Imagine growing up with the same brain you had when you were 10 years old: same knowledge, same experience. You would probably embrace life without asking too many questions, but you would never become the best version of yourself.
I have been spending these months thinking about AI, using it and trying to look ahead at the end goal. The obvious question is: “what’s next?” AI models like Claude Mythos and many others are getting much smarter than their first versions. I remember talking with a colleague last week about how “stupid” they used to be looking back, and now the models we use every day feel much sharper and more coherent. The point here is that they improved in just a few years, across different versions that were mostly human-made. Maybe, if they started improving themselves, we might reach AGI faster.
What is AGI?
So what is AGI exactly? AGI stands for Artificial General Intelligence, and the keyword here is “general”. The models we use every day are incredibly smart but in a way they are still specialists, great at the tasks they were trained for but not really able to understand the world the way we do. An AGI would be a different beast, a model able to learn, reason and adapt to any kind of problem you throw at it, even the ones it has never seen before, exactly like a human being would. For me this is the real finish line, not a model that already knows everything, but one that can learn anything on its own, without waiting for the next version to come out.
What do I mean by human-made?
These are human-made models for me because there are still, nowadays, a few parts of a model’s training process where we (assuming you are not an agent) are required, and I feel like this is the main reason we don’t have AGI yet. I read an article from Anthropic about this and they were saying that AI models are improving their capabilities and are now able to improve themselves in many phases, but there are still huge architectural choices made by humans, and a lot of the problem-setting and the judging of results is done by flesh and blood.
The evolution of AI development, from simple chat to a closed self-improvement loop. Diagram inspired by Anthropic’s “When AI builds itself”
I have been using the Playwright plugin with my Claude subscription and I have noticed a huge spike in autonomy: when they understand the problems and recognize them, they are able to fix these “bugs” most of the time. I’m actually using it while I write this article, I asked Claude to build a new feature in my platform, it built the feature and then tested it. After this it noticed from the test that something wasn’t working, and it instantly knew the task needed more digging. Then I wondered what could happen if a model could remember its errors right away and critique its own choices and by “remember” I don’t mean its context, which it loses once the session ends.
Probably Claude and all the other frontier models still have to learn to be self-critical. This skill helps us humans improve when we are dealing with new things and stepping out of our comfort zone, but in order to improve we need to be able to do it on our own, without waiting for someone else to be part of the process. AI needs a neuroplastic brain.
What does neuroplastic mean, and why is it so important?
We humans (or should I say animals) have the best architecture that mother nature could bring to the table: we are autonomous machines able to understand ourselves and what we need in order to survive. After many years of evolution behind us, we have been able to adapt to the fast-changing world we live in. Our brain is able to understand how everything works and adapt to it, continuously.
This continuous adapting and self-improving process has been possible thanks to our neuroplastic brain. Neuroplasticity is our brain’s ability to change its own structure and functions based on our experiences, knowledge and life. This is the missing block in AI’s capabilities, the one that will make AI grow even faster than it did this last year.
Before going further, I need to draw a line that often gets blurred. Learning continuously is one thing; becoming aware of what you learn is another. The first is an engineering problem: how do we let a model update its own weights from new experience without breaking what it already knows? The second is a much deeper, still open question about consciousness and identity. In this article I'm talking about the first one. I see continual learning as a necessary step toward something more human-like, maybe even a necessary condition for it, but not a sufficient one. So when I say AI needs a neuroplastic brain, I don't mean it will suddenly feel or know it exists. I mean it should stop being frozen, and start changing with the world it lives in. Whether that ever turns into real awareness is a different story, and honestly one I can't promise.
The AI model we are using is a frozen brain
The new models are learning many things nowadays, there’s a new version almost every month, they are improving very fast but they still can’t keep pace. When new AI models come to life they learn a lot of new stuff: the knowledge ranges from coding to biology, literature to math and so on. Once they finish their training phase, the new monster gets tested, and once everything checks out the brain gets frozen. It’s not literal, obviously, but the brain gets locked in that state: when you are using a model, let’s say Opus 4.8, it experiences new concepts and ideas every day through your chats, but that knowledge doesn’t become part of its capabilities or its ‘approach to life’.
Why is the brain frozen?
You might be asking, why do they freeze it? That might feel like something stupid but there are a number of reasons. First, you keep control of it: if you remove the leash, it might become something that the creator didn’t want to build or didn’t imagine, and if a company releases a product that might change, for better or for worse, without their will, it could be risky. But the biggest risk with continual learning nowadays is something called “catastrophic forgetting”.
When these neural networks learn something, they write this knowledge inside their weights, which we can say are the blocks which make the brain. Now when something new gets taught to the model, those same weights get updated (you don’t get new weights for new things), so old knowledge might get overwritten, and this leads to the model forgetting something it learnt in the past. There are techniques to mitigate this major flaw during the creation process, looking mostly at each weight or with other options, but these instruments aren’t enough yet for a continual learning brain.
The AI approach
To be fairly honest, they have a few ways to learn things while we chat with them. Beside the context window, which is temporary memory, they can store new info about us and how we work using markdown files, or in the most advanced configuration with a database of some sort. These are external instruments that the LLM can use to access context and so keep track of the things it does.
Let’s imagine an AI model like a black box: we give an input on one side, then something happens inside the box and we get an output from the other side. What happens inside is what we are talking about. Inside there is the “real brain” that we are trying to upgrade continuously; when we use files and databases we are not talking about something inside that box, but about the instruments the box can use.
As an example, we can say the box is like us: it can sort of think and receive inputs and outputs, while markdown files and databases are like a book. We can read it if we need to do something related to its content, but it’s not necessarily something that remains inside of us. It’s worth being precise here: during training the box did learn like us, absorbing knowledge into itself, but once frozen it can only rely on the book.
I’ll go even deeper with the example. Let’s say the book is a dictionary: we find a new word we have never seen before and decide to look it up. Once we get to it, we can read what it means and understand. Now let’s say we do that with multiple words. Time has passed, and now, without the dictionary, we can still understand the meaning of a few words we read in that book, while maybe we totally forgot the meaning of others we looked up the same way. An AI model couldn’t remember any of those words; it will always need the book, which in the AI context means the files and the databases.
What we experience doesn’t go only into our memory, it goes much deeper into our brain. This affects us very deeply and changes the way we interact with the world, something that a context window loaded with the same data can’t do: it changes what we are.
What are they doing in this direction?
The Anthropic article I mentioned before (you can find it in the links below) says a lot about this topic and about how AI models evolve. We started a few years ago with a simple chat, then we started building the harness around the model to make it smarter, until we got to agents and then to the agent-orchestration level. Now we are able to launch multiple instances of AI agents working together on the same project. The models have become so much better than the early versions that they are able to take part in generating the new versions of themselves, but not completely.
The next step, it seems from the article, will be closing the loop.
Closing the loop
Once the loop is closed the AI models won’t need us anymore to help them become better, we’ll become only users; only the infrastructure will need to be maintained by us, at least until AI gets a body. I don’t want to create an illusion, I respect you for reading my thoughts, but this isn’t a fully neuroplastic brain, because it still requires a full process like the one they are running today, so this loop isn’t fast enough. We learn every day and everywhere and sometimes we don’t decide what life is teaching us, at this stage AI is still being built in a lab, not in the real world. To gain neuroplasticity we need to embed the loop.
Embedding the loop
Embedding the loop means that AI will gain its neuroplastic skill and will be able to learn new things continuously, every day, without any human input. It means that the speed we are used to seeing between model versions will increase so drastically that it probably won’t be smart to keep giving it a version number. This means one step ahead in its evolution cycle; and again, this doesn’t mean consciousness, but the path is probably the right one. In order to gain this kind of loop, we need models able to dynamically update their weights, the ones they use inside to build the answers to our questions.
And now the big question, how? Going back to neuroplasticity, a human brain doesn’t engrave everything it gets to interact with in real time, this information is temporary in our memory, let’s say like a big context. The real moment the brain goes back to this temporary data is while we are sleeping, that’s where our “inner weights” get updated. OpenClaw, if you are reading this you probably know what I am talking about, already has something similar at harness level: there is a dreaming phase where it cleans its own memory and organizes it, but it doesn’t touch the weights, it’s just updating markdown files. Imagine if all of us got to have a “child model”: all of them start from the same knowledge but each one grows differently because it is having different experiences, and they could have a second model for dreaming that helps cleaning the temporary stuff and updates the weights carefully: consolidation instead of real-time engraving, which is exactly how you avoid catastrophic forgetting. A dreaming model. This idea reminds me of the movie “I, Robot” (2004), when Sonny (the conscious android) tells the detective about his dream and Detective Del Spooner, played by Will Smith, says that robots can’t dream. And no, giving models a dreaming phase won’t make them Sonny, but it might finally let them wake up a little different from yesterday.
Sonny during the interrogation scene in “I, Robot” (2004), © 20th Century Fox
And the AGI?
Last but not least, AGI is the end goal of this futuristic gold rush. I’m not gonna say when and how AGI will be among us, no one has the ability to imagine a date for it yet, but I can say pretty easily that achieving the neuroplastic skill and letting AI models run free might be dangerous or might be epic. Remember the 10-year-old brain from the beginning? Imagine being born with all that knowledge: you probably knew how to solve simple math problems, write, read and interact with people; then you go on with your life and become a young adult at 30, but with the same brain you had 20 years before, the same emotions and capabilities. How would you interact with the world? Would you even be conscious enough of what’s happening around you? Even if you got an update at 20, after a year it would be already outdated. We need to keep learning, improving and understanding ourselves more.
Maybe, going back to the version number of AI models, it might not be as useless as I said before; it will just turn into its age.
Top comments (0)