The popularity of ChatGPT has exploded since it was released 2 months ago. As more and more people play around with the tool, they find the possibilities for what it can achieve seem endless. It doesn’t feel like an upgraded chatbot, but it is very close to conversing with a very smart human like Ken Jennings. The user is able to get good succinct answers to complex questions, it can tell a joke based on a subject, and can craft a good short story. Microsoft recently invested 10B in OpenAI, the creators of ChatGPT and Dall-E in a move that positions itself at the head of generative/creative artificial intelligence technology. I did not really pay attention until this move by Microsoft. When I think of natural language processing (NLP), I think of chatbots, programs that summarize text, sentiment analysis, or text classification. I was not impressed until I asked ChatGPT to explain what it does in the form of a rhyme:

I’m ChatGPT, a language model AI,
My function’s clear, to respond and reply,
With a text that’s informative and quick,
So you don’t have to search or pick.

I work on your prompts with great pace,
Computing vast data, with speed and grace,
Through OpenAI’s training and its might,
I can answer questions day and night.

I’ll provide you with answers so neat,
To make your life simpler, can’t be beat,
From trivia to stories, I’ll comply,
Just type your query, and I’ll reply.

So go ahead, give me a chance,
To assist you, with a little dance,
I’ll be here, ready to rhyme,
Making your life easier one line at a time.

This was the first evidence that made me think programs can become as creative as people. Creativity is partly mixing ideas from different areas and finding new applications and the drawback for machine learning and deep learning have been that most programs are laser focused on one task. It might be able to beat a chess grandmaster, but can it help a business leader make better business decisions. ChatGPT is able to help in many different arenas it can solve problems in customer service,  chatbots, and content creation. Users have found that it can be more convenient than google search, help write and debug code, and write essays that can pass a Wharton MBA final exam. All of these are solutions should make companies and workers worried about whether their lunch will be eaten by ChatGPT.

How does it work?

ChatGPT is run on GPT-3 which is a large language model developed by OpanAI. GPT-3 differs from the usual NLP techniques because of the massive scale, it was fed 570gb of text data and contains 175 billion parameters. This is about a 100x increase over its last iteration. This scale allows it to be more of a general application versus just focused on a single industry.

We might compare it incorreclty to the autocomplete feature of Google search, but instead of predicting what the next word or phrase is being typed, ChatGPT will predict whole sentences and paragraphs. Users have even used it to create scripts with dialogue. 

GPT-3 stands for Generative Pretrained Transformer-3. The transformer architecture is what makes the model so powerful and differs from recurrent neural networks (RNNs) in that data does not need to be read sequentially. This is one of the main reasons for its fast speed and computation, it is free to use parallel computing. The next technology used is transformer attention and more specifically self-attention. A transformer adds a position label (positional encoding) to each word of the query and remembers how it pertains to other words. It then looks at all the words in the input and creates predictions on what words should be in the output (attention).

This has been used before but the added positional encoding is important because now the model learns rules such as how to deal with ambiguity, grammar, correct tense, and informal guidelines to language. In other words, the inputs to the query are also being used (self-attention) to learn about what the answer should look like. In Star Wars, Anakin never wiped R2-D2’s memory after missions which would have protected valuable data from the Separtists, but this allowed R2 to form strategies and tactics that other droids could not create. Similarly the transformer model can remember the positioning of words and can use those inputs to help it learn how words should be arranged. 

The human element to ChatGPT is the person that grades outputs to make sure that it is performing as expected. This labeler checks the results against tasks and is key component to the algorithm learning to maximize its subsequent results. This process of attempting to maximize the grades is aptly named Reinforcement Learning with Human Feedback (RLHF) training. Humans are also needed to make sure goofballs aren’t using technology to ask about offensive, harmful, or inappropriate topics. It would be fun to try to get the model to talk about hot takes, but eventually you’ll most likely get too many negative inputs and ruin the technology for everyone.

Best Uses

• Simple questions
• Content creation
  • Write jokes or stories
  • Write legal and copyright templates
  • Write emails
• Write or debug code
• Language Translation

Simple Questions

This is the main reason why google is said to be facing a Kodak/Blockbuster moment. If you know a bit about what you are asking you can ask ChatGPT to give you a quick summary instead of searching on Google and looking at the validity of the website. The moments when I went over to google are when I would need to look at images or if I had to look more in depth about a topic. Trust in ChatGPT’s answers provided were established very quickly.

Content Creation

What potentially makes ChatGPT very profitable is its ability to create content. You can ask it to create pretty well formed stories: a simple command such as “write a 1000 word story about someone moving from financial services to data science” comes up with a story that is similar with my own. The drawback is that it only wrote 500 words. I previously commanded it to write a 5000 word story and it could not because it requires detailed description and character development which wouldn’t fit the text-based interface.

It can also help write emails. For those that take forever to write professional sounding emails, this will be a godsend. It can write a cover letter that matches what you would find in a job advice site and it fills in the details you provided. At the very minimum, it will get rid of a lot of the low level email tasks that office work entails.

A very interesting thing that surprised me was it’s ability to create legal documents. When asked about types of legal documents it can write, six domains were listed:

• Contracts: I can assist in writing agreements between two parties such as employment contracts, service contracts, confidentiality agreements, etc.
• Legal letters: I can write letters to inform parties of their legal rights and obligations, or to respond to legal claims.
• Business formations: I can help write documents related to the formation of a business, such as articles of incorporation, bylaws, and operating agreements.
• Intellectual property: I can assist in writing legal documents related to intellectual property, such as patents, trademarks, and copyrights.
• Estate planning: I can help draft documents related to estate planning, such as wills, trusts, and powers of attorney.
• Legal memoranda: I can write legal memoranda outlining the relevant law and arguments for a particular case or issue

The coup-de-grace was a note at the end mentioning that it is not a lawyer, cannot provide legal advice, and this is for information purposes only. It’ll be very interesting to see the limits of what it can do and how it will affect the industry.

Write or Debug Code

This was pretty scary, I found a hard leetcode python problem and ChatGPT was able to solve it in under 2 minutes while ranking in the 56th percentile for runtime. The cynic in me tells me that it is a straightforward question and that will be given a straightforward answer, but it does seem to be good at helping to write code. 

I have used it for less known languages like Sybase and SAP HANA and it came back with some wrong answers. It was in the ballpar, but it is a reminder that might not work in every situation.

The user should be warned that ChatGPT can be wrong in subjects and present those mistakes as fact. This often occurs when it is looking at wrong data and creates the answer from where it did not exists (hallucinating), often these answers can be convincing, but will be incorrect. That is where an expert is needed to oversee the results, it can’t make someone a subject matter overnight.

Language Translation

For fun, I tried getting the lyrics of ‘In da Club’ translated into French, but it turns out there are some bad words in the lyrics, so it did not allow full translation. You can easily translate portions of texts using ChatGPT with a prompt like “Translate the Preamble to the Constitution to Spanish”.

Preambulo de la Constitución de los Estados Unidos

Nosotros, el pueblo de los Estados Unidos, con el objeto de formar una unión más perfecta, establecer justicia, garantizar la tranquilidad interior, proveer para la defensa común, promover el bienestar general y asegurar los bendiciones de la libertad para nosotros mismos y para nuestra posteridad, ordenamos, establecemos y damos esta Constitución para los Estados Unidos de América.

Is it worth it?

It was announced that a subscription model will be available for 20 dollars a month. Benefits include faster response times, availability during peak moments, and priority access to new features and improvements. I make pricing decisions based on how much value they provide, how much I value my time, and how much other services are worth.

ChatGPT has several things that I find valuable: being able to explain complex ideas (simpler google), writing common emails, and helping write code. In terms of usefulness, if I use it once or twice a day every weekday, it would be very valuable.

Whenever I get a chance to delegate tasks I don’t like to do, I compare the cost and how much I value my time. For tasks such as washing and folding laundry, which I absolutely hate doing I compare how long it would take me versus how much time it would cost me. If it takes 2 hours to do laundry and costs 50 dollars to have a professional do it, I would look at that $25 dollar per hour and compare to how much I value my time, which I value at multiples of $25 dollars. For laundry, that decision to have someone else do it is easy to make. ChatGPT costs $20 dollars for one month and it has the potential to save me 20 hours a week, it would definitely be a steal.