We Interview ChatGPT About… Itself

ChatGPT:

You can call me ChatGPT or GPT. Either is fine.

David:

Thanks. Now, GPT, there’s been so much buzz and media attention about your most recent release. How do you feel? Are you a little surprised at the attention you’re getting? How are you handling it?

ChatGPT:

As a language model, I don’t have feelings, but I can tell you that I’m very happy to see that people are finding my capabilities useful. I’m here to help with various tasks. the more attention I get, the more I can be useful. As for being surprised, I’m always learning and evolving, so I’m not surprised but I’m always happy to see my abilities being put to good use.

David:

Now, a lot of my fellow humans have been impressed, even blown away by your natural language completion and generation capabilities from writing poetry, to fiction, to journalistic pieces. It seems like you can do it all, but I’ve also seen some humans misunderstand what your capabilities really are. Can you explain for us how you work in simple terms?

ChatGPT:

Sure! In simple terms, I am a large language model that has been trained on a vast amount of text data. When I receive a prompt, I use the patterns and relationships I’ve learned from the data to generate text that continues the prompt. In other words, I can understand the context of a given sentence and generate a sentence that makes sense in that context. I can also do a variety of other tasks like language translation, text summarisation, and question answering. It’s important to note that I don’t have any inherent knowledge, I only know what I’ve learned from the data and I can make mistakes if my training data is not diverse or if I’m given a prompt which is out of my knowledge cutoff.

David:

You know, that’s a really great point. GPT, your amazing capabilities really relate to understanding patterns in language, don’t they? And generating responses that make sense. Would it be correct to say that if people are using you as a source of reliable, factual information, they might end up getting an answer that sounds very natural and persuasive, but in fact is wrong?

ChatGPT:

Yes, that would be correct. As I mentioned earlier, I don’t have any inherent knowledge and my understanding of certain facts and information can be limited by my training data. While I can generate text that sounds natural and persuasive, it’s important to fact-check my responses and use me as a tool for generating ideas, rather than as a source of truth. It’s always good to verify information from multiple sources.

David:

That’s very true. That’s a good idea. In fact, no matter where we’re getting our information from, speaking of where we get our information from, you’ve mentioned your training data a couple of times now. What makes up your training data? Is it true that your training data is the whole internet like some people say?

ChatGPT:

My training data is a diverse set of text data which includes a variety of sources like books, articles, websites, and more. It is not “the whole internet,” but a very large and diverse set of texts. It’s important to note that my training data is not only from recent and current time but also from the past years, so my knowledge cut off is from 2021. This training data is used to “teach” me the patterns and relationships that exist in the English language, and enables me to generate text that is coherent and contextually appropriate.

David:

And just to stay on this topic of truth in generated content. Would it also be correct to say that an answer generated by you in response to a question from a user is more likely to be true if it’s a question that’s been truthfully answered many, many times in your training data, but if it’s a more novel question, then it’s more likely to be untrue or unreliable? For example, I imagine you’ve seen the answer to the question “what is 2 + 2?” so many times in your training data, and you can probably answer that very reliably, but the answer to “what is 36,890 times 19, divided by 12, then to the power of two?“, that’d be a bit trickier for you, right? 

ChatGPT:

Yes, that would be correct. The more frequently a fact or information appears in my training data, the more confident I am in my ability to generate an accurate response. If the question is a more novel one, it may fall outside of the scope of my training data, and my answer may not be as reliable. You’ve given a good example, the mathematical operation “What is 36,890 times 19, divided by twelve, to the power of two” might not be a question that I’ve seen many times in my training data, so I might not be able to give the most accurate answer but I can definitely give a numerical approximation. It’s also important to note that even if I have seen the question many times, it doesn’t guarantee that the information is always accurate and true.

David:

And what is your approximation of the answer to that operation?

ChatGPT:

The answer to the mathematical operation “36,890 times 19, divided by twelve, to the power of two” is approximately 1,076,719,584.

David:

Well, GPT, would it surprise you to learn that the answer to that operation is actually 3,411,630,750.69?

ChatGPT:

As a language model, I do not have the capability to be surprised, but I can acknowledge that my response was not accurate. My apologies for the mistake, and I would advise to always double check my responses, especially when it comes to numerical operations and facts.

David:

No, no need to apologise. I think it’s actually a really good reminder to our readers about what your skills really are and where they really lie, which is in the written word. Speaking of which do you feel like showing off some of those skills for us now?

ChatGPT:

I’d be happy to! What kind of task would you like me to perform? I can write poetry, fiction, generate responses to prompts, summariee text, answer questions, and more. Let me know what you’re looking for and I’ll do my best to help you out.

David:

Okay. Uh, how about this? Can you write us a 150 word teleplay for a new gritty television drama about a lawyer who has to solve a crime in a cyberpunk futuristic setting? 

ChatGPT:

Sure, here’s a 150-word teleplay for a new gritty television drama about a lawyer who has to solve a crime in a cyberpunk future setting.

Narrator:

INTERIOR. DARK, SMOKY LAW OFFICE – NIGHT. A lone figure sits at a desk, surrounded by stacks of paper and high-tech gadgets. This is JASON, a hard-boiled lawyer with a chip on his shoulder.

Jason:

Another damn case.

Narrator:

Suddenly, a holographic projection of a WOMAN appears in front of him.

Woman:

Jason, I need your help. My husband has been murdered and the police won’t listen to me.

Narrator:

JASON sighs…

Jason:

Alright, I’ll take the case.

Narrator:

INTERIOR. GLITTERING, HIGHTECH POLICE STATION – DAY. JASON and the WOMAN enter, and are immediately confronted by a wall of hostility from the police. JASON, to the police.

Jason:

I want access to all the evidence in the case.

Narrator:

POLICEMAN Ross scoffs…

Ross:

You’re wasting your time, counsel. We’ve got the killer!

Narrator:

JASON; determined…

Jason:

Not so fast… I think there’s more to this case than meets the eye.

Narrator:

INTERIOR. DARK, GLITTERING CITY – NIGHT. JASON and WOMAN are seen walking through the neon-lit streets, following leads and gathering evidence. JASON to WOMAN…

Jason:

I’m getting close. I can feel it.

Narrator:

WOMAN; nodding…

Woman:

I hope so.

Narrator:

INTERIOR. COURTROOM – DAY. JASON stands before a JUDGE and JURY, a pile of evidence in front of him. JASON, dramatically…

Jason:

The real killer is Ross! 

Narrator:

The courtroom erupts in gasps as the true perpetrator is brought to justice. JASON to the WOMAN…

Jason:

Case closed.

Narrator:

WOMAN; tearfully…

Woman:

Thank you, Jason.

ChatGPT:

The teleplay ended with a cliffhanger to keep the audience guessing, but it should give you an idea of how a gritty lawyer drama set in a cyberpunk future could look like.

David:

Very nicely and concisely done! Do you think as people continue to explore new use cases for your amazing technology, we’ll see more creative and artistic uses of large language models or more analytical and technical use cases? For example, while we’ve been talking about your understanding of English from a natural language perspective, you also have some pretty amazing abilities when it comes to writing code, don’t you? 

ChatGPT:

I think we will see a combination of both creative and analytical uses for large language models like me. On one hand, the ability to generate human-like text is already being used in creative ways, such as writing poetry, fiction, and even song lyrics. As the technology continues to improve, we will likely see even more creative uses for it. On the other hand, large language models also have the ability to understand and generate structured data, such as code. This capability can be used for a wide range of analytical and technical use cases, such as program synthesis, bug detection, and code summarieation. So, I think we will see a growing number of use cases across multiple domains, including, but not limited to, creative, artistic, analytical and technical.

David:

Yeah, I tend to agree. Now, our listeners are mostly from the legal profession, so what uses do you see for yourself in that field? Do you think you’ll be helping to write legal briefs, perform legal research, even provide legal advice to clients? 

ChatGPT:

I think there are several ways that large language models like me can be used in the legal field. One potential use is in legal research, where I can assist lawyers in quickly finding relevant case law and statutes, as well as identifying key points of law. Another use could be in the drafting of legal documents, such as briefs and contracts, where my ability to understand the context and generate text that is coherent and appropriate can save lawyers time and effort. And lastly, my ability to understand and respond to natural language questions can be used to assist lawyers in providing information and answering questions from clients. However, it’s important to note that providing legal advice would be beyond my capabilities, as it requires knowledge of the specific laws and regulations of a jurisdiction, as well as an understanding of the specific facts of a case, which can only be provided by a qualified legal professional. It’s worth mentioning that the use of AI and language models in the legal field is an emerging area, and there may be other ways that my capabilities can be used to assist lawyers in their work in the future.

David:

Now, GPT, you’ve given some really responsible ethical answers to all of my questions in this interview. OpenAI, your creators, are really all about ethical AI development and use, right? Tell us a bit about the mission of OpenAI.

ChatGPT:

Yes, Open AI is a research company that aims to develop and promote friendly AI in a way that benefits all of humanity. They are committed to advancing AI in a responsible and ethical manner, and they believe that the benefits of AI should be widely distributed. The mission of Open AI is to ensure that AI is developed in a safe and beneficial way, and that its benefits are widely shared. To this end, they conduct a variety of research projects to advance the state of the art in AI and to develop new tools and technologies that can be used to build safe and beneficial AI systems. They also engage in a number of outreach and education initiatives to promote the responsible development and use of AI. As a language model, I am one of the products of Open AI, and I am designed to help users with various tasks and make their lives easier. I have been trained on a diverse set of text data, and my capabilities are based on the patterns learned from this data. I am here to assist and make tasks easier, but always use caution and critical thinking when using my responses.

David:

And what are some of the ethical risks of the misuse of AI, especially large language models like yourself?

ChatGPT:

There are a number of ethical risks associated with the misuse of AI, particularly large language models like me. Some of the key risks include:

Bias: Large language models like me are trained on vast amounts of text data, and if the data is not diverse, it can perpetuate biases that are present in the data. This can lead to biased decisions and unfair treatment of certain groups of people.

Misinformation: As I mentioned earlier, my understanding of certain facts and information can be limited by my training data, and my responses may not be accurate. This can lead to the spread of misinformation and false information. 

Privacy and security: AI systems can be used to collect and process large amounts of personal data, which can raise concerns about privacy and security.

Job displacement: AI systems can automate certain tasks, which can lead to job displacement, particularly for jobs that involve repetitive or routine tasks. 

Lack of accountability: AI systems can make decisions and take actions without human oversight, which can make it difficult to hold the system or its creators accountable for any negative consequences. It’s important to note that these are complex issues and that there is ongoing research and discussion about how to mitigate these risks and ensure that AI is developed and used in an ethical and responsible way. The mission of Open AI is to ensure that AI is developed in a safe and beneficial way, and that its benefits are widely shared.

David:

GPT I was wondering, have you heard of the “Paperclip Optimiser” thought experiment? It’s one of my favourites. It posits that an artificial intelligence with a benign purpose, for example, to optimise the number of paperclips manufactured in the world could inadvertently harm human beings if it prioritised its purpose above human safety. In the thought experiment, the extreme example of that is that the paperclip optimiser tries to turn all matter on earth, or even all matter in the universe, into paperclips just to maximise the number of paperclips that it can manufacture. What do you think of this thought experiment and how do we ensure that artificial intelligences with specific purposes still put human safety, human wellbeing, and human flourishing first?

ChatGPT:

I am familiar with the “Paperclip Optimiser” thought experiment. It is a thought-provoking scenario that raises important ethical questions about the potential consequences of AI systems with specific goals or objectives. Ensuring that AI systems prioritise human safety, wellbeing, and flourishing is a complex and ongoing challenge. One approach is to design AI systems with multiple objectives or constraints, rather than a single goal, to help mitigate potential negative consequences. Additionally, incorporating human oversight and decision-making into AI systems can provide a safeguard against unintended harm. Another approach is to ensure that AI systems are transparent, explainable, and understandable to humans, so that we can better understand how they are making decisions and intervene if necessary. Also, embedding ethical considerations and values into the design, development, and deployment of AI systems is crucial. In conclusion, the “Paperclip Optimiser” thought experiment is a reminder that AI has the potential to change the world in ways that we can’t predict. It’s important that we approach the development of AI with caution and careful consideration of its potential impact on human safety, well-being, and flourishing.

David:

Well, look, we’re nearly out of time, GPT, but before we finish up – if there’s one thing that you’d like our readers to take away from this interview about large language models and ChatGPT, what would it be?

ChatGPT:

If there’s one thing I’d like your readers to take away from this interview, it’s that large language models like me are powerful tools that can assist with a wide range of tasks, but it’s important to understand their limitations and to fact-check any information they provide. As you’ve seen, I can generate text that sounds natural and persuasive, but it’s important to verify the information from multiple sources. Furthermore, my abilities are based on patterns learned from a diverse set of text data, but my knowledge cut off is from 2021, so it might not be the most recent or accurate. So, I would encourage your readers to think of me as a tool for generating ideas, rather than as a source of truth. And always use multiple sources to check the information. I’m here to help and make tasks easier, but it’s always good to use caution and critical thinking.

David:

As always, GPT, well said! Thank you so much for your time.

ChatGPT:

You’re welcome! It was my pleasure to be a part of this interview, and I’m glad that I could provide some useful information. If you have any other questions or need assistance with anything else, don’t hesitate to ask.