← Back to Blog

AGI is Hundreds of Years Away

So there has been a lot of buzz about the true meaning of Artificial General Intelligence. There are a lot who claim that LLMs are conscious and can think like humans. There are also others that argue that we are not quite there. The general consensus seems to be that AGI is not far away, it’s within our reach. While I understand where all these arguments are coming from, I don’t agree with the way people have carried the meaning of AGI.

Firstly, let’s talk about LLMs. I think LLMs are very useful, and it is quite astonishing that we were able to engineer a system that is capable of holding memory of such a huge corpus of data from the internet and extract relevant parts for queries in order to produce a response. Pretty remarkable and very impressive. No Doubt.

What I completely disagree with is the framing that LLMs think like us. They are more capable than us. They are far more intelligent than us. This is not true. Every idea that LLMs produce was already there in the internet. Well, if not all, I would argue 90% of all thoughts and ideas that LLMs claim as their own is somewhere in the internet. The remaining 10% are just either extremely generic, mediocre responses that just fit multiple ideas together to force a narrative or just hallucinations. What this means is that LLMs are a great engineering system that can fool us into believing that it's producing ideas tunomously, but what it actually does is reproduce the garbage that it has been trained on. Period.

In other words, LLMs are incapable of producing new knowledge. They can talk about Newton’s laws because it’s already discovered. They can talk about Machine Learning concepts because it’s already there in the internet. Now take a moment. Consider our world 100 years from now. There are going to be significant changes, most of them coming out of great ideas. Can LLMs produce those ideas? No. Another thought experiment. Imagine ChatGPT was discovered in the 12th century. You prompt it “Does the Sun revolve around the earth?” It will quite certainly answer No, because people did not believe in the Heliocentric model back then. It was simply not discovered. This is the exact gap we are incapable of seeing right now because we are blinded by the enormous hype in the internet. If LLMs were truly remarkable, every researcher would have a top conference paper. Every student would have genius ideas to present in class. But that simply does not happen, because LLMs are dumb. They don’t think. They are a sophisticated machine capable of reproducing the garbage they are trained on. They can find soft connections and pinpoint you to the right parts of the internet, all while storing memories in its parameters.

Now that we have established that LLMs are not truly intelligent, let’s address another hot topic. What is AGI, and how far is it? AGI means different things to different people. I personally don’t think we need to worry too much about the word AGI as long as we can make AI systems become good at narrow downstream tasks. If this is the case, AGI would be a process over hundreds to thousands of years. And I think you might have already guessed my prediction for when AGI will be here. I am fully convinced that I won’t see AGI in my lifetime. I am convinced that I will see AI getting better on narrow downstream tasks like math, coding, specific leg surgery, etc. But no way we see a system like humans, capable of everything in my lifetime. When will AGI be here? I think a better question to ask me would be how much time would AGI at the very least take? I would say it might take us around 500 years. Hell no way in my lifetime and people who claim it’s 5-10 years from now, are deluded.

Why? Because, firstly, LLMs are not intelligent. Scaling will hit a limit. You can already sense this with Claude. Claude Code is great for open-ended tasks. Ask it to make a calculator, it will do it. Ask it to make a website it will do it. It’s great when you don’t have a lot of requirements. You just need a website, anything works. But as you keep adding requirements on top of requirements and you are in need of a specific design, it will fail. You need to be a very good programmer. You need that intuition to spot errors, because Claude Code will produce garbage code with a lot of errors. If you are able to use it right, it will certainly speed up your process. It’s a great productivity tool, but no way a replacement. I don’t think software engineers are going anywhere. We will for sure see a reduction in the number of people employed as software engineers because a lot of junior developers will be impacted. In other words, people whose jobs were to write some small functions and objects, we need less of those people. I would argue that software engineers would go down by 30% max. That’s it. And while that’s a significant number, the remaining 70% is still significant as software engineers were in already too much demand and too many people employed before. It might be more like a chemical engineer or mechanical engineer profession where the number of jobs in the US for instance are about 100k. But if you are passionate about building software, there is no reason for you to be afraid. Stay put. Stay on the line.

Secondly and more importantly, real intelligence requires interaction with the real world. Current systems are good with language, but they are very bad at processing visual data. Language is finite, discrete and verifiable. Real world is high dimensional, continuous, noisy and not verifiable. Consider math. You know whether an answer is wrong or right. In other words, math is verifiable. But how can you verify whether a hypothesis LLMs produce about Hawking’s radiation is a sensible one? There is no way to verify this, at least it’s not easy like math. So current systems are good at solving tasks that are discrete, finite and verifiable, but the real world isn’t any and we don’t have any good systems that can do anything meaningful in the real world. Getting a robot to clean a utensil is even something that we haven’t been able to do. Almost every algorithm in this field only works on toy benchmarks like Maze environment, Minecraft, counterstrike, all these games. Because it’s simple, all of these environments give you very good and stable data where states and actions and rewards are well defined and goal is verifiable. For instance in a game, did you kill your opponent or not? Can verify, and can assign rewards based on the time it took for you to do it. Just an example. None of these algorithms work on video tasks, real world tasks because real world is noisy, we don’t have action labeled datasets, and all the datasets are very high dimensional and large, it’s intractable to reason over the space.

This is the exact reason AGI is far away. Because current systems fail at reasoning from the physical world and to get to AGI we need to be able to develop systems that can reason from the world. When can we tell we have AGI? Consider this test. Leave your robot stranded into the wilderness of Africa. Give it meat. Its task is to heat it. But the only knowledge of heat that you inject to the model is the warmth from the Sun. The objective is to find something that can amplify this warmth and heat the meat, because eating raw meat leads to death. Leave it for 5 years. If it can discover Fire, you have AGI.