Executive Guide to ChatGPT

Within just a few days, the beta release of ChatGPT has demonstrated its potential - a conversational AI that can surprise, impress and amuse. Unlike its previous language model siblings, ChatGPT comes with an interface which means anyone (not just developers) can use it. That's why it the preview version has gained users faster than other services. The chat interface has spawned a whole new phenomenon of prompt making (or prompt engineering as some call it.)

Underneath the layer of its well-known shortcomings (see the section below), there are clear signs that ChatGPT will cause major change on how we work with information, content and tasks. The element of surprise and creativity that it can generate is unparalleled. This is not too different from the text-to-image, diffusion based models like DALL-E-2 (can we please have better model names!) have also created quite some buzz.

This guide is meant for executives who are interested in knowing the facts about ChatGPT and learning why ChatGPT is relevant for them. To keep it concise, I've skipped the technical details in the post but included them in the Resources section at the end. I provide a list of generic use cases, not specific to any industry but it's easy to apply them to different industries, roles and situations.

There are important lessons for leaders and executives who are responsible for AI strategy, investing and hiring. Executives must not ignore the new developments because of their current limitations. Leaders should proceed with caution but urgency, should allocate resources to learn the new developments and evaluate how ChatGPT / other LLMs work for their specific problems and situations.

Key Points for ChatGPT:

• What: ChatGPT builds on the family of LLMs (Large Language Model) of natural language models (but with 100x smaller parameters) which are the most powerful models for predicting language. ChatGPT was further trained using Reinforcement Learning with Human Feedback. It is probability based so it is likely to tell you whatever it saw during the training and what the humans taught it. The OpenAI blog post mentions: "We trained this model using Reinforcement Learning from Human Feedback (RLHF), using the same methods as?InstructGPT, but with slight differences in the data collection setup."

fine-tuning with human feedback is a promising direction for aligning language models with human intent.

• Impressive intent understanding and interactive conversation capability: learning from the large corpus of training and ability to take follow up cues and instructions such as examples. It can fine-tune its response based on feedback.
• Show don't tell. Picks up on subtle instructions and is very good when you provide a few examples. For instance, you can ask it be a customer service assistant with a description of its personality: helpful and polite or witty and sarcastic. Before this, copywriters had to craft the conversations to reflect or match a certain type of tone and personality.
• Ability to extract and organize information - excel macros, technical information etc can be well-formatted. Contrast this with the other solutions which are either too simple (text/image) or require custom development work to format nicely (as in bullets, tables etc). Another example - it can generate a quiz from a source.

• Look for rapid speed of improvement: The popularity of the beta version will help the creators make it better (virtuous cycle), just like the popular search engine of the past benefited from the economies of scale and feedback.

Few Noteworthy Use Cases:

• Understanding, Summarization and Explanation (e.g. explain quantum computing to a 6 year old)
• Language translations (limited to the natural languages it knows at this time)
• Creating / generating text: prose (essays, articles, copywriting for ads, taglines or promotions etc), poetry, jokes, fun (movie names to emojis!)
• Brainstorming, idea generation, imitation of examples, helps with the blank slate / writer's block problem
• Automating common online tasks and chores - for instance, making meal plans
• Customer Service (matching the requested tone and personality)
• Technical tasks of various types e.g. writing Excel functions, connecting different online tasks
• Software tasks: explaining code (e.g. regex term), reviewing code, generating code (may not be a working piece of code but it can sure get you started) / recommending technical steps or approach
• Many of the use cases are also mentioned here in the OpenAI API page are also applicable here - see below.

I have not included answering questions here because of the current limitations. I am also certain that there are many many more interesting examples and use cases that are not included above - the idea was to show the vast potential.

Known Limitations

As with all AI, the model has bias from training data, human supervision issues and algorithmic shortcomings. There are 5 major limitations of the first release of ChatGPT at this time. So it needs all kinds of guardrails and caution. And of course, subject to change: these could be fixed over time or new issues could emerge.

1. Model has been trained using human trainers till 2021 end so that marks the limit of its knowledge for events info. For instance, it does not know what happened in World Cup 2022.
2. Truthfulness for a given answer is unknown. For situations where ChatGPT does not the answer, many times, it makes up an answer, even with authentic-looking (but fake) references. The OpenAI post says: "ChatGPT sometimes writes plausible-sounding but incorrect or nonsensical answers." What it means: don't use ChatGPT for fact finding or for use cases where accuracy is top of mind.
3. Disambiguation is missing. Model guesses the intent instead of asking / verifying.
4. Filters for inappropriate requests can be evaded.
5. Intent understanding is impressive but not perfect. It may not be able to answer a question for one input but a tweak in input (prompt) could produce an answer for the same concept.

Also note that "during the research preview, usage of ChatGPT is free." It is almost a given that like other APIs, there will be fee-based model for ChatGPT in near future.

I hope I have convinced you that you cannot ignore ChatGPT, regardless of its current shortcomings and flaws. Drop me a line, if you disagree - would love to engage and hear your point of view.

Remember, the approach of RLHF allows for improvements and there's a good mechanism for fine-tuning the model. So I expect this post to become outdated sooner than later!

Useful Resources and Links:

• Sample Twitter Threads on ChatGPT: Here, this, another one, and this one
• Technical: NLP Technology Guide by DeepLearning.AI, Large Language Models such as GPT-X (where X=2,3, 3.5 ..)
• Official blog post about ChatGPT, Product notes
• Official FAQ - https://help.openai.com/en/collections/3742473-chatgpt
• Explore further with the Playground: https://beta.openai.com/playground and with fine tuning the models