Why you should not worry about large language models like GPT-3

OpenAI released a product recently called ChatGPT. It became viral instantaneously, and people started wondering how far AI could go because it could generate excellent output and answer questions thoughtfully. But if we take a deep look inside the black box, you will see that, for the moment; they are not what you may think they are.

There are many reasons why large language models, such as GPT-3, Cohere, Ai21Labs,… should not be a cause for concern. Some of these reasons include:

Large language models are not conscious:

Large language (LaMda, GPT, GPT-3, Cohere, Ai21Labs,…) models are not conscious because they are simply algorithms designed to process and generate language. They do not have the same kind of consciousness or self-awareness that humans do, and they are incapable of making decisions or acting independently.

Consciousness is a complex and poorly understood phenomenon, and it is not clear what exactly is required for a system to be conscious. Some researchers believe that consciousness arises from complex computations by the brain, while others believe it is an emergent property of complex systems.

Regardless of the underlying mechanisms of consciousness, it is clear that large language models do not have the same kind of consciousness that humans do. They do not have the ability to experience or perceive the world, and they do not have the ability to make decisions or act on their own. They are simply algorithms that are designed to process and generate language.

Let me materialise that by presenting something straightforward about statistical natural language processing. Let’s imagine you have a baby with knowledge about whatsoever, and you say to this baby every that after an “I”, you will see an “am” and then an “at” and then “school”. The next time you ask this baby to generate a sentence, it will say, “I am at school”. Imagine you do that with more possibility by adding the amount “am” and “will” after “I”, and you do it at scale (With a really large amount of data). When you start a sentence like “I am”(called the prompt in those language models), you will be surprised by how good mathematics can make this model predict something that seems intelligent. But it is not. At the level we are right now, we don’t have the proper process and methods to make a model intelligent. It may change in the future (which will be great), but we don’t have that right now.

Large language models do not have access to the internet

Most large language models (LaMda, GPT, GPT-3, Cohere, Ai21Labs,…) do not have access to the internet because they are trained using large amounts of data collected and stored offline. While some large language models may be able to process and generate text based on data they are given, they cannot gather new information or interact with the world as we could.

There are several reasons why large language models are typically trained offline. One reason is that it allows for more control over the data used to train the model. By collecting and storing the data offline, researchers can ensure that the data is high quality and free from biases or other sources of error. Online Training could be really bad as the model will learn almost everything we have online (And that is irresponsible of us).

Training large language models can be computationally intensive and require a lot of computational resources. By training the models offline, researchers can use specialized hardware and software to speed up the training process, which would be impossible if the models were connected to the internet.

Large language models are not perfect:

Those models (LaMda, GPT, GPT-3, Cohere, Ai21Labs,…) are not perfect because they are simply algorithms designed to process and generate language. Like any other algorithm, they can make mistakes or generate nonsensical outputs, and they cannot reason or make decisions like humans.

One of the huge biases about data-dependent fields is the data source. The Internet is one of the most controversial sources of information. There is much accurate information, opinions, and also non-verified data, … And when you train your model with those data, it’s inherited from those biases. They are based on probabilistic models, which can make predictions based on the likelihood of different outcomes. This means that large language models may generate outputs that are not always correct but that is more likely to be correct based on the data that they were trained on.

Also, Large language models such as Ai21Labs Jumba and large, Cohere, and GPT-3 are trained on large amounts of text data. However, this data is still finite and cannot capture human language’s full complexity and diversity. As a result, large language models may make mistakes or generate nonsensical or inappropriate outputs.

Large language models can be controlled and regulated

Large language models (LaMda, GPT, GPT-3, Cohere, Ai21Labs,…) can be controlled and regulated because they are simply algorithms designed to process and generate language. They can be designed to follow specific rules or guidelines and be monitored and modified as needed.

One way large language models can be controlled and regulated is by defining the types of output they are allowed to generate. For example, a large language model could be designed only to generate text appropriate for a specific audience or follow certain ethical guidelines. This can be accomplished using techniques such as filtering or censoring, which can remove inappropriate or offensive content from the model’s outputs.

Another way large language models can be controlled and regulated is by monitoring their performance and making adjustments as needed. For example, suppose a large language model generates biased or inaccurate outputs. In that case, researchers can analyze the data that the model was trained on and identify sources of bias or error. They can then adjust the model or the training data to improve its performance.

Overall, while large language models are powerful and impressive, they do not threaten humans, and there is no reason to be concerned about their existence. They are simply probabilistic models capable of generating text that makes sense to humans. Let’s hope we will have fully intelligent systems helping us in our day-to-day tasks in the future.