​ChatGPT - Hype or Breakthrough?

Lots of hype and excitement erupted with the release of OpenAI’s ChatGPT just a few months ago.  It can write a play in the style of Shakespeare.  College students can generate essays in a matter of minutes, freeing precious time for drinking beer.  A variant can create software code at lightening speed.  ChatGPT looked set to revolutionize the entire AI-space.

But that was soon followed by dismay at the litany of mistakes, limitations, and ‘hallucinations’ resulting in unreal conversations experienced by public testers.  There was the text chat with a journalist in which ChatGPT proposed love.  There were the examples of ChatGPT getting factual answers wrong and not fessing up to the mistake.

Now that it has been a few months, it’s time to take a step back from the excitement and the controversy for an attempt at sober reflection.

Definition

What is ChatGPT?GPT stands for Generative Pre-trained Transformer (GPT) which is a neural network form of machine learning that has been trained on vast quantities of text data to be able to generate sophisticated output responses with just a little input text.  The existing model reportedly has 175 billion parameters built in.  ChatGPT is capable of creating not just human language text, but also programming code.

Training the model is a labor intensive process in which vast quantities of text data are ingested from major internet sources (Wikipedia, etc.).  There follows a reinforcement training phase in which human engineers ask questions, the model responds, and the engineers rate the quality of the response with a score which helps reinforce/improve the model.

The model then uses this vast library of text and reinforced training parameters on what a good response is to generate new text in response to a question.

What all of this means is that ChatGPT is incredibly powerful and a genuine-seeming breakthrough in creating real-time, life-like conversations or written text outputs.

There are, however, some caveats.

The model is pre-trained which means it does not learn from each interaction it has with a consumer or customer.  It only learns when responses are scored during a supervised training phase and then formally upgraded into the model.

Because the model is a neural network with many layers of processing that are generated by the model itself in an iterative fashion (self-learning), it is very difficult for engineers to explain how and why specific inputs result in specific outputs.

When trained on the vast corpus of human created content, the model is at risk of reinforcing bias.  A play by Shakespeare and a manifesto from the Unibomber are both bodies of text.  ChatGPT has all of it in its training data.  Humans can remove known example of bad text, but there are simply too many intolerant screeds on the internet to do this for everything.   Therefore, ChatGPT is at risk of drawing on this type of material and generating a response that is highly biased when it is creating a text response.

Cost

It takes a lot of resources to create a model this powerful. Estimates are $500 million to a billion.  That means large models like this will be mostly big undertakings by large organizations with significant resources rather than hatched in a garage by a small 2-pizza team anytime soon.

Microsoft has invested billions in OpenAI and sees a vast potential to commercialize and incorporate ChatGPT into its own products.  They are probably not wrong.  

But the real question revolves around the operating model and development costs for ChatGPT.  If Microsoft and OpenAI are able to use ChatGPT has a foundational model with modest upkeep and maintenance into which domain specific cartridges can be plugged for specific business outcomes to create a range of products, then this could be a very profitable business for both OpenAI and Microsoft.

But that's not the only avenue for innovation.  

Innovation

The model is considered a foundational model. The underlying model can be transferred and used with domain specific data to create a more focused tool.

And that means there is potentially opportunity for those 2-pizza team startup to create disruptive new businesses.  Right now, ChatGPT has inspired a rush of startups that are plugging into OpenAI's API and starting to experiment with how to use ChatGPT to create new businesses.  

We'll soon start to see whether commercial business models can be created on top of ChatGPT and how disruptive this technology really is.    

Competition

If ChatGPT is a wonderful, unique, incredibly powerful tool that corners the market, then OpenAI and Microsoft can charge a premium and recoup all of those development costs.

But that is probably not the case.  Competition is going to pour into this field.  There will be more models that enter the race from large, well-financed competitors.  Google.  Meta.  Apple.  And more.  After all, the basic approach is pretty well understood.  Other firms will offer a similar capability.  Competition will drive down what can be charged if not the underlying cost.

Testing Runs

I have an OpenAI account and have done some basic test runs to generate ChatGPT outputs.  I prompted ChatGTP for 1) An essay comparing the similarities and differences of Elon Musk and Jeff Bezos' approaches to space colonization, 2) requested a 2-page synopsis for a mystery thriller set in Brazil involving an investigation into an eco-crime (a novel I've written), and 3) asked it to generate parental advice for a teenager recently diagnosed with a medical condition.

In each of these cases, ChatGPT came back within moments with a well-thought out, coherent and logically structured essay, synopsis, or recommendation which was factually correct.  These outputs each had a few elements that seemed original and interesting and reflected a point of view.   

But they were also formulaic.  They quickly generated basic content.  They helped me see another perspective in each case, whether fiction or nonfiction.  But they didn't astonish me and they didn't make me feel that there was a lot of creativity in any of the responses.  It was pretty much what you would expect.  It summarized the generally acceptable responses based on thousands, if not millions, of similar products.    

Summary

ChatGPT doesn’t have a brain or a soul.  It has a brute-force model with nearly 200 billion parameters to help guide it in generating text outputs.  Many of these outputs are quite impressive.  Future versions and competitors will have trillions or even hundreds of trillions of parameters.  These models have vast stores of text from a vast number of sources - more than any human could ever ingest in 10,000 lifetimes.  They are capable of finding and creating new content based on these relationships.  This is a lot of power.

But these models still don’t have judgment, morals, or a conscience.  They can’t tell right from wrong.  They don’t have empathy.  They can’t independently understand the context of the question, only a few words entered as a request.  They will need guardrails and constraints.  They will need maintenance and monitoring.  They will create great stuff and awful stuff - but potentially more devastating - they may just create a lot of 'good, average' stuff.

For many routine, specific activities, this tool is going to be powerful and disruptive.  Woe to the software coder.  Apologies to the news copy writer.  Not so good for call center workers and the back-end accounts receivable clerks.  Anything that follows a template and is easy to reproduce will be automated fast.  Basic customer chat scenarios with clear cut one-dimensional responses will fall quickly to ChatGPT or similar products.

And, yet, what ChatGPT seems to represent does not yet seem like a full revolution, but the acceleration of an ongoing process of change that is going to disrupt and amplify existing trends.  It's a cool new tool.  It does some cool new tricks.

It will, however, inspire a vast wave of innovation as new startups begin testing thousands of possible use cases in search of profitable business models.  New products will get created.  Some of these will not have been possible before.

As this process unfolds, we will start to get a better answer on whether this is a revolution, a new general purpose technology, or just another step in a longer journey.  Either way, it is going to be exciting.