ChatGPT: Implications for Business

ChatGPT, a new service from Open AI, has taken the internet by storm with its ability to generate natural language in impressive demonstrations that we have not seen before. ChatGPT can write essays, create research papers and even solve some math problems. Open AI has racked up more than 1 million users in just five days after the release. With this unprecedented response, it is worth taking time to understand what it is, what it’s good for and where it should potentially be avoided.

What is it?

ChatGPT is a large language model. Advances in large language models have led to an explosion of new capabilities in natural language processing (NLP) in the last 5 years. All machine learning models, learn from ground truth data (data we know to be true). It can be slow and expensive to have experts assign labels to enough data to train a machine learning model. A pivotal change occurred in NLP with the idea of masked language model training. With this technique, we take known sentences, hide a few words then build a model to try and predict the masked word.

I am going to New York, and I am going to see a ________

What is the blank?: show, play, musical … Yankees game

A language model is able to learn that any of these options are viable solutions for the missing word. With enough examples, the language model can learn the word or phrase that will have the highest probability of correctly filling the blank. Also, from the way that the words (show, play, musical) are used, in other text, the model can learn that they are synonyms. Whereas the Yankees are used in different contexts and different characteristics are learned. The Yankees and musicals are both associated with New York, but musicals are synonymous with plays while Yankees are associated to baseball.

When we use a masked language model training, all written words become our ground truth. We can apply this to all the articles in Wikipedia, the New York Times and product reviews on Amazon. Large language models trained on a virtually infinite amount of text with more than 1 trillion trainable parameters are able to do some truly impressive things.

The language model has learned more than the definitions of the words, it has memorized certain facts about people and places.

What has ChatGPT done differently?

ChatGPT is not the first language model and it’s not the largest language model. The significant innovation of ChatGPT is the application of human in the loop reinforcement learning to the output of its earlier language model named GPT-3. Prior models have produced impressive results, but the language sounded a little clunky, like a robot. Open AI showed multiple model responses to people and had them rank the responses in order of their perceived appropriateness. The model gradually learned from the responses; the model learned what characteristics of language people preferred to see. The combination of learning from all the text on the internet, then fine tuning with learned human preference has produced results that are more fluent than any prior work in the field.

What is it good for?

ChatGPT is going to produce well written language in a logical format.

This example is not specific enough to the conditions of a particular company to be used as-is, but it definitely covers the topics in a logical order that an enterprise ai program may want to consider as part of their plans. It is a useful jumpstart to a well written plan.

So, if you need to create a document and you can do so without sending proprietary information in the prompt, use of ChatGPT as a writing assistant can be productive. Ethan Mollick, a professor at the University of Pennsylvania’s Wharton School of Business, told NPR in an interview on Morning Edition, that He’s used it as his own teacher’s assistant, for help with crafting a syllabus, a lecture, an assignment and a grading rubric for MBA students. Similar to how we use Google today, use of AI assistants to multiply personal productivity is a trend that is only going to increase.

ChatGPT only runs on servers hosted by Open AI; So, any data that you use to interact with the system is recorded on their servers. This is a pretty severe limitation to what we can do with ChatGPT in a corporate setting. For example, it would be interesting to send customer service questions to ChatGPT to see how the service would answer, but it would not be appropriate to put our proprietary data on Open AI servers.

Risk of False Positive

ChatGPT is designed to mimic human writing and it has learned some facts about the world. However, it will still produce confident sounding responses that are completely false. ChatGPT is unlikely to produce an ‘I don’t know’ answer. It is more likely to produce answers that sound convincingly plausible.

For example, Teresa Kubacka, a data scientist based in Zurich, Switzerland, tested the tool by asking it about a made-up physical phenomenon. “I deliberately asked it about something that that I know doesn’t exist so that I could judge whether it actually also has the notion of what exists and what doesn’t exist,” she said. ChatGPT produced an answer so specific and plausible sounding, backed with citations, that she had to investigate whether the fake phenomenon, “a cycloidal inverted electromagnon,” was actually real. ChatGPT produced names of well-known physics experts but the titles of the publications they supposedly authored, did not exist.

So, at this stage, it is important to keep a human in the loop for any ChatGPT use and be willing the check the facts produced by its responses.

Accelerated Learning

Another good use case for ChatGPT is learning about a new topic. It can put together a concise summary of a topic or produce a summary of a more complex paper in simpler terms. ChatGPT remembers your last few interactions, so you can ask follow-up questions. This ability to support follow-up questions that may go deeper than the initial query sets ChatGPT apart from just reading the Wikipedia page. ChatGPT can get you a quick primer on a topic but remember that you still need to check the facts to make sure it hasn’t told you any tall tales.

Code Snippets

Software developers and data scientists do their work today with a bunch of tabs open in their browser, many of them pointed at Stack Overflow, a crowd source tool where people share solutions to coding problems. We all do this because no one remembers all of the syntax for every function in every programming language that we have to use. ChatGPT has learned from the data in Stack Overflow and other sites like it. So, ChatGPT can efficiently produce working code examples that you can use in your projects. In this use case we again need to be careful not to share proprietary source code with Open AI. As tempting as it may be to paste a block of code into ChatGPT and ask it to find the source of the bug, that would send your source code to Open AI, probably not what you want to do.

Availability and Performance

At this stage, ChatGPT is being offered as a completely free service. So, it is a victim of its own popularity. At any given time, users cannot be confident that the service will have enough capacity to support the demand. Users often get an error message indicating there is not enough capacity to support more users at this time. In a future state, where dedicated hardware in the cloud could be allocated based upon budgeted demand, this constraint can be resolved. It is worth noting that performing model inference with models as large as ChatGPT takes a lot of GPU processing time. It is estimated that Open AI is spending $3M per day to keep the current demo running.

Alternatives

ChatGPT is not the only large language model available; it’s just the model that is getting the most attention in the press today. Google has produced a model called Flan-T5 which is competitive with ChatGPT and has a few advantages. Flan-T5 lacks ChatGPT’s fluency and ease of use, but it makes up for it in that it is completely open source and can be run on your local hardware without sharing any data with Google or anyone else.

Conclusion

ChatGPT is an impressive new development for NLP, and I fully expect the speed of delivery of AI solutions like it to continue at a rapid rate. ChatGPT is a great tool for starting a writing task or finding specific syntax in a popular programming language. When used as intended, it can deliver a significant productivity boost. You have to check the facts generated by the chat bot, because it will occasionally make things up and you need to be careful not to send any proprietary data into Open AI’s servers. Meanwhile, you can take advantage of fully opensource solutions which are similar to ChatGPT to continue to safely advance the NLP capabilities at your company.