the other modern AIs use. This technique seems to be much better than the chain of thought technique in tasks that involve non-linguistic reasoning, like spatial puzzles, symbolic patterns, mazes and other abstract visual tasks.*
*Hierarchical Reasoning Model* <https://arxiv.org/pdf/2506.21734> *HRM is completely open source so you could download it on GitHub at https://github.com/sapientinc/HRM <https://github.com/sapientinc/HRM> and it's so small you can run it locally and off-line if you have a high end consumer PC.* *The first thing that's unusual about HRM is that it's really fast and very small, it only has 27 million parameters, GPT3 had 175 billion, OpenAI has never said how many GPT4 or GPT5 have but it's probably in the trillions. In spite of its tiny size, on tasks that involve logic not knowledge it can outperform state of the art AI's that are 100 to 1000 times larger, and it needed less than 1000 trading examples to do so, while the huge musclebound AI's need millions. * *For example, in spite of their large amount of training, in a test that involves solving sudoku puzzles that were picked because they were unusually difficult, Claude 3.7, OpenAI's o3 mini-high, and Deepseek R1, all scored a big fat ZERO, they couldn't solve a single one, but HRM could solve 55% of them. And NONE of those huge AIs could find the shortest path between two points in a complex 30 x 30 maze, but HPM could find it 74.5% of the time. On the ARC AGI-1 test **OpenAI's o3 mini-high beat the others with the score of 34.5 but HPM got 40.3. Nobody did very well on the far more difficult **ARC AGI-2 test, **HPM only scored a 5 but even here it beat it's nearest competitor **o3 mini-high, it only got a 3. **Deepseek R1 got 1.3 and **Claude 3.7 got a 0.9 * *All the large conventionally AIs use chain of thought reasoning using explicit linguistic steps, and that makes it very good at language but it also requires a vast amount of training data. But having a running inner dialogue composed of words is not the only way humans can think, we also use a series of non-verbal, abstract computations that cannot be expressed in any human language. * *The paper says, “the brain sustains lengthy, coherent chains of reasoning with remarkable efficiency in a latent space, without constant translation back to language. HRM can perform a sequence of nested computations in a biologically credible way. " * *HRM has two main parts:* *A High-level module (H) Thinks abstractly and slowly, and plans out a strategy. * *Low-Level Module (L): This is the fast-working intern, crunching numbers or details quickly, guided by the high-level plan. * *H sends out an order to L to look for a solution at a particular place and awaits further orders, H reads what L has found and on that basis sends modified orders back to L, this back-and-forth iteration continues until a small part of H called the "Q-head" decides that this iteration is close enough to the truth and stops things. Conventional neural network AIs remember every step but that requires a huge and ever-growing amount of memory, but HRM only remembers a one-step gradient approximation, it calculates that gradient only at the end. As a result memory use is constant, even for deep reasoning.* *All the large conventional AIs that use Transformers and chain of thought reasoning use explicit linguistic steps, and that makes it very good at language but it also requires a vast amount of training data and electrical power. HRM doesn't think with words or tokens, its inner states are continuous vectors and reasoning is not an inner dialogue but more like neurons firing in patterns.* * Disadvantages of HRM.* *AIs that use transformers and chain-of-thought think in words, therefore they are very good at language, so when you ask them a question they can not only provide an answer they can also eloquently explain how they obtained it. You could train a HRM to input and output text but, because it doesn't think in words, its intermediate representations won't align perfectly with human sentences, and telling others what our intermediate states were is how we explain to others how we arrived at our conclusion. * *So the HRM AI might produce answers that are 100% correct but won't be able to justify them in a way that most humans would find coherent. It might be like the famous Indian mathematician Ramanujan who almost always seemed to know if a mathematical statement was true or not but could not provide a proof or even give a good informal explanation of how he knew what he knew, in some cases it took decades for other mathematicians to provide proofs, and it was only then that they realize that Ramanujan was right. * *Even with conventional AIs it's often very hard to figure out exactly why AIs do what they do but HRM makes things even worse. In a normal neural net you just have one black box to interpret. In HRM You have a black box (high-level) giving tasks to another black box (low-level). And the low-level is trained in a way that forgets all but the final state, and that is only an approximation. So if the AI develops a mental pathology, like suicidal depression or homicidal rage or an inability to metaphorically get out of bed and actually do something, an AI psychiatrist (a human who has an AI as a patient) will have great difficulty in determining if the problem is in the content of the high-level plan, or in the execution at the low level, or in the interpretation and interaction between the two. The reasoning chain is nested and partially invisible.* *John K Clark See what's on my new list at Extropolis <https://groups.google.com/g/extropolis>* z3h -- You received this message because you are subscribed to the Google Groups "Everything List" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion visit https://groups.google.com/d/msgid/everything-list/CAJPayv12Nek%3DrDCtmRJYpXzhzEDk4bu-tDFDZtwe4tq1PC%2BN1w%40mail.gmail.com.
