Why ChatGPT will never be an Artifical General Intelligence
Understand GPT’s inner Workings and the Difference to AGI
I engange with generative AI since several years now.
The first GPT (Generative Pre-trained Transformer) large language model was released in 2017 by OpenAI. You could already play around in a graphical editor with the GPT playground, way before ChatGPT was released.
The first blog post from OpenAI regarding GPT-3 was published March 2021 — and today it was only until a few weeks that every user was able to even use GPT-4 in ChatGPT, instead of GPT-3.
Still it feels like the whole world discovered something completely new at the end of 2022, when ChatGPT was released.
Who could blame them.
The UI was slick and easy to use, the output fine-tuned to give a natural, human feel than the usual GPT, and several resourceful developers and entrepreneurs spit out AI services with one marketing headline catchier than the other.
Along with this newfound attention, however, came a surge of speculation that we were on the brink of creating an Artificial General Intelligence (AGI).
Suddenly, the public consciousness was abuzz with the notion that we had birthed an AI that could do anything and everything.
But, that’s not the case.
Especially not with ChatGPT or other large language models (LLMs).
Let me explain why.
Understanding the difference between AI and AGI
Before we dive into the depths of ChatGPT’s workings, let’s clarify the distinction between AI and AGI.
AI, in the sense we commonly discuss it today, refers to systems designed for a specific task. They’re like gifted specialists, excellent in their niche but unable to branch out (dumbies..).
AGI, on the other hand, is a hypothetical intelligence that matches or surpasses human capability across all economically valuable work. It’s a universal genius, capable of learning any intellectual task that a human can do.
There are huge difference between the two.
Task-specific vs. General Purpose
- AI: Is designed to perform a specific task, such as language translation or image recognition. It cannot perform tasks outside of its specific domain.
- AGI: Is designed to perform any intellectual task that a human can do. It can understand, learn, adapt, and implement knowledge across a wide array of tasks.
Learning and Adaptability
- AI: Narrow AI systems can learn and adapt, but only within their specific domain. They require a large amount of training data and cannot easily transfer knowledge between different domains.
- AGI: AGI would have the ability to learn from experience, reason, plan, and solve problems in ways that are not predefined by its programming. It would have the ability to transfer learning from one domain to another, similar to how humans can.
Understanding and Awareness
- AI: Narrow AI does not truly understand the tasks it performs. It can analyze and respond to input data based on its programming and training, but it does not possess consciousness or self-awareness.
- AGI: In theory, AGI would have a level of understanding and potentially even self-awareness, allowing it to comprehend complex concepts, contexts, and abstract ideas, similar to a human.
Development Status
- AI: Many forms of narrow AI have been developed and are in use today in various applications, such as speech recognition, image recognition, and natural language processing.
- AGI: AGI is still theoretical and has not yet been created. Scientists and researchers are unsure when or if it will be possible to develop AGI.
How GPT Language Models work
In order to understand why GPT won’t reach the status of an AGI, it helps to understand how these transformers work.
Think of GPT as an incredibly well-read author who has read and memorized millions of books, articles, and websites but doesn’t truly understand any of it.
When you ask it a question, it doesn’t "know" the answer in the way a human does. Instead, it looks at what you asked and compares it to all the sentences and responses it has seen in the past. Then it comes up with a reply that matches the patterns it has seen before.
Let’s imagine you’re playing a game of word association with this author. You say a word, and the author responds with another word that it thinks is closely related based on all the texts it has memorized.
For example, if you say "beach," the author might say "sand" because those two words often appear together in sentences.
In the same way, when you give GPT a prompt or question, it looks at all the "words" (or rather, chunks of words) you’ve given it and generates a piece of text that follows a similar pattern to what it has seen before in its training data.
It’s important to note that ChatGPT doesn’t truly "understand" the sentences it generates.
It’s following patterns based on its training, not drawing on a personal understanding of the world or the concepts involved.
It doesn’t really understand what the books are about.
And while ChatGPT may seem like it remembers past interactions, it actually doesn’t. Each question you ask is responded to in isolation. It sends the whole conversation history with every new prompt.
But why can’t ChatGPT reach the status of AGI?
Okay, now we have some fundamental knowledge about the differences of AI and AIG and how ChatGPT works.
This lets us simply break the answers down into five key reasons.
Reason 1: Lack of True Understanding
As discussed earlier, our well-read author, who has devoured and memorized millions of books, doesn’t truly comprehend them.
When asked a question, ChatGPT sifts through its mental library, locates related information, and weaves an answer without a genuine understanding of the content.
It doesn’t comprehend the text it generates; instead, it identifies patterns based on its training data and generates a response mirroring those patterns.
Reason 2: Absence of Transfer Learning
ChatGPT, like our well-read author, remains confined to its expertise in text generation and manipulation. Its ability to learn new skills outside its specific domain is non-existent.
Unlike humans who can apply lessons from one area of their life to another, ChatGPT can’t transfer knowledge or skills across various domains.
It can generate and execute code, use that code on data like images and audio. Still, it only uses the transformer concept — the additionsl functionality is programmed.
Reason 3: No Consciousness or Self-awareness
This one might seem obvious, but it’s worth mentioning. ChatGPT doesn’t have a sense of self, emotions, beliefs, or consciousness.
Any appearance of such is an illusion created by the patterns it’s learned to generate and especially by fine-tuning it the respond in a more human-like style.
Our author may write heartfelt poetry, but there’s no emotion behind those words, only patterns learned from other poems.
Reason 4: Limited Adaptability
ChatGPT operates within the confines of its training, making its adaptability limited.
While it generates a myriad of responses based on the inputs it receives, it can’t evolve or improve its capabilities outside its training.
It can’t learn from past interactions as humans do, and each question is responded to in isolation.
Embedding of vector databases can be used to provide internal data and your conversations will be used to further fine-tune ChatGPT. Still, it doesn’t adapt other than increasing its already large amount of text data.
Reason 5: No Goal Setting or Planning
Finally, ChatGPT can’t set goals or make plans.
Unlike a human who might plan out their day or set goals for their work, ChatGPT responds to each input with no broader objective beyond generating a plausible-seeming text response.
There’s no foresight or planning involved in its operations.
With the rise of LLM driven agents it seems like this argument won’t be valid in the future. But in the end these agents are only programs with logic to use these LLMs to fulfill specific tasks.
In Conclusion
Despite these limitations, ChatGPT is still a testament to the power of AI and its potential to revolutionize our world.
However, like the author who doesn’t truly understand the content they’re writing, ChatGPT serves as a reminder of the gap that still exists between AI and AGI.
We can extend its features by programming plugins and code execution, give it the ability to recognize images using a Python machine learning library and extend its knowledge to internal data via vector database embedding.
But all of this are only different systems, programmed to work together to extend capability as a whole — but it’s far from an Artificial General Intelligence.
And maybe that’s better for all of us!
