You’ve Been Brainwashed: ChatGPT Is Stupider Than You Think!

Author(s): Koza Kurumlu

Originally published on Towards AI.

You’ve Been Brainwashed: ChatGPT Is Stupider Than You Think!

The incredible advancements in artificial intelligence, with some help from the media, may have you convinced that we’re closer to achieving artificial general intelligence (AGI) than ever before. But hold on for a moment — AI is actually much dumber than you might think. Despite its impressive accomplishments, such as defeating the world-class “Go” champion and acing college admission tests, AI still lacks common sense and is far from reaching AGI.

In this article, we’ll dive into the limitations of AI, particularly large language models like ChatGPT, and examine why they’re not as intelligent as they appear. We’ll also explore the nature of intelligence, the possibility of pairing AI with a logic machine, and the potential future of AI beyond AGI.

Limitations of AI and the Myth of Scale:

AI has grown tremendously in recent years, with some models, like ChatGPT, trained on tens of thousands of GPUs and a trillion words. These extreme-scale AI models seem to demonstrate sparks of AGI, but they often make small, silly mistakes.

The popular belief is that simply increasing the scale and adding more resources will fix these errors. However, this brute-force approach to AI development has several negative consequences.

First, the extreme expense of training large-scale AI models leads to a concentration of power in a few tech companies, making independent research difficult. Second, the massive carbon footprint of these models has significant environmental impacts. Finally, AI without robust common sense will never reach the potential you think. Models such as ChatGPT have reached their peak. Let’s take a look at the graph below. It demonstrates the size of these models, displaying how many parameters they have, and also plotting their accuracy, i.e., the model’s performance.

If we draw a curve of best fit through this, it will be even more apparent that there is a cap to the number of parameters that will bring about a massive increase in performance. It’s not a linear relationship. So now we’ve established massive change isn’t coming using language models such as ChatGPT; let’s discuss ChatGPT itself.

What is Intelligence, Then?

To understand why AI is not as intelligent as it may seem, we must first define intelligence.

Intelligence is the ability to learn, understand, reason, and adapt to new situations. It involves problem-solving, creativity, and the application of knowledge.

Intelligence requires explanation based on a causal understanding of the world and not just prediction. Human intelligence is not just predictive processing because it also excels at pattern recognition, explanation, evaluation, selective memory, and communication. Which ChatGPT doesn’t — so let’s look at why.

Why Does ChatGPT Lack Intelligence?

ChatGPT, while an impressive language model, lacks the common sense and understanding of the context that characterizes human intelligence. This is primarily because it relies on pattern-matching from the vast amount of data it has been trained on, rather than understanding the underlying concepts. In short, how it works is that incredible amounts of text are passed into the AI model and it learns the relationship between these words.

More specifically it learns the probability of one specific word following another one. Therefore when it’s writing it uses entirely probabilistic methods to determine the next word.

Let’s look at an example.

Here we can see the percentage probability of a corresponding word occurring after the previous, and of course, full stops and just stopping have corresponding probabilities as well. As a result, it often produces answers that sound plausible but are incorrect or nonsensical when analyzed closely. One reason for this shortcoming is the absence of a proper grounding in reality. While ChatGPT is trained on a massive amount of text data, it does not have direct experiences or interactions with the real world: it’s just these probabilities. Consequently, it cannot form a coherent understanding of the world and is unable to reason about it effectively.

This leads to ChatGPT just being a guesser based on memory, meaning it’s unreliable, inconsistent and lacks understanding. In addition, we also know that it won’t be able to advance significantly, therefore there is no future for ChatGPT or any other LLMs other than writing aids, or is there?

Can ChatGPT Be Paired with a Logic Machine, to Create Intelligence?

So we know that ChatGPT lacks intelligence but is impressive with language, so why can’t we add the intelligence on? And yes, it is possible to combine the capabilities of ChatGPT with a more logical system, like Wolfram Alpha, to create a more intelligent AI. Wolfram Alpha specializes in computational knowledge and can process complex mathematical and scientific queries. By combining the strengths of both systems, a more intelligent and versatile AI could be created.

However, this would still not constitute AGI, as the system would still have limitations and lack the full range of human-like understanding and reasoning. For such a hybrid system to be effective, it would need to overcome several challenges.

1. It would require seamless integration between the language model and the logical reasoning engine. This would necessitate a shared understanding of the context and meaning of user queries, as well as a mechanism for coordinating the capabilities of both systems. And honestly, this is where my doubts lie. Logic models, Wolfram Alpha, will have trouble dealing with such inconsistent LLMs, as it requires consistent inputs.
2. The hybrid system would need to be capable of learning and adapting over time. This could involve incorporating feedback from user interactions, updating the knowledge base, and refining the reasoning and language generation capabilities. Ultimately, the goal would be to create a system that can truly understand and respond to complex queries in a way that demonstrates intelligence rather than simply regurgitating pre-existing knowledge.

The Future of Large Language Models (LLMs) if Not AGI:

Even if LLMs don’t achieve AGI, they still have significant potential for practical applications.

They can be used in areas such as customer support, content creation, translation, and many other fields. By refining the technology and addressing its limitations, LLMs can become increasingly valuable tools for a wide range of industries.

One potential avenue for improving LLMs is to focus on incorporating other types of data, such as visual or auditory information, into their training. This could help AI models develop a more comprehensive understanding of the world and improve their reasoning and problem-solving capabilities. Additionally, incorporating mechanisms for learning and adapting over time could help make LLMs more contextually aware and better able to engage in meaningful interactions. Another critical area of focus should be on ensuring that AI models are ethical, safe, and aligned with human values. This will involve developing better methods for curating and filtering training data, as well as designing algorithms that can recognize and correct biased or harmful outputs.

Conclusion

Although I have slandered ChatGPT massively throughout this article, it’s not all bad. As I’ve mentioned, it’s text-generating abilities are unmatched and can blow people away, and this is where the danger lies.

Those who don’t truly understand the inner workings of LLMs may assume that it has identical results with logic-based queries, and rely on it for certain tasks.

But hopefully, in this article, I’ve persuaded you otherwise and given you a glimpse of what is currently going on in the AI world behind all the flash media headlines.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI