Deep Learning appears to be different from symbolic learning. A human can store "If A and A -> B then B" relatively fast and apply this as a rule of symbol manipulation to "X and X -> Y" to generate new knowledge. This kind of learning is far from gradually adjusting neural network weights and thresholds.

Regards,

Joachim

In general I agree with your statement, that on a higher level, human and machine learning are quite similar. However, AGI requires a broader understanding of the world (combining language, memories, history, sensory information, social and cultural understanding), which is currently not feasibly with existing architectures.

I think, this is more of a data problem. There are to few examples of handling-data. For example, if you ask an LLM the following Question: "You put a marble in a cup and put the cup on a table with the bottom of the cup facing upwards, what happens to the marble, if you take the cup and put the cup in the microwave." Most LLM will tell you, that the marble will getting heated up. That is because LLM currently have to less of handling-data. So the understanding of how things work often is wrong. I think this can be fixed either with a special dataset or learning through embodiment. High-level LLMs develop a world-model inside to understand the world and predict an outcome. The ability to generate new knowledge with existing knowledge is technically given through the high dimensional vector-space in the NN. These can be shown by asking new and never seen before questions. So I see no technical limit in this case.

@Sonic2k Dbs it seems ai does not understand the basic principle of gravity. Any suggestions to solve this?

Bao Pham Thank you for your question. Yes, I have 3 possible solutions to this. Two like I mentioned above and a third one. But first let me describe the problem in detail. LLMs are currently like kids, who have only read but not learned by growing up before. But because we have, our text data is already incomplete and compressed. Given the above mentioned example, uncompressed text would be: “You are in a kitchen, You reach to the cupboard and grab a cup with your left hand, you lift up your hand a bit by lifting your left arm up a bit and pull your hand back. You hold the cup in your left hand. You reach to an open glass nearby with your right hand and put your fingers in by lowering your right arm to grab a marble. You grab one marble with your fingers. You feel the temperature. The marble is room temperature. You feel the surface. It is bend and slick. You take your hand out of the glass by lifting your right arm. You move your right arm to place your hand above the cup. You let go the marble by spreading your fingers to put the marble in the cup. The marble falls in the cup. This makes a sound, as the marble reaches the bottom inside of the cup...” and so on. These steps are not needed for us to understand, but for the LLM, they are completely missing. They are not in our text data. It can't know. How do the LLM think it works? So if the LLM reads the question: "What happens to the marble, if you take the cup and put the cup in the microwave?", the LLM will exchange the Table with the Microwave instantly, without any steps. Steps, we always add in mind because of our trained world image. And also no stepping time is seen here by the LLM. So nothing changed with the cup and the marble for the LLMs Mind. So the LLM thinks the whole construct is in the microwave now and says, it would heat up. But the truth is, nothing would happen, even if the marble would still be in the the cup in the microwave. Because no one told the LLM, that Microwave oven only heat things up, if you close the door and switch them on. Because we are skipping this step also in our texts. And you can only switch them on, if they have power. And because no one writes text like such. LLMs can not learn such steps. Without prior knowledge the text is incomplete. But text is all the LLM can learn from. Its world-image (embedded vector-structure of NN weight relations in this case) is incomplete then. That is the problem, why LLMs struggle with physics in their mind. A self-image would also help. At least an image of having hands, able to grabbing and moving something. Having hands is not a given thing for an LLM. So we need at least a more complete world-image and even better, also a simple self-image for better physical understanding. Once the knowledge is there, LLMs will integrate these embedded vector-structure of NN weight relations for physics by connecting it to the whole NN, because LLMs try to predict correct answers, who fit to the training data and this knowledge gives advantage to do so. Also applying it pretty early in the training could give an advantage for better categorizing following training data. Training in stages could be an advantage here.

The 3 possible solutions I can see: 1) Special training data-set: physical steps as step by step details. (I searched for this and couldn't find anything like that) 2) Embodiment: develop this understanding naturally by using a physical body. (probably best, but expensive and time consuming) 3) Video data (or image series) for multimodal LLMs, because here exists a frame by frame explanation in it - for the LLM, that gives best details. For example how a ball accelerates by falling and rolling, the materials colors, firmness and weight, It can see how the ball is held above the ground, because it doesn't just float in the air and the LLM can learn how the hands who hold the ball are releasing it, how it accelerates until it hits the ground, how the grass bends under the weight of the ball, how it bounces back, how it moves and how far it rolls downhill in the grass, and that it takes some time steps to do so. (for learning, one LLM could describe the video frames in detail, while the other LLM could learn from the text data) Best practice would be to combine these 3 solutions, to build up the best understanding of practical physics for the LLM, for its world-image. Thanks again. I hope my answer is useful to you and gives some ideas.

@Sonic2k Dbs thank you very much for your wonderful and detailed suggestions. I am just beginning to learn basic ai but i think your 3rd suggestion is in the right direction. Let's see what the future holds