Lessons learned from ChatGPT about marketing new technologies

This article attempts to unpick what has made ChatGPT so successful in marketing itself as an AI language model. As part of an extended conversation, I ask ChatGPT about its own marketing strategy.

Interviewer ?

Let us consider another question pertinent to responsibilities as a CTO — that of marketing and publicising technical innovations.

ChatGPT ?

As a CTO, marketing and publicizing technical innovations is an important responsibility...

. . .

Overall, ChatGPT can be a valuable tool for CTOs looking to market and publicize technical innovations in a way that is clear, engaging, and easy for non-technical audiences to understand.

Interviewer ?

That is undoubtedly true. However, I am most interested in the extent to which ChatGPT’s success is fundamentally the skill with which this model has been publicised. In particular, the human-like conversation of ChatGPT seems like an exceptionally powerful way to market the innovation which is mainly inherent in the underlying GPT models.

ChatGPT ?

You are correct that the success of ChatGPT, like any technology, is also dependent on how well it is marketed and publicised. One of the key features of ChatGPT that makes it particularly well-suited for marketing and publicizing technical innovations is its ability to generate human-like conversation.

. . .

In summary, ChatGPT’s ability to generate human-like conversation makes it an effective tool for marketing and publicizing technical innovations, as it can help to create clear, engaging and easy-to-understand content that highlights the benefits of the product or service, and can be used to generate interactive demos, chatbots and virtual assistants that can help educate customers and answer their questions in a natural way, helping to build trust and interest in the product or service.

As expected, ChatGPT presents itself as either an assistant to producing marketing materials, or as an augmented interface providing information to clients. It also reflects on the importance of human-like interaction in promoting itself as a technology.

What distinguishes ChatGPT from GPT3?

ChatGPT is a new member of the GPT3 family of models and products. The underlying GPT3 technology was well received. But, ChatGPT has been far more successful. Why the difference?

To understand the marketing technique behind ChatGPT, we must first understand the relationship between the various offerings.

Interviewer ?

Which of ChatGPT’s capabilities are available directly from the underlying GPT model, rather than the find-tuning inherent to ChatGPT?

ChatGPT ?

ChatGPT is a variant of the GPT (Generative Pre-trained Transformer) model, which is an autoregressive language model that has been pre-trained on a large dataset of text from the internet. The main capabilities of ChatGPT are inherited from the underlying GPT model and are available directly from it.

Some of the core capabilities of GPT that ChatGPT inherits include:

1. Language understanding:?GPT is trained to understand natural language inputs and respond in a human-like way. This allows it to understand and answer a wide range of questions in a natural language format.

2.?Language generation:?GPT is able to generate high-quality, natural-sounding language responses to a wide range of questions. This allows it to create engaging and informative content that highlights the benefits of a product or service.

3.?Text completion:?GPT is able to complete a given text prompt with high quality and coherent text based on the patterns it has learned from the training data.

4.?Language translation:?GPT can be fine-tuned to perform language translation tasks.

5.?Language summarization:?GPT can be fine-tuned to perform text summarization tasks.

ChatGPT builds on these core capabilities by fine-tuning the model on a specific task such as conversational language understanding and generation, which allows it to be more effective in generating human-like responses and handling the conversational context.

It’s important to note that while ChatGPT is fine-tuned for a specific task, it still inherits all the capabilities from the underlying GPT model and can be used for other tasks as well, depending on the fine-tuning process and the data it’s trained on.

Seeing the abilities of GPT, one wonders what ChatGPT is actually adding. It notes the augmentation of handling conversational context, and of generating human-like responses. Both of these aspects are focused on interaction rather than core language skills.

The power of the underlying GPT model means that ChatGPT’s extra abilities may not always be necessary. If so, why has ChatGPT been so much more successful in generating market awareness?

Interviewer ?

When would it make sense to use a GPT model rather than ChatGPT?

ChatGPT ?

It may make sense to use a GPT model rather than ChatGPT depending on the specific use case and the desired outcome.

GPT models can be useful in a wide range of tasks that involve understanding and generating natural language, such as:

1. Text generation:?GPT can be used to generate high-quality, coherent and natural-sounding text in various forms like articles, stories, poems, etc.

2. Text completion:?GPT can be used to complete a given text prompt with high-quality and coherent text based on the patterns it has learned from the training data.

3. Language translation:?GPT can be fine-tuned to perform language translation tasks.

4. Language summarization:?GPT can be fine-tuned to perform text summarization tasks.

5. Language understanding:?GPT can be fine-tuned to understand natural language inputs and generate appropriate responses.

Additionally, GPT models are not restricted to conversational context and can be trained on any type of text-based data, so it can be used for tasks such as text classification, sentiment analysis, and more, whereas ChatGPT is mainly fine-tuned for the conversational context.

In summary, if the task at hand is not conversational in nature, or if the main focus is on text generation, text completion or other natural language tasks that don’t require human-like conversation, then it might make more sense to use a GPT model rather than ChatGPT.

Although, not specifically answering our question, the key differentiator is clear. The conversational interface makes ChatGPT a more compelling marketing story.

Interviewer ?

So, as a CTO, what is the main takeaway in terms of marketing our technical capabilities and innovations? Can we generalise the strategy used so successfully to market GPT via ChatGPT?

ChatGPT ?

As a CTO, one of the main takeaways when it comes to marketing your technical capabilities and innovations is to use clear, concise, and engaging language that highlights the benefits of your products and services, and that is easy for non-technical audiences to understand.

. . .

It’s also important to note that the success of GPT models like ChatGPT in marketing and publicizing technical innovations is not only due to the technology itself, but also the skillful communication and storytelling that allows the audience to understand the technical capabilities and benefits of the technology...

. . .

ChatGPT latches on to a key takeaway in the third paragraph. “Skillful communication and storytelling” is key to what works in ChatGPT’s favour. The conversational interface encourages us to involve GPT in our everyday stories. This novel level of emotional engagement differentiates ChatGPT as a powerful marketing tool.

Is ChatGPT trained specifically for marketing purposes?

It’s worth considering which of ChatGPT’s human-like qualities are inherent to its understanding of language. Which qualities are inherent to the model, and which are deliberately crafted to make a compelling product?

ChatGPT’s conversational interface is deliberately trained to appear human. Does it merely imitate, or does it truly exhibit human-like attributes (such as correcting mistakes)? Is this due to the model’s complexity, or is it merely designed to drive engagement?

…You could use the API provided by OpenAI to generate the responses,?and the?store the conversation logs on your own backend…

Interviewer ?

I note that in the first sentence of the second paragraph of your previous response, you used an apparently grammatically incorrect sentence. The error appears to be starting the second half of the sentence with “and the” instead of “and then”. Is this a side-effect of the training?

ChatGPT ?

I apologize for the error in my previous response. As a language model, I am trained on a large corpus of text, and sometimes mistakes can slip through despite my best efforts to avoid them. The error you pointed out is a mistake on my part as a language model and it should have been “and then store the conversation logs on your own backend” instead of “and the store the conversation logs on your own backend”.

ChatGPT acknowledges a mistake. Mathematical functions don’t make mistakes. What is going on here?

Interviewer ?

What enables ChatGPT to analyse its own responses in order to verify an error? Computers generally produce the same output given a given starting point.

ChatGPT ?

As a language model, ChatGPT is trained on a large corpus of text, and it has been trained to generate text that is coherent and grammatically correct to the best of its abilities. However, since it’s a machine, it can make mistakes, just like any other computer program.

When I analyze my own responses, I am not “thinking” or “verifying” in the way that humans do. Instead, I am using the rules and patterns that I have learned during my training to identify errors in my output. The errors are identified by comparing the generated text to the training data and looking for patterns that do not match, or by using pre-defined grammar checker or spell checker.

It’s also worth noting that the ability to identify errors in my own responses is limited by the quality and diversity of the data that I was trained on, so it’s possible that I may miss some errors or produce incorrect outputs.

In summary, ChatGPT is not analyzing its own responses in the way that a human would, but it uses the rules and patterns that it has learned during its training to identify errors in its output. But it’s also important to keep in mind that my ability to identify errors is limited by the quality and diversity of the data that I was trained on.

So, ChatGPT appears to analyse the entire conversation as input along with each prompt. This recursive aspect of processing makes sense. Many other disciplines (e.g.?Chaos Theory,?Dynamical Systems) derive from feedback loops.

But, is the mistake more akin to a spelling error, or to language parsing?

Interviewer ?

In this example, did you mis-spell a word (omitting the “n” at the end of the word), or did you just match a different sentence structure?

ChatGPT ?

In the sentence “and the store the conversation logs on your own backend” the mistake is that I used “and the” instead of “and then” which is a grammatical error that is caused by the way the model combines different pieces of text from its training data. I apologize for any confusion caused by my previous response.

The response is inconclusive. But, the type of mistake seems to show that the pattern combination is complex. This is an unusual mode of failure for a traditional computer program. We expect a word, if understood at all, to be a fixed token identifying a semantic concept. But here ChatGPT appears to be using words in a more advanced way. It may leverage the statistical likelihood of a word being present at a given point in the context of all the other words.

ChatGPT sits behind a cleverly designed fa?ade

So, truly advanced conversational abilities make ChatGPT technically compelling. But other parts of product design also contribute. I ask ChatGPT about the way in which the UI reveals text (one word at a time). Why would a language model reveal one word at a time?

Interviewer ?

I also note that you reveal your responses one word at a time, as though a human were spelling out the response one letter at a time on a keyboard. Is this merely an effect to make the chat more accessible, or is it part of how you work?

ChatGPT ?

The way my responses are revealed one word at a time is not a side effect of how I work, but it’s a choice made by the developers of the API or platform you’re using to interact with me. The reason behind this choice is to mimic the way humans type which makes the interaction more natural and similar to a human-to-human conversation. This can improve the overall user experience, as well as making it more accessible for users with different abilities. But please note that it does not affect my ability to process and generate responses.

Interviewer ?

Can you respond to me with the entire response all at once for the rest of this conversation?

ChatGPT ?

Sure, I can respond to you with the entire response all at once for the rest of this conversation. However, please keep in mind that receiving a response all at once may not be as natural as receiving it one character at a time, and it may affect the overall user experience.

Note — ChatGPT agreed to stop delivering conversation one word at a time, but it continues to do so for the rest of the interview. This illustrates a fracture between the UI product design and the underlying model.

Conclusion

ChatGPT is a masterclass in marketing. As a technology, the conversational abilities of ChatGPT make it the perfect vehicle for marketing GPT.

It is designed to be compelling to human beings. The UI makes the interface approachable to almost anyone. OpenAI drive engagement to the model through design decisions such as:

1. Increasing the sense of intimacy by revealing conversation one word at a time
2. Flourishes such as auto-generating titles for each conversation

As technical leaders, this technology is a powerful use case about how to market innovation. Well designed products can generate their own engagement. But, massive investment in making a technology engaging and approachable pays off. If we want to take technology to the market, engaging with emotions and narrative is likely non-negotiable.

This article is part of a series. The original article, and links to the series can be found here: https://tinycode2.medium.com/lessons-learned-from-chatgpt-about-marketing-new-technologies-ff9f0b0bfe74?source=friends_link&sk=2acaf142b67b79b5c072e2736667a71e