My Unedited Conversation with ChatGPT about Artificial General Intelligence

With the arrival of ChatGPT, the availability of Gato, and the soon to be released GPT-4, are we about to witness the point by which machines achieve general intelligence or is it all just fakery? I had a conversation with ChatGPT about this very same thing.

I asked him/her/it to “Write a short blog post that describes AGI?”

Here’s what I got back:

#ChatGPT: “Artificial General Intelligence, also known as AGI, is a type of artificial intelligence that has the ability to understand or learn any intellectual task that a human being can. It is a form of AI that has the capacity to think and reason like a human, and has the potential to perform a wide range of tasks without being specifically programmed for each one.

AGI is different from other forms of AI, such as narrow AI, which is designed to perform a specific task and lacks the ability to generalize its knowledge to new situations. AGI is not just a tool, it is a system with general cognitive abilities.

AGI is still in the early stages of development, but it has the potential to revolutionize the way we live and work. It could help us solve some of the world's most pressing problems, such as climate change, poverty, and disease. It could also lead to new technologies and industries that we can't yet imagine.

However, AGI also poses some risks, such as the possibility of superintelligent machines that could outsmart or outcompete humans. There is a need for careful research and development to ensure that AGI is developed in a responsible and ethical way.

Overall, AGI is an exciting and rapidly advancing field of AI research with the potential to bring about significant benefits to society. While there are concerns, we can mitigate these by having an open and transparent dialogue about the development of AGI and its implications.”

Me: I followed up by asking: Are you AGI?

#ChatGPT: No, I am not a true AGI. I am a large language model known as GPT (Generative Pre-trained Transformer) that is trained by OpenAI. While I have a large amount of general knowledge and can perform a wide range of language-based tasks, I lack the ability to understand or learn any intellectual task that a human being can. I am a narrow AI, specifically designed to perform certain language-based tasks such as answering questions, writing text, and so on. AGI is still a topic of research and it's not yet fully developed, but some researchers and companies are trying to develop AGI.

Me: I followed up again by asking: What is Gato?

#ChatGPT: Gato (Generative Adversarial Text-to-speech) is a type of Generative Adversarial Network (GAN) model for Text-to-Speech (TTS) synthesis. Gato is a neural network model that is trained to generate speech from text. It is composed of two main components: a generator network, which generates speech from text, and a discriminator network, which evaluates the speech generated by the generator and tries to distinguish it from real speech. The generator and discriminator networks are trained together in a process called adversarial training, where the generator tries to produce speech that is indistinguishable from real speech, while the discriminator tries to correctly identify whether a given speech sample is real or generated.

Gato is a relatively recent development in TTS synthesis and its goal is to improve the quality of the generated speech by training the model with a much larger dataset and by using a more powerful architecture. Also, it aims to make the generated speech more natural and closer to human-like speech.

Me: Not done playing around, I followed up with another question: How are you different from Gato?

#ChatGPT: I am a large language model, while Gato is a text-to-speech (TTS) model. I am trained to generate text based on the input I receive, while Gato is trained to generate speech from text. My main purpose is to understand and generate human-like text, such as answering questions, writing essays, summarizing, and so on. Gato, on the other hand, is focused on the generation of human-like speech from text, making the speech more natural and closer to human speech.

Another difference is the way we are trained, I am trained using a technique called unsupervised learning, where the model is trained on a large dataset without explicit labels, instead it learns to recognize patterns in the data and generate text that follows those patterns. Gato, on the other hand, is trained using a technique called adversarial training, it uses two neural networks, one that generates speech and other that tries to discriminate the generated speech from real speech, the goal is to train the generator to produce speech that is indistinguishable from real speech.

In short, I am a language model, while Gato is a TTS model, and we are trained for different purposes and in different ways.

Tristan Greene , editor of Neural at The Next Web, persuasively argues that it’s time for a reality check. https://undark.org/2022/06/02/the-long-uncertain-road-to-artificial-general-intelligence/

Tristan rightly says that "Gato is cool and there’s nothing quite like it. But, essentially, it is a clever package that arguably presents the illusion of a general AI through the expert use of big data. Its giant database, for example, probably contains datasets built on the entire contents of websites such as Reddit and Wikipedia. It’s amazing that humans have managed to do so much with simple algorithms just by forcing them to parse more data.

In fact, Gato is such an impressive way to fake general intelligence, it makes me wonder if we might be barking up the wrong tree. Many of the tasks Gato is capable of today were once believed to be something only an AGI could do. It feels like the more we accomplish with regular AI, the harder the challenge of building a general agent appears to be."

In conclusion, I gave ChatGPT the last word: How close are we to achieving AGI?

#ChatGPT: Artificial General Intelligence (AGI) is still a topic of active research and it is currently considered as a long-term goal in the field of AI. While significant progress has been made in recent years in areas such as machine learning, natural language processing, and computer vision, we are still far from achieving AGI.

Currently, the AI systems that exist are what is known as narrow AI, which are designed to perform specific tasks and lack the ability to generalize their knowledge to new situations. In contrast, AGI would have the capacity to think and reason like a human, and perform a wide range of tasks without being specifically programmed for each one.

Achieving AGI would require solving a number of technical challenges, such as developing algorithms that can learn from a wide range of experiences, understanding common sense and reasoning, and developing systems that can navigate the real world and its complexity. Additionally, AGI also poses some ethical and societal risks, which need to be considered and addressed.

It's hard to give a precise estimate of when AGI will be achieved, but some experts believe it could be decades or even centuries away. It's important to note that the development of AGI is a long-term goal that requires a significant amount of research and development, as well as collaboration between experts in various fields such as computer science, psychology, neuroscience, and philosophy.