• ☆ Yσɠƚԋσʂ ☆OP
      link
      English
      01 year ago

      Absolutely, but we have to be careful not to confuse stuff like chatgpt with AGI. This is an excellent explanation of what tech like chatgpt actually does.

      In my opinion this is the surest path to creating genuine understanding is to use embodiment as seen here. The model needs to be trained to interact with the actual world so that it creates an internal representation of how the world it’s interacting with behaves through reinforcement.

      This will create correct expectations of how the physical world works, and how to interact with it. This understanding creates a shared context we can use to communicate with the AI. Next, we can start teaching it language based on this understanding the way we would teach a child.

      At that point if I tell a robot to go get a cup from the table, it will have an expectation of what a table is, what a cup is that’s similar to my own. It will know what to expect when it touches the cup, when it tries to pick it up, etc.

      I think we’re still a little ways off from that right now, but people are starting to experiment with these ideas already.

        • ☆ Yσɠƚԋσʂ ☆OP
          link
          English
          01 year ago

          Yeah, I imagine it’s going to be a combination of different approaches in the end. A couple of great books I can recommend are On Intelligence by Jeff Hawkins and Society of Mind Marvin Minsky. I found both to be really insightful regarding what constitutes intelligence and how it may be implemented in practice. Both books are very accessible as they focus on high level concepts as opposed to implementation.

          Personally, I think there are three general approaches towards AGI. First is to try and figure out what it is we mean by human style intelligence, and then try to come up with algorithms that implement that definition. We’re not really close to this right now as we don’t have a clear definition of what we mean by intelligence and consciousness. The two books I mentioned tackle defining these ideas.

          Another approach is to try and reverse engineer what biological brains do. We don’t necessarily have to start with human brains either. For example, bee brains only contains about a million neurons. And bees can do a lot of interesting things like counting, learning, navigation, and communication. Creating an equivalent of a bee brain would be a huge AI breakthrough. This approach can work with the first approach where we can map out brain structures in different animals and then try to figure out the algorithms that can do similar kind of work. On Intelligence has a good discussion about this approach. There’s been some progress in biologically inspired neural networks, but there’s still a long way to go here a well.

          Finally, there’s a possibility of evolving an intelligent agent. This is sort of a brute force solution to the problem where we could create a high fidelity environment that selects for intelligent behaviors. The tricky part here is figuring out what kind of environment would select for human style intelligence. After all, even in the real world life got on just fine for many millions of years without breeding human style intelligence.

          I do suspect that one or combination of these approaches will start producing results that can be called intelligence sooner than people think though.

          Regarding governments, I imagine that there is already a race to build AI systems that can do large scale analytics and predictions. Any government that’s able to create a system that could model the direction of global events better than others would have an incredible advantage. For example, imagine if a government is able to accurately predict when economic crashes or other important events happen before they occur. If a country is able to see what will happen further into the future than their competitors then they have a huge advantage.

          Another huge aspect of this would be creating a large scale planned economy. You could have an AI dynamically adjust inputs and outputs of your domestic economy to ensure that things are produced to meet the needs of the population, to avoid shortages, and so on.