Since OpenAI introduced ChatGPT-3 in late November 2022, it seems everyone is talking about ChatGPT.
Earlier in the NamePros Blog, I ‘interviewed’ ChatGPT on domain topics. I think most will agree that ChatGPT did an impressive job on most answers.
Many investors are registering domain names in this sector, and a few have already sold. It is important to understand how GPT fits within the wider AI field, including limitations and possible applications, to help evaluate which names have most potential.
In this article, I take a deeper look at what GPT-3 is, and is not, and what it may mean for the domain name world.
ChatGPT and OpenAI
ChatGPT is one product from OpenAI. With interest in ChatGPT, OpenAI has rocketed to the 98th top website globally according to SimilarWeb.
OpenAI started as a capped-profit initiative, but has transformed into two parts, the non-profit OpenAI Incorporated, and for-profit, OpenAI Limited Partnership.
OpenAI was founded in 2015 by a number of investors who pledged $1 billion in startup funds. You can read more about OpenAI history here, and their charter at OpenAI Charter.
Sam Altman is the CEO of OpenAI. He had earlier served as President at Y Combinator.
On January 23, 2023 Microsoft and OpenAI announced an extension of their partnership. That includes a number of AI initiatives, and is expected to integrate ChatGPT into Microsoft products Bing, Word and PowerPoint.
ChatGPT was introduced as a free and open offering, although you do need to sign up for an OpenAI account. They have now introduced a ChatGPT Plus at $20 per month, with a restricted free version still available.
What Is GPT-3?
GPT is an example of Natural Language Processing, NLP, an active field since the 1950’s. The key idea of natural language processing is the system learns to interpret conversational prompts, and to reply in a way that mimics how a human would respond.
Over the last decade, most developments in NLP have used neural networks for machine learning.
The plan for the OpenAI GPT-3 implementation was outlined in a preprint document in 2020. You can freely download Language Models Are Few-Shot Learners, 75 pages long, by 31 authors from, or associated with, OpenAI.
GPT-3 was not well known to the general public until the November 30, 2022 release of ChatGPT by OpenAI, but it was in beta testing as early as July 2020.
GPT stands for Generative Pre-Trained Transformer. We explore below the meaning of each of the three terms.
Since many who use the term don’t relate it to the meaning of the acronym, GTP is a common typo of GPT.
What Was GPT-3 Trained On?
GPT-3 was pre-trained on a huge set of material. When you ask it a question, it is not searching for an answer, nor is it coded in the traditional sense. Rather, it mimics how someone might respond who had read almost everything, learned and internalized the information, and had developed the ability to generate text output.
So what was GPT-3 pre-trained on? That included essentially all of Wikipedia, but that only accounts for 3% of its training. About 6% is from the two huge sets of digitized book content. The biggest part, though, is from a filtered crawl of the online web, Common Crawl, represented 60% of the GPT-3 training material.
Read more in GPT-3 All You Need to Know About the AI Language Model.
How Does GPT-3 Work?
While the field of machine learning continues to evolve, the software is very good at knowing what terms come in what order, because it has learned from ‘reading’ so much what, statistically, are the most common sequences for words, and the implications of different placements.
That is why it writes so fluently - effective writers usually read a lot, and it has ‘read’ almost all of the digitized world. In a simplified fashion, generative means it does a great job at generating text, after prompts.
It has learned to key on prompts, to process the language more than just knowing what each word means out of context. It has learned language and response in much the way that infants and children do, or you learn a second language. It notices patterns, and through trial and error, during training, gets better at responding in the ‘correct’ manner.
What Is A Transformer?
The term transformer is not what you might associate with the term in toys, media, or electrical devices, but rather is a specific neural network term. Dale Markowitz wrote a good article: Transformers Explained: Understand the Model Behind GPT-3, BERT and T5.
An excellent technical, but still readable, introduction to the neural network transformer is What Is A Transformer Model?, written by Rick Merritt. That includes the origin of the term.
The neural net transformer idea was developed at University of Toronto and Google Labs primarily to do language translation, but now they can apply the same principles to interpret language. Their preprint is available free, and has the simple title Attention Is All You Need.
The idea of transformers in neural network implementations is that they come, through learning and attention, to understand the patterns of language. It turns out, compared to other ways to process language, the transformer method is both simpler and very well suited to parallel processing.
Just to show how important the transformer idea is, one estimate is that 70% of all AI preprint literature in the last few years has involved transformer ideas.
What Does GPT-3 Do Well That Could Help Domain Sellers?
I think most would agree that GPT-3 does a good job of ‘understanding’ most prompts, and writes fluent and grammatically correct output. Because it was trained on a huge amount of content, it can create a detailed answer on most topics.
It is not surprising that it is already very good at writing domain name descriptions. You can use SquadHelp, standard or premium listings, to see how effective AI-generated descriptions can be. I commented a couple of weeks ago that I have written a lot of domain name descriptions over the past five years, and, in most cases the SquadHelp AI implementation does a better job than I would in this task.
GPT-3 would also be well suited to writing other types of promotional material, such as website content, press releases, outbound emails, or social media posts. I have not tried it personally, but SquadHelp have a social media creation tool currently operational.
While one would need to human check prior to sending, GPT-3 implementations could possibly be as effective, or more effective, in replying to email inquiries on a domain name.
The whole point of a ChatBot is to mimic human speech interaction, and in many, not all, circumstances it does this pretty well. As such, it could, if trained on all documents at the marketplace, be very effective in answering questions, replacing FAQs of a registrar or domain marketplace with a more engaging style.
If trained on the specific inventory, not possible directly with ChatGPT which has been pre-trained, I think it could be a far superior system for searching the millions of domain names at a marketplace, and quickly leading potential buyers to the ones of greatest interest. SquadHelp have introduced this already, although still in a pretty simplistic fashion.
I think whichever company optimizes AI for inventory search and domain selection will have a significant competitive advantage.
Since GPT-3 is trained to know what sequence of words are natural together, it should be ideal at creating multi-word domain names. Because it is pre-trained, ChatGPT has no way of knowing which of those names are available, however.
It can generate results in what would be considered creative applications, such as writing poetry, lyrics or short stories. Expert views vary on the quality of these. Creative works might be effective in promoting certain types of names.
The Chat implementation of GPT-3 is generally fast at responding, and can give improved answers from additional clarification prompts.
What Can’t It Do Well?
The fact that it is pre-trained, and not actively connected to the Internet, means that, at least the current OpenAI ChatGPT formulation, fails at anything requiring near real-time information.
It makes mistakes, and should never be used without human supervision in any error-critical task. When asked to write research papers with references, it not only got some things wrong, but worse seemed to at times simply make up paper references. What it probably did was report references that somewhere in online world had been incorrectly reported. Similarly, when asked for top domain sales in a certain category, it gives some results not in NameBio and not clearly valid.
But to understand the most critical limitation of Chat GPT-3, it is important to stress again that it learned by looking at a huge collection of text. Along the way, it developed ability to key in on patterns within that text, and applied it to generate fluent response text. Young children learn language that way too.
But then, as youth and then adults, we also develop a second system, one built not just on recognizing patterns but developing abstract abilities, such as complex mathematical operations. If you ask math questions of ChatGPT it will get simple arithmetic right, since it has seen the question before, but will soon fail. Similarly, people have tricked it with logical deductions that most humans would get right.
GPT alone systems will fail at many tasks involving sophisticated logical and mathematical operations.
Will There Be GPT-4?
Yes, GPT-4 is already well in development, and plans are in place for GPT-5. While a specific release date has not been announced, a fair amount already seems clear about GPT-4 – see What We Know About GPT-4.
By the way, strictly speaking the ChatGPT was based on GPT-3.5, although it is simplified to GPT-3.
So What About Domain Names?
There has been a flood to register domain names containing the term GPT, often with chat as well. These numbers will already be dated when you read this, but the exact phrase GPT was, according to DotDB, registered in 164 extensions, not nearly as many as NFT or meta, but almost all are recent registrations. Also, it is part of more than 21,000 longer domain names, mainly in .com or .ai.
Here is a list of NameBio-recorded sales from the last two years that include the term ‘GPT’. Note that at least one sale on list includes GPT but is not a GPT domain sale. A few other sales have been reported on social media, and at time of article UnreportedSales showed 5 sales with ‘GPT’, 3 in 4-figures and one at $10,000.
Dofo Advanced Search indicated 670 domain names for sale including ‘GPT’, fewer than I would have thought. The number will probably be higher as you read this. Of those, 460 were .com.
@Centaur has started a NamePros Discussion Thread for GPT domain names.
Possible Domain Name Market for GPT
Let’s first start with who might be the market for domain names related to GPT, or AI more broadly. This is certainly not an exclusive list, but a few ideas:
Keep In Mind
Here are a few points to keep in mind.
On Feb 6, 2023 Sundar Pichai, Google CEO, announced AI integration with Google search, called Bard.
Their product is based on an alternative to GPT called Language Model for Dialogue Applications (LaMDA). LaMDA has been around for a few years, and you can read more about it in LaMDA: Our Breakthrough Conversation Technology.
Originally, the bard from Celtic culture, was a story teller, composer, oral historian, or musician, usually having a patron. It seems a flexible name, allowing for the integration of text dialogue, music and sound, images and video, all of which are about to be fundamentally changed through AI.
According to DotDB, the day I checked bard was registered in 176 TLDs, and in almost 60,000 longer names.
The
The term bard is part of the name, or also known as name, for over 2200 companies and organizations, according to OpenCorporates. Since the name is so related to Shakespeare, a number are theatre related, although almost every sector finds some use.
With the reach of Google, there is no doubt we will hear a lot about Bard in the weeks ahead.
A Future Scenario
Imagine a scenario in which you tick a preference box, and your registrar lists your domain name for sale on marketplaces.
But what does that leave for us to do?
More Information
Here are some articles that I found helpful in researching this topic:
Earlier in the NamePros Blog, I ‘interviewed’ ChatGPT on domain topics. I think most will agree that ChatGPT did an impressive job on most answers.
Many investors are registering domain names in this sector, and a few have already sold. It is important to understand how GPT fits within the wider AI field, including limitations and possible applications, to help evaluate which names have most potential.
In this article, I take a deeper look at what GPT-3 is, and is not, and what it may mean for the domain name world.
ChatGPT and OpenAI
ChatGPT is one product from OpenAI. With interest in ChatGPT, OpenAI has rocketed to the 98th top website globally according to SimilarWeb.
OpenAI started as a capped-profit initiative, but has transformed into two parts, the non-profit OpenAI Incorporated, and for-profit, OpenAI Limited Partnership.
OpenAI was founded in 2015 by a number of investors who pledged $1 billion in startup funds. You can read more about OpenAI history here, and their charter at OpenAI Charter.
Sam Altman is the CEO of OpenAI. He had earlier served as President at Y Combinator.
On January 23, 2023 Microsoft and OpenAI announced an extension of their partnership. That includes a number of AI initiatives, and is expected to integrate ChatGPT into Microsoft products Bing, Word and PowerPoint.
ChatGPT was introduced as a free and open offering, although you do need to sign up for an OpenAI account. They have now introduced a ChatGPT Plus at $20 per month, with a restricted free version still available.
What Is GPT-3?
GPT is an example of Natural Language Processing, NLP, an active field since the 1950’s. The key idea of natural language processing is the system learns to interpret conversational prompts, and to reply in a way that mimics how a human would respond.
Over the last decade, most developments in NLP have used neural networks for machine learning.
The plan for the OpenAI GPT-3 implementation was outlined in a preprint document in 2020. You can freely download Language Models Are Few-Shot Learners, 75 pages long, by 31 authors from, or associated with, OpenAI.
GPT-3 was not well known to the general public until the November 30, 2022 release of ChatGPT by OpenAI, but it was in beta testing as early as July 2020.
GPT stands for Generative Pre-Trained Transformer. We explore below the meaning of each of the three terms.
Since many who use the term don’t relate it to the meaning of the acronym, GTP is a common typo of GPT.
What Was GPT-3 Trained On?
GPT-3 was pre-trained on a huge set of material. When you ask it a question, it is not searching for an answer, nor is it coded in the traditional sense. Rather, it mimics how someone might respond who had read almost everything, learned and internalized the information, and had developed the ability to generate text output.
So what was GPT-3 pre-trained on? That included essentially all of Wikipedia, but that only accounts for 3% of its training. About 6% is from the two huge sets of digitized book content. The biggest part, though, is from a filtered crawl of the online web, Common Crawl, represented 60% of the GPT-3 training material.
Read more in GPT-3 All You Need to Know About the AI Language Model.
How Does GPT-3 Work?
While the field of machine learning continues to evolve, the software is very good at knowing what terms come in what order, because it has learned from ‘reading’ so much what, statistically, are the most common sequences for words, and the implications of different placements.
That is why it writes so fluently - effective writers usually read a lot, and it has ‘read’ almost all of the digitized world. In a simplified fashion, generative means it does a great job at generating text, after prompts.
It has learned to key on prompts, to process the language more than just knowing what each word means out of context. It has learned language and response in much the way that infants and children do, or you learn a second language. It notices patterns, and through trial and error, during training, gets better at responding in the ‘correct’ manner.
What Is A Transformer?
The term transformer is not what you might associate with the term in toys, media, or electrical devices, but rather is a specific neural network term. Dale Markowitz wrote a good article: Transformers Explained: Understand the Model Behind GPT-3, BERT and T5.
An excellent technical, but still readable, introduction to the neural network transformer is What Is A Transformer Model?, written by Rick Merritt. That includes the origin of the term.
The neural net transformer idea was developed at University of Toronto and Google Labs primarily to do language translation, but now they can apply the same principles to interpret language. Their preprint is available free, and has the simple title Attention Is All You Need.
The idea of transformers in neural network implementations is that they come, through learning and attention, to understand the patterns of language. It turns out, compared to other ways to process language, the transformer method is both simpler and very well suited to parallel processing.
Just to show how important the transformer idea is, one estimate is that 70% of all AI preprint literature in the last few years has involved transformer ideas.
What Does GPT-3 Do Well That Could Help Domain Sellers?
I think most would agree that GPT-3 does a good job of ‘understanding’ most prompts, and writes fluent and grammatically correct output. Because it was trained on a huge amount of content, it can create a detailed answer on most topics.
It is not surprising that it is already very good at writing domain name descriptions. You can use SquadHelp, standard or premium listings, to see how effective AI-generated descriptions can be. I commented a couple of weeks ago that I have written a lot of domain name descriptions over the past five years, and, in most cases the SquadHelp AI implementation does a better job than I would in this task.
GPT-3 would also be well suited to writing other types of promotional material, such as website content, press releases, outbound emails, or social media posts. I have not tried it personally, but SquadHelp have a social media creation tool currently operational.
While one would need to human check prior to sending, GPT-3 implementations could possibly be as effective, or more effective, in replying to email inquiries on a domain name.
The whole point of a ChatBot is to mimic human speech interaction, and in many, not all, circumstances it does this pretty well. As such, it could, if trained on all documents at the marketplace, be very effective in answering questions, replacing FAQs of a registrar or domain marketplace with a more engaging style.
If trained on the specific inventory, not possible directly with ChatGPT which has been pre-trained, I think it could be a far superior system for searching the millions of domain names at a marketplace, and quickly leading potential buyers to the ones of greatest interest. SquadHelp have introduced this already, although still in a pretty simplistic fashion.
I think whichever company optimizes AI for inventory search and domain selection will have a significant competitive advantage.
Since GPT-3 is trained to know what sequence of words are natural together, it should be ideal at creating multi-word domain names. Because it is pre-trained, ChatGPT has no way of knowing which of those names are available, however.
It can generate results in what would be considered creative applications, such as writing poetry, lyrics or short stories. Expert views vary on the quality of these. Creative works might be effective in promoting certain types of names.
The Chat implementation of GPT-3 is generally fast at responding, and can give improved answers from additional clarification prompts.
What Can’t It Do Well?
The fact that it is pre-trained, and not actively connected to the Internet, means that, at least the current OpenAI ChatGPT formulation, fails at anything requiring near real-time information.
It makes mistakes, and should never be used without human supervision in any error-critical task. When asked to write research papers with references, it not only got some things wrong, but worse seemed to at times simply make up paper references. What it probably did was report references that somewhere in online world had been incorrectly reported. Similarly, when asked for top domain sales in a certain category, it gives some results not in NameBio and not clearly valid.
But to understand the most critical limitation of Chat GPT-3, it is important to stress again that it learned by looking at a huge collection of text. Along the way, it developed ability to key in on patterns within that text, and applied it to generate fluent response text. Young children learn language that way too.
But then, as youth and then adults, we also develop a second system, one built not just on recognizing patterns but developing abstract abilities, such as complex mathematical operations. If you ask math questions of ChatGPT it will get simple arithmetic right, since it has seen the question before, but will soon fail. Similarly, people have tricked it with logical deductions that most humans would get right.
GPT alone systems will fail at many tasks involving sophisticated logical and mathematical operations.
Will There Be GPT-4?
Yes, GPT-4 is already well in development, and plans are in place for GPT-5. While a specific release date has not been announced, a fair amount already seems clear about GPT-4 – see What We Know About GPT-4.
By the way, strictly speaking the ChatGPT was based on GPT-3.5, although it is simplified to GPT-3.
So What About Domain Names?
There has been a flood to register domain names containing the term GPT, often with chat as well. These numbers will already be dated when you read this, but the exact phrase GPT was, according to DotDB, registered in 164 extensions, not nearly as many as NFT or meta, but almost all are recent registrations. Also, it is part of more than 21,000 longer domain names, mainly in .com or .ai.
Here is a list of NameBio-recorded sales from the last two years that include the term ‘GPT’. Note that at least one sale on list includes GPT but is not a GPT domain sale. A few other sales have been reported on social media, and at time of article UnreportedSales showed 5 sales with ‘GPT’, 3 in 4-figures and one at $10,000.
Dofo Advanced Search indicated 670 domain names for sale including ‘GPT’, fewer than I would have thought. The number will probably be higher as you read this. Of those, 460 were .com.
@Centaur has started a NamePros Discussion Thread for GPT domain names.
Possible Domain Name Market for GPT
Let’s first start with who might be the market for domain names related to GPT, or AI more broadly. This is certainly not an exclusive list, but a few ideas:
- Better prompts mean better results with ChatGPT. That means that the immediate market might be to solopreneurs who hope to cash in on the sudden interest by offering prompt services tailored to a specific sector.
- We will probably quickly see consultants and small agencies that will help clients get up and running using AI tools, including, but not restricted to, ChatGPT.
- What about monetized reference and information sites? There may be some market, probably over a short time horizon building on the current momentum.
- As Microsoft implement OpenAI products into Word and PowerPoint, there may be a mass-market opportunity for education and training.
- Are there new services only feasible with AI? That is, not simply applying AI to an existing service, but something entirely novel. If so, what are they? This is a harder segment of the market to predict, but potentially selling names as brands for higher amounts than the applications mentioned earlier.
- Will AI Chatbot become good enough, and trustworthy enough, to use in support, training and therapy roles? If so, the number of possible end use cases is very large.
- Educational institutions are already scrambling to prevent academic dishonesty from students misuse of tools like ChatGPT. There is probably a market there for a few businesses. OpenAI has introduced a way to detect AI-written text, and announced that more is on the way.
- GPT was first developed for language and translation, and it is natural that we will see many applications in sectors related to language: translation, editing, content generation, etc.
- There are a host of legal, equity, copyright and other issues associated with AI technology. Will there be opportunities for domain names related to those sectors?
- Clearly some great AI names, not necessarily with AI specifically in them, and definitely not with GPT, might sell for a lot. The most obvious extensions seem .com or .ai.
Keep In Mind
Here are a few points to keep in mind.
- We know that GPT-4 is in development, so names involving GPT-3 may have a short shelf-life.
- In all sectors, most businesses do not brand on the specific, generic term, but rather something that is more unique and more easily trademarked and defended.
- OpenAI has made noise about renaming ChatGPT. If that were to happen, domain names with GPT are suddenly less relevant.
- OpenAI filed a trademark application for the term ChatGPT on Dec 27, 2022. An apparently unrelated party also has a trademark application, dated Dec 15, 2022.
- Making a great chatbot interface in a free product has captured public attention. But longer term, most uses will probably not be operating through a chat interface.
- Some of the most exciting developments in AI are related to images, so while everyone is talking language AI now, that is just one part. You can read about OpenAI DALL-E 2 project here.
- As I was completing this article, Microsoft announced AI integration into Bing search. This will further increase public awareness and appreciation of AI, but at same time, as Microsoft implements features into Bing and Office components, some of the market for stand-alone applications will dry up.
- Those investing in this area should delve at some depth into the topic, so created names are relevant. Invest in what you know, and know what you invest in, is usually good advice.
On Feb 6, 2023 Sundar Pichai, Google CEO, announced AI integration with Google search, called Bard.
Their product is based on an alternative to GPT called Language Model for Dialogue Applications (LaMDA). LaMDA has been around for a few years, and you can read more about it in LaMDA: Our Breakthrough Conversation Technology.
Originally, the bard from Celtic culture, was a story teller, composer, oral historian, or musician, usually having a patron. It seems a flexible name, allowing for the integration of text dialogue, music and sound, images and video, all of which are about to be fundamentally changed through AI.
According to DotDB, the day I checked bard was registered in 176 TLDs, and in almost 60,000 longer names.
The
bard.com name is used by Becton, Dickinson and Company, while the .ai and .xyz are both currently for sale.The term bard is part of the name, or also known as name, for over 2200 companies and organizations, according to OpenCorporates. Since the name is so related to Shakespeare, a number are theatre related, although almost every sector finds some use.
With the reach of Google, there is no doubt we will hear a lot about Bard in the weeks ahead.
A Future Scenario
Imagine a scenario in which you tick a preference box, and your registrar lists your domain name for sale on marketplaces.
- AI automatically generates categories and potential uses for each domain name. It also checks for trademark issues.
- AI assesses various characteristics, and determines retail pricing.
- AI automatically generates a logo for the domain listing, and writes an eloquent description.
- Perhaps it will generate still images or a video presentation specific to that name. For example, call for an AI-generated image of a couple beside a vintage car in front of a sandy beach.
- AI will interact with visitors to the marketplaces, and through conversational dialog quickly guide them to just what they are looking for.
- If you have the box ticked for outbound, it will research potential buyers, generate an email outreach, and communicate with them.
- AI will, following your directions, generate a list of potential acquisitions, determine their wholesale worth, and bid in auctions on your behalf.
But what does that leave for us to do?
More Information
Here are some articles that I found helpful in researching this topic:
- Wikipedia Article on GPT-3.
- An excellent article which does a great job of explaining how it compares to previous efforts is at Sigmoid: GPT-3 All You Need to Know About the AI Language Model.
- A very readable outline is GPT-3.
- Everything We Know About GPT-4
- You can’t truly understand GPT without understanding the idea of a neural network transformer, and Dale Markowitz’s article, Transformers Explained: Understand the Model Behind GPT-3, BERT and T5 is really helpful.
- Also on the neural network transformer is What Is A Transformer Model? by Rick Merritt.
- If wondering how AI can transform search, Microsoft explains the AI integration in Bing here, along with some interesting examples.
- I recently attended a great lecture by Dr. Dan Lametti on AI and Academia: The End of the Essay? Despite the title, most of the lecture was a superb introduction to how ChatGPT works. It is available on YouTube.
Last edited:
