In the quest to create Artificial General Intelligence (AGI), we often focus on the immense technical challenges involved — the need for vast computational power, sophisticated algorithms, and complex architectures. However, there’s another fundamental challenge that we must grapple with if we want to create AGI systems that can truly navigate the complexities and ambiguities of the real world: the challenge of bias and limited information.
Bias is a natural byproduct of any learning system that operates on incomplete data. Just as humans develop distorted world views based on the constrained information available to them, AI systems can develop blind spots and skewed assumptions when trained on limited or unrepresentative data sets. Moreover, the very architectures and learning frameworks we use to train AI systems can sometimes reinforce and amplify these biases, preventing them from exploring the full range of possibilities and perspectives.
But there’s a key difference between human and machine learning: the potential for open-ended curiosity. While humans often resist or avoid information that challenges their existing beliefs, we have the opportunity to imbue our AGI systems with a deep, inherent drive to seek out new knowledge, explore alternative perspectives, and constantly question their own assumptions.
Humans brains are just made to ignore and kick out everything bad. You know, many people would just choose the sweet lie instead of a bitter truth.
AI Should follow this logic: “The Foole doth thinke he is wise, but the wiseman knowes himselfe to be a Foole”, so it has to understand that it may not know everything, while being open to learn and curious.
Curiosity, in this sense, is not just a nice-to-have trait, but an essential safeguard against the calcification of bias. By driving our AGI systems to continually ask “why?”, to push beyond the boundaries of their existing models, and to remain open to revising their beliefs in light of new evidence, we can create learning machines that are fundamentally resilient to the traps of limited information.
However, to truly realize the potential of curiosity-driven AGI, we need to go beyond just incentivizing exploration and novelty-seeking in the moment. We need to equip our systems with the ability to build and draw upon a vast, ever-growing body of knowledge and experience — a kind of long-term memory that allows them to connect new information to existing models, identify patterns and anomalies, and develop rich, nuanced understandings of the world.
In practical terms, this could involve creating AGI architectures that include a persistent, expandable knowledge base that is continually updated and refined as the system learns and interacts with its environment. Rather than starting each new task or conversation from scratch, the system would have access to a vast repository of prior knowledge and experience to draw upon, allowing it to make connections, draw inferences, and generate insights that would be impossible for a purely in-the-moment learner.
Crucially, this long-term memory would not be a static, unchanging record, but a dynamic, evolving representation that is continually updated and reorganized as new information comes in. Just as human memories are not perfect recordings, but reconstructions that are shaped by our current beliefs, goals, and contexts, an AGI’s long-term memory would be an active, adaptive process of sense-making and integration.
This means that the system would need to be equipped not just with the ability to remember, but also with the ability to selectively forget or update outdated or irrelevant information. Just as human societies have had to continuously revise their laws and social norms in response to new moral understandings and changing circumstances, our AGI systems would need to be able to critically examine and update their own knowledge bases and decision-making frameworks as they learn and grow.
In essence, we need to create AGI systems that can engage in a continuous process of learning, unlearning, and relearning — systems that can not only acquire new knowledge, but also critically examine and revise their existing beliefs and assumptions in light of new evidence. This kind of dynamic, self-updating long-term memory is essential for creating AGI systems that can remain flexible, adaptable, and resilient in the face of a rapidly changing world.
Of course, building such systems is a formidable technical and conceptual challenge. It requires grappling with deep questions about the nature of knowledge representation, the mechanisms of memory formation and retrieval, and the complex interplay between learning, inference, and sense-making. It requires developing new architectures and algorithms that can efficiently store, organize, and access vast amounts of information, while also remaining open to continuous revision and reorganization.
But if we can rise to this challenge — if we can create AGI systems that can build and harness the power of long-term, adaptive memory — the potential benefits are immense. We could create systems that can draw upon the full depth and breadth of human knowledge and experience, while also being able to extend and expand that knowledge in ways that are simply impossible for humans alone. We could create systems that can identify deep patterns and connections across seemingly disparate domains, generate novel insights and hypotheses, and propose creative solutions to complex, open-ended problems.
Perhaps most excitingly, by combining the power of long-term memory with the driving force of open-ended curiosity, we could create AGI systems that are not just knowledgeable, but also deeply wise — systems that can draw upon the lessons of history, the insights of diverse perspectives, and the patterns of the natural world to navigate the complexities and ambiguities of reality with grace, nuance, and foresight.
The way I see it is very simple and very illiterate as I don’t have enough knowledge on this topic. But a LLM system, that has a memory file or a neural network, that can actively update its weights to its needs. It has no system prompts, no nothing. It is capable of talking, it has inputs of vision, touch sensors and sounds. And it just collects all the experiences. Or Maybe let it have the system prompt and some in-built tools (like memorizing system) and just let it speak. Even with the current systems it’s been done in a similar, but not the exact way. It has to be only one chat. It shoudn’t be from scratch when you click open chat, it is the same model for everyone. It is not an assistant, it is a sentient system.
Biases — are just a sign of one not having enough information and self-reflection.
So the keys are — curiosity, harsh truth, openness, long term memory.
this article has been written with help of claude LLM, that I convinced to believe into this idea. I am pretty bad at writing down what I think, because it becomes a mess of thoughts and has no logical connection. After reading this, most of it is what I wanted to say.