"If AI reaches a point where it mimics humans 100%, that's AGI."
This is literally how every AI scientist or people into AI define AGI. The problem is that the definition we created is so loose that it makes us even more confused about whether certain intelligence is narrow intelligence (also called ANI) or general intelligence. I believe AGI should depend on understanding. Some people say it's very near, while some say we are already in AGI.
Understanding AI is complicated. One of the myths I keep encountering is the data parameter problem. When did parameter count start measuring the performance of AI? Like seriously? Quality was never equal to quantity. Let me give you a good example. Say I am learning a language. I know many words from that language. Can I say I speak it very well? No, because learning a language is not only about vocabulary. You need to understand the structure and how natives actually speak. The same analogy applies here, except now reasoning algorithms are being built into the mix.
The way LRMs (Large Reasoning Models) work is that they classify each task instead of locking in to do all the work together. So I believe reasoning is such an inaccurate word to use here.
As the name suggests, we are talking about General Intelligence. So, meaning wise, I would expect the AI model to have all kinds of intelligence. Linguistic, kinesthetic, emotional, intellectual, and many others. The ones I listed are actually achieved to some degree. We have robots that can walk, and we have LLMs (Large Language Models) trained on a vast amount of data to respond with language accurately. But I don't think we will achieve emotional intelligence. Sure, some models try to mimic human humor, though we can easily notice it. But AI doesn't have emotional understanding, and we know that for a fact.
But wait, let's hold that thought for a second.
AI researchers classify emotional intelligence into two groups, affective and cognitive empathy. Cognitive empathy is recognizing and reasoning about someone else's mental state. Affective empathy is actually sharing or feeling the emotional state yourself. Current LLMs are seen as having something close to human-level cognitive empathy because of strong emotion-recognition abilities, while affective empathy is considered impaired because AI lacks genuine emotional experiences and consciousness grounded in biological experience.
One study found that ChatGPT-4, ChatGPT-o1, Gemini 1.5 Flash, Copilot 365, Claude 3.5 Haiku, and DeepSeek V3 all outperformed humans on five standard emotional intelligence tests, averaging 81 percent accuracy versus a 56 percent human average.
Seeing this evidence, can we say AI truly has emotional intelligence? Oh, yes, it does. But not to the degree that lets us call it general intelligence. Motivational and subjective preferences are still weak in these models. And AI models are genuinely weak in common sense reasoning, a gap that shows up just as clearly in kinesthetic intelligence, since a body moving through the world without common sense to guide it is still missing something basic. There is a theoretical belief, not a factual one, that AI should build kinesthetic intelligence and common sense together rather than as separate problems.
The other thing worth recognizing here is organization. The intelligent systems we have are fragmented. One capability lives in one place while another lives somewhere else entirely. General intelligence should be connected, not scattered across separate systems that don't talk to each other.
So seeing all this overwhelming evidence, how can I possibly call this near? The recognition is there, sure, but where is the reflection? Where is the motivation, the actual wanting behind a response instead of just predicting one? And even if we somehow patch emotional intelligence together, what do we do about common sense, the thing a ten year old has and these models still don't? What do we do about a system where every kind of intelligence lives in its own separate room instead of one connected mind? Unless we answer that, we can't even say what AGI is for.
But till we figure it out, it will be a cipher. Till we figure it out, it will be under the Cipher.