GUEST POST from Geoffrey A. Moore
I recently finished reading Stephen Wolfram’s very approachable introduction to ChatGPT, What is ChatGPT Doing . . . And Why Does It Work?, and I encourage you to do the same. It has sparked a number of thoughts that I want to share in this post.
First, if I have understood Wolfram correctly, what ChatGPT does can be summarized as follows:
- Ingest an enormous corpus of text from every available digitized source.
- While so doing, assign to each unique word a unique identifier, a number that will serve as a token to represent that word.
- Within the confines of each text, record the location of every token relative to every other token.
- Using just these two elements—token and location—determine for every word in the entire corpus the probability of it being adjacent to, or in the vicinity of, every other word.
- Feed these probabilities into a neural network to cluster words and build a map of relationships.
- Leveraging this map, given any string of words as a prompt, use the neural network to predict the next word (just like AutoCorrect).
- Based on feedback from so doing, adjust the internal parameters of the neural network to improve its performance.
- As performance improves, extend the reach of prediction from the next word to the next phrase, then to the next clause, the next sentence, the next paragraph, and so on, improving performance at each stage by using feedback to further adjust its internal parameters.
- Based on all of the above, generate text responses to user questions and prompts that reviewers agree are appropriate and useful.
OK, I concede this is a radical oversimplification, but for the purposes of this post, I do not think I am misrepresenting what is going on, specifically when it comes to making what I think is the most important point to register when it comes to understanding ChatGPT. That point is a simple one. ChatGPT has no idea what it is talking about.
Indeed, ChatGPT has no ideas of any kind—no knowledge or expertise—because it has no semantic information. It is all math. Math has been used to strip words of their meaning, and that meaning is not restored until a reader or user engages with the output to do so, using their own brain, not ChatGPT’s. ChatGPT is operating entirely on form and not a whit on content. By processing the entirety of its corpus, it can generate the most probable sequence of words that correlates with the input prompt it had been fed. Additionally, it can modify that sequence based on subsequent interactions with an end user. As human beings participating in that interaction, we process these interactions as a natural language conversation with an intelligent agent, but that is not what is happening at all. ChatGPT is using our prompts to initiate a mathematical exercise using tokens and locations as its sole variables.
OK, so what? I mean, if it works, isn’t that all that matters? Not really. Here are some key concerns.
First, and most importantly, ChatGPT cannot be expected to be self-governing when it comes to content. It has no knowledge of content. So, whatever guardrails one has in mind would have to be put in place either before the data gets into ChatGPT or afterward to intercept its answers prior to passing them along to users. The latter approach, however, would defeat the whole purpose of using it in the first place by undermining one of ChatGPT’s most attractive attributes—namely, its extraordinary scalability. So, if guardrails are required, they need to be put in place at the input end of the funnel, not the output end. That is, by restricting the datasets to trustworthy sources, one can ensure that the output will be trustworthy, or at least not malicious. Fortunately, this is a practical solution for a reasonably large set of use cases. To be fair, reducing the size of the input dataset diminishes the number of examples ChatGPT can draw upon, so its output is likely to be a little less polished from a rhetorical point of view. Still, for many use cases, this is a small price to pay.
Second, we need to stop thinking of ChatGPT as artificial intelligence. It creates the illusion of intelligence, but it has no semantic component. It is all form and no content. It is a like a spider that can spin an amazing web, but it has no knowledge of what it is doing. As a consequence, while its artifacts have authority, based on their roots in authoritative texts in the data corpus validated by an extraordinary amount of cross-checking computing, the engine itself has none. ChatGPT is a vehicle for transmitting the wisdom of crowds, but it has no wisdom itself.
Third, we need to fully appreciate why interacting with ChatGPT is so seductive. To do so, understand that because it constructs its replies based solely on formal properties, it is selecting for rhetoric, not logic. It is delivering the optimal rhetorical answer to your prompt, not the most expert one. It is the one that is the most popular, not the one that is the most profound. In short, it has a great bedside manner, and that is why we feel so comfortable engaging with it.
Now, given all of the above, it is clear that for any form of user support services, ChatGPT is nothing less than a godsend, especially where people need help learning how to do something. It is the most patient of teachers, and it is incredibly well-informed. As such, it can revolutionize technical support, patient care, claims processing, social services, language learning, and a host of other disciplines where users are engaging with a technical corpus of information or a system of regulated procedures. In all such domains, enterprises should pursue its deployment as fast as possible.
Conversely, wherever ambiguity is paramount, wherever judgment is required, or wherever moral values are at stake, one must not expect ChatGPT to be the final arbiter. That is simply not what it is designed to do. It can be an input, but it cannot be trusted to be the final output.
That’s what I think. What do you think?
Image Credit: Pixabay
Sign up here to join 17,000+ leaders getting Human-Centered Change & Innovation Weekly delivered to their inbox every week.