ChatGPT Will Change Everything
At this point, nearly everyone has written about ChatGPT, but I figured I’d offer a short take of my own. There is surely a much longer story to be told, but for those that may be newer to this tool, this serves as a brief introduction.
What is ChatGPT?
ChatGPT Link: https://openai.com
ChatGPT is a tool built on the GPT3 (Generative Pre-trained Transformer 3), which is a language generation model (machine learning neural network) built and trained by OpenAI. The model has 175 Billion parameters, which is an incredible size and at a practical level, means it probably costs quite a lot of energy in computation for each question it answers.
The input takes common language text questions and will deliver answers in conversational text that’s pretty much indistinguishable from a human respondent, which is what sets it apart from other AI models that have made their way into the world so far. It works by taking your question as input to the neural net and augments it by utilizing the context around your question to generate the most probabilistically appropriate response based on its training.
The user interface is a basic chat window that has text input and text output. If you didn’t know what it was when you logged on to the site, you’d probably think you were chatting with a professional support person. The responses don’t have a lot of personality, but they are well written, comprehensive and accurate the vast majority of the time.
Where Do We Go From Here?
The tool is still just in beta so it’s still being purposely constrained in its application, but its real potential is in the ability to integrate the model into some domain specific business or societal use. Imagine the use of a model this sophisticated with a more carefully guided set of training data?
Someone will surely figure out how to hone those models for specific fields and build businesses on top of it that have incredible value in specialized domains. What if you were to use the internal financial records and all public competitive information from a real business to train the network to answer questions about what competitive moves would be most economically beneficial for the company? Could this be your corporate strategy team?
As we’re all probably aware, there has been significant investment over the last decade in sophisticated ML models in many domains. Financial trading firms have been at the forefront of this trend for a long time, hiring physicists and programmers in an attempt to build models that can beat the market. So, does GPT3 improve on these types of use cases? My guess is probably not given the specialty of this model is language generation, but who knows.
There is certainly a major opportunity in accelerating the productivity in technical business domains. As we get further along in our careers, the availability of some foundational knowledge we learned in school may fade. We know the ideas conceptually and functionally, but we don’t recall the specific derivations or mechanics. This becomes a wonderful tool to dig up these details to assist technical roles to deliver more comprehensive value to their organizations as they age through their careers. It’s a heck of a lot simpler than digging through our old textbooks (if we even have them) and if we have a follow-on question, we can interact with ChatGPT to find it.
Another major benefit is its ability to distill and produce a response to a question that you may typically search on Google for. For example, when I asked ChatGPT to write me some Python code to plot a particular mathematical function, it produced a completely functional script including the use of appropriate package imports, code documentation and use of package-imported functions. In this case, ChatGPT is saving me a huge amount of time searching just the right question, clicking on a bunch of forums like StackOverflow and reading all the responses to questions that are almost the same as mine, but not quite. And the best part is that if the response is not quite what I was looking for, the prior response becomes the baseline context for both my follow-on query as well as the next response. “No, what I was really looking for is a green dashed plot line, not red.” This type of ongoing discussion works very well in this tool and acts to further hone the response to be exactly what you are looking for.
Taking this a step further, you can imagine ChatGPT as an incredible tool for assisted learning, effectively acting as an interactive tutor. When you don’t understand something, you can ask. When the response still doesn’t make sense, you can ask for clarification and so on. This tool can meet you where you are in your learning state. It’s a level of EQ that’s better than many human tutors since they often only understand concepts in particular ways and aren’t able to step back to first principles and draw you another path to the conclusion.
This assisted learning option is what I immediately gravitated toward as I found out about this tool. I am an infinitely curious person and I have an interest to learn about so many different things that I naturally need someone (or some thing) to help me traverse the domain of knowledge. ChatGPT can be just that.
On The Other Hand
There are detractors, of course. I listen to Cara Swisher and Scott Galloway on the Pivot podcast, and Cara’s take on ChatGPT is very different than most. She claims the output is bad quality writing and we should stop highlighting it. While I do like these two and their podcast, this seems like the classic view from someone who may feel their livelihood is threatened. While that’s understandable, I think her take is misguided. I think ChatGPT will only serve to elevate the value of high quality journalism and writing. As humans we want connection. We want to know who wrote the opinion, not just what the opinion is. If we’re looking for facts, then getting a textbook answer seems appropriate, and what’s wrong with that?
What about errors? The model obviously can’t be correct in every response (I received one incorrect one already), so should we care? While we should probably be more impressed that it gets the vast majority of answers right, which is certainly a feat, the risks associated with getting erroneous responses that the user is considering as fact, is fairly high. It’s a statistically small quantity of errors, but if the assumption is accuracy, then the impact of errors is nonlinear.
So, how do we mitigate the inclusion of erroneous responses? I definitely don’t know the answer to that, but as more people rely on AI models like this to generate thoughts in their stead, the more garbage there will be online, which is the depot of knowledge from which these tools consume. Unmanaged, this could spiral into a negative feedback loop that renders the tools useless. I don’t believe that’s likely to happen, but some care should be taken to mitigate such outcomes.
Final Thought
To me, ChatGPT and the GPT model more broadly is a tool that will multiply the power and intelligence in a human or team by making a massive amount of knowledge immediately accessible to them. This most certainly does not obviate the intelligence of the humans that utilize this tool, it only enhances it. Those that benefit the most will be those with the best questions.
-Chris