New AI Discovery Changes Everything We Know About ChatGPTS Brain

A new AI research paper has revealed some surprising geometric structures in the LLM Learned concepts. Max Tegmark has done a recent paper that reveals the surprising structure of AI brains. He talks about how our new AI paper reveals a surprising geometric structure in the LLM Learned Concepts. He talks about the fact that they form brain like lobes. They form semantic crystals, much more precise than it seems. And the concept cloud is more fractal than round. So if you aren't familiar with the fact that large language models like GPT have been something of a black box, we know that they work. But understanding exactly how they work has been challenging. It's quite like having a car that runs perfectly but not being able to see under the hood. Now, recently, scientists developed tools called sparse autoencoders. They essentially just act like X ray machines for AI, helping us peek inside and see how these models organize information. Think of it like finally getting to look under that car's hood and discovering an unexpectedly organized engine. Now, the researchers found three distinct levels of organization, each more complex than the last. And this layered organization wasn't programmed in. It emerged naturally as the AI learned, which is what is particularly fascinating. So first we had the level one, which is the atomic structure. So here's what they found. The researchers discovered that at its most basic level, the AI organizes concepts in geometric patterns. Imagine a giant 3D connect the dots puzzle where related concepts are linked together in specific shapes. The classic example that they used is how the AI understands the relationship between words. If you plot the concepts man, woman, king, Queen in the AI's mental space, they form a perfect parallelogram, which is essentially just a shape, like a stretched rectangle. And the distance from man to woman is the same as the distance from king to queen. And this basically shows that the AI has learned that adding femininity works in the same way, whether you're talking about royalty or regular people. So now here's where we do see something rather fascinating. You can see that I've managed to create a short artifact with Claude that helps you visualize exactly what I'm talking about. We can see that the relationship between man and woman parallels the same relationship between king and queen. And it's supposed to be a parallelogram, but Claude actually struggles with this quite a lot. Now, this is also the same for capital cities, which you can see right here. You can see that Claude countries and their capital cities form similar relationships in the AI's understanding. And in the past tense, you can see that these verb tenses show parallel relationships, though not as Always as regular as other examples. This was something that was discussed in the paper, but I find this to be really cool. Now, initially, these patterns were hard to see because they were obscured by irrelevant information. It's like trying to see a constellation while there's light pollution. You have to filter out the distracting lights first. Now, the researchers found that things like word length were creating noise that had to be filtered out to see the true patterns. And then this is where we get onto level two, which is the brain structure. So the most exciting finding was that the AI's knowledge is organized into distinct regions, or lobes, similar to how a human brain has different areas for different functions. This organization wasn't programmed, it emerged naturally. Now, there's three main lobes. There's the code math lobe. This is where it specializes in programming and mathematical concepts. And this activates strongly when the AI deals with a coding task or mathematical problem. And this shows clear boundaries from other types of knowledge. Then we have the general language lobe. This handles regular English text processing and deals with the general knowledge and comprehensive understanding.

And the general language lobe processes longer form content like articles and papers. Now, of course, we have the dialog lobe, which specializes in conversational text and short messages. This is particularly active during chat interactions. And this handles different communication styles than the general language lobe. Now, the researchers actually proved that these lobes weren't random by showing that features that often activate together are physically close together in the AI's mental space, which is actually remarkably similar to how neurons that fire together in a brain tend to be physically close to each other. Now then we had the level three of the brain, which is the galaxy structure. So this is where, looking at the overall entire system, this is where the researchers found that the AI's knowledge is organized in a specific way that follows mathematical patterns. Now, this isn't random, it's highly structured, particularly in the middle layers of the AI. Now, some of the key findings from this is that the organization is most pronounced in the middle layers of the AI, where the representations become highly abstract and consolidated, reflecting the most significant transformations of the input data. This layer acts like a kind of information bottleneck, where only the most essential features are retained for further processing, enhancing the model's ability to generalize. Now, information in this AI brain also appears to be compressed efficiently, quite like how our brains compress sensory information. Just as the human brain reduces complex sensory inputs into manageable chunks for efficient processing, the middle layers of AI condense vast amounts of data into simpler, high level representations. And this efficient compression helps the model to Focus on key features while discarding irrelevant details, leading to more robust performance across tasks. Now, the structure follows specific mathematical patterns, which are essentially power laws that suggest optimal organization. And these power laws indicate that the distribution of of feature importance is not random, but follows an organized predictable pattern. And the largest principal components dominate the representation, while the smaller components taper off, creating a natural hierarchy of information. This hierarchical structure not only makes AI more efficient, but also aligns with how biological systems, including human brains, tend to prioritize and process information. Now, all of this is pretty cool, but why does this matter? Why is this even significant to AI research? Well, we can now see how AI systems organize information. And this has far reaching implications for both research and practical applications. This helps explain why they're so effective at various tasks. By understanding the internal organization of these systems, we can see how models are able to generalize, identify patterns and adapt to a wide range of different challenges, from language translation to problem solving. This detailed organization is what gives AI its versatility and efficiency. Now this also gives us insights into potential improvements. If we actually know how these systems organize concepts, this allows us to make targeted enhancements. For example, we could refine the way features are learned in in order to improve the clarity of specific lobes, or even develop new training methods that enhance certain capabilities. Now this insight can actually also help in reducing biases, optimizing computational efficiency and making AI models more interpretable, which is critical for deploying AI in sensitive areas like healthcare and finance. Now it's actually pretty crazy because it parallels with human cognition. Now this, this suggests that there might be a fundamental principle about how intelligence organizes information. The fact that AI and biological brains independently arrived at similar ways of structuring knowledge implies that there could be universal rules governing efficient information processing. This has the potential to influence how we design future AI systems, making them more aligned with with natural intelligence. And of course, this could help us better understand both AI and human cognition. By studying these structures in AI, we can draw parallels to the human brain and potentially learn more about how our own minds work. Understanding why AI develops these lobes and what functions they serve could lead to better models of human cognition, offering insights into how we think and solve problems. And this knowledge might help in addressing cognitive impairments, designing advanced learning tools, or developing AI that collaborates even more effectively with humans. Now, there are some limitations and caveats to any AI research. This doesn't mean that AIs are like human brains. The similarities are organizational, not biological. While the AI's structures resemble those of a brain, they are fundamentally different. The AI is built using layers of mathematical functions and weights. Unlike biological neurons that are compressed of cells and synaptic connections, the AI isn't conscious or thinking like a human. It processes inputs and produces outputs based on learned patterns. But there's no awareness or subjective experience involved. Its responses are derived purely from the data it has been trained on, without any understandings or feelings. Now, these are mathematical patterns, like I said before, not biological ones. The structures that emerge in AI are the result of optimization processes that find the most efficient way to represent information mathematically. Unlike the biological processes that evolved over millions of years, these patterns are engineered to maximize efficiency in computation and task solving, not to replicate the functioning of a human brain. Now, of course, this is just the beginning of understanding these structures. More research is needed to fully understand the implications. We're only really right now starting to understand how these structures form and what they mean for the overall functioning of AI systems. There are many open questions, such as how these structures change as models grow larger, and whether they can be influenced actively during training to improve performance or interpretability. Now, the field is rapidly evolving. We've got many different advances happening in AI and new techniques and discoveries emerging all the time. And the study of these brain like structures is likely to lead to even more breakthroughs, not only in building better AI, but also understanding fundamental aspects of intelligence itself. And as more researchers explore these concepts, we might find parallels that extend beyond AI, potentially offering new insights into cognitive science, neuroscience and even the philosophy of the mind.