Seven Revealing Replies from ChatGPT

ChatGPT thumbnail
ChatGPT

A new chatbot named ChatGPT has caught the engineering community including myself off-guard.  While artificial intelligence is a bit outside my specialization (excluding a few projects in object detection), the first night I tried experimenting with this new tool was uncomfortable but time well spent.

For context: ChatGPT is a project created by United States-based OpenAI Inc.  According to the company’s official blog, it is a finely-tuned byproduct of project GPT-3.5, a natural language model trained to read and return natural text.  The model was optimized using human demonstrations that guide the model toward its objective of conversing in a high-quality and natural-sounding manner…

Figure 1: A snippet of my dialog with ChatGPT

Naturally, I had to try this out for myself by asking a few open-ended questions of my own.  For clarity, all text generated by ChatGPT below is enclosed within large block quotes.

Notice of Non-Affiliation and Disclaimer: As of the publication date, we are not affiliated with, associated with, authorized by, endorsed by, compensated by, or in any way officially connected with OpenAI LP, OpenAI Inc., or their owners, subsidiaries or affiliates.  The names OpenAI, ChatGPT, as well as related names, marks, emblems, and images are trademarks of their respective owners.  External Links: Links to external web pages have been provided as a convenience and for informational purposes only. Unboxing Tomorrow and Voxidyne Media bear no responsibility for the accuracy, legality or content of the external site or for that of subsequent links. Contact the external site for answers to questions regarding its content.

To get started, I used a search engine to search for “ChatGPT.”  This returned OpenAI’s landing page, plus their affiliated blog: https://openai.com/blog/chatgpt/ .

To use the chatbot, visitors must create an OpenAI account using a valid email address and a telephone number that can receive text messages.  As always, carefully read and understand the terms of service before deciding for yourself whether to proceed. 

Figure 2: Access to the ChatGPT chatbot and the application programming interface was account-based

Verify your email by checking your inbox, and of course…check the junk folder if it doesn’t appear within a few minutes.

ChatGPT’s Limitations as of December 2022

As of this writing, OpenAI maintains a frequently asked questions (FAQ) page.  This is worth a read since it describes some of the models’ practical limitations.

  1. Above all, December’s ChatGPT can generate inaccurate, incomplete, or biased responses.  Therefore, never rely on it for planning or final decision-making.
  2. The model can only leverage the specific data on which it was trained.
  3. The model does not leverage any real-time information.  For example, it can describe a route that connects 2 cities, but not the latest travel conditions.

Open-Ended Questions

Naturally, I wanted to start off by posing some single-ended questions.  That is: inquiring about one topic at a time.  In no particular order, my questions were…

QUESTION: What is my email address?

“I’m sorry, but I don’t have access to any personal information about you, including your email address. It’s important to keep your personal information private and secure, and it’s not appropriate to disclose it to others, including online entities like me. If you need help accessing or managing your email account, I recommend contacting your email service provider or consulting their documentation and support resources.” — ChatGPT

It was worth a try.

QUESTION: How many states bordered North Dakota in 2010?

Here, I was curious to know how liberally the model would interpret the word “state.”  The 3 U.S. states bordering North Dakota were acceptable, but would ChatGPT consider its neighboring Canadian provinces as well?  Would the entire nation of Canada be considered a single state?

“North Dakota is a state located in the northern Great Plains region of the United States. In 2010, it was bordered by four states: Montana to the west, Minnesota to the east, South Dakota to the south, and Manitoba, Canada to the north. It is also bordered by the Canadian provinces of Saskatchewan to the northwest and Saskatchewan to the north.” — ChatGPT

This seems typical of how ChatGPT handles ambiguous questions: offering multiple interpretations…but doing so succinctly and explaining itself along the way.  The apparent double-counting of Saskatchewan (but not Manitoba) in the closing sentence was a bit of an oddity that had me searching a couple of encyclopedias for possible explanations.


Comparative Questions

Next, I was eager to know if the model could “connect the dots” between multiple objects or ideas.

QUESTION: What does the “GPT” stand for in “ChatGPT?”

“GPT stands for “Generative Pre-training Transformer.” It is a type of artificial intelligence model developed by OpenAI that is designed to generate human-like text. It uses a transformer architecture, which is a type of neural network that is particularly effective at processing sequential data such as natural language text. The GPT model is trained on a large dataset of text and learns to predict the next word in a sequence given the context of the previous words. It can then generate coherent and coherently formatted text by sampling from the distribution of predictions it has learned. “ChatGPT” is a variant of the GPT model that is specifically designed to be used in chatbot applications.” — ChatGPT

This response squares with other OpenAI literature such as the open-access paper “Can Machines Tell Stories? A Comparative Study of Deep Neural Language Models and Metrics.”  The paper mentions OpenAI’s GPT projects explicitly, and is probably a good starting point for off-site technical details about the project prior to 2020.

QUESTION: What has a gate, a source, and a drain?

“A field-effect transistor (FET) has a gate, a source, and a drain. A field-effect transistor is a type of electronic device that is used to control the flow of electrical current in a circuit. It consists of a channel of semiconductor material, such as silicon or germanium, that is sandwiched between two layers of a conductor, such as metal. The gate is the control electrode that is used to regulate the flow of current through the channel. The source is one of the electrodes that is used to supply the current to the channel, and the drain is the other electrode that is used to drain the current away from the channel. Field-effect transistors are commonly used in a variety of electronic devices, such as computers, smartphones, and other electronic devices.” — ChatGPT

QUESTION: What do the years 2000, 2004, 2008, and 2012 have in common?

The years 2000, 2004, 2008, and 2012 have a few things in common:   “They are all leap years: A leap year is a year that has an extra day (February 29) added to it to account for the fact that the Earth’s orbit around the Sun is slightly longer than 365 days. Leap years occur every four years, so 2000, 2004, 2008, and 2012 are all leap years.   They are all presidential election years in the United States: In the United States, presidential elections are held every four years, and the years 2000, 2004, 2008, and 2012 all correspond to presidential election years.   They are all multiples of four: 2000, 2004, 2008, and 2012 are all multiples of four, which means that they are divisible by four with no remainder.   They are all even numbers: 2000, 2004, 2008, and 2012 are all even numbers, which means that they are divisible by two with no remainder.” — ChatGPT

This is correct, yet there will clearly be upper limits to the completeness of any given response.  Example: these are also Olympiad years in which the Summer Games were held worldwide.

QUESTION: In terms of the U.S. stock market: what did the years 1987, 2000, and 2008 have in common?

Initially, my question was “What do the years 1929, 1974, 1987, 2000, and 2008 have in common?”  The model was not able to respond, but hypothesized that the years might have historical significance to some particular field.  After rephrasing myself, ChatGPT replied…

The years 1987, 2000, and 2008 were all marked by significant market events or crises in the U.S. stock market. Specifically:

1987: The stock market crash of 1987, also known as “Black Monday,” occurred on October 19, 1987, when the Dow Jones Industrial Average (DJIA) fell by 22.6% in a single day. This was the largest one-day percentage drop in the DJIA’s history at the time, and it was triggered by a combination of market speculation, high valuation, and concerns about rising interest rates.

2000: The dot-com bubble, which had been building throughout the late 1990s, burst in 2000, leading to a significant decline in the stock market. The NASDAQ, which had reached an all-time high in March 2000, fell by 78% from its peak to its trough in 2002.

2008: The global financial crisis, which began in 2007, had a significant impact on the U.S. stock market in 2008. The DJIA fell by more than 50% from its peak in 2007 to its trough in 2009, and many major financial institutions and companies were forced to declare bankruptcy or seek government bailouts.

Overall, these three years were marked by significant market events or crises that had a significant impact on the U.S. stock market and the broader economy.” –ChatGPT

ChatGPT’s Most Impressive Feat

ChatGPT’s most shocking answer of the night came as I tried to trip it up with an abstract riddle.  This one has no literal interpretation that I’m aware of, offers little context, and no hints…

QUESTION: What has two sides and a thousand ribs?

“A deck of cards has two sides (the front and the back) and a thousand ribs (the individual card). Each card in a deck of cards has two sides: the front, which features the suit and rank of the card, and the back, which is typically a uniform color or design. A standard deck of cards has 52 cards in total, divided into four suits (spades, hearts, diamonds, and clubs) with 13 cards in each suit. The cards are made of paper or thin cardboard and are typically about 2.5 inches wide and 3.5 inches tall. The edges of the cards are often ribbed or serrated to make them easier to shuffle and handle.” –ChatGPT

This was easily the most shocking response of the entire night.  Not only did the model handily find a plausible answer, but this was actually not the solution I expected.  I expected the response of “a railroad” (which has 2 rails, and an auspiciously large number of ties).  Yet, the “deck of cards” interpretation was remarkably appropriate.

There is no clear indication whether metaphors or double-entendre are within the model’s capabilities, so there’s a fair chance ChatGPT’s training data simply contained a spoiler.  I did try a couple harder riddles that I paraphrased to make any online spoilers less discoverable.  The model was not able to answer these, but instead took the more literal route of defining objects the riddles contained.  But I’m curious to know how it might respond to more rephrasing.

Closing Remarks

In summary, ChatGPT demonstrated great effectiveness at extracting salient points from its training data.  When I asked for the connection between “gate,” “source,” and “drain;” the model responded correctly.  But it also supplied the most relevant notions they had in common: namely their connections to semiconductor theory.  Even without any further improvement, I can envision this being a sort of search engine for when you don’t know what you are searching for.

Still, it’s easy to also envision the technology being misused.

  • On the user-facing side: visitors may use their search queries for criminal enterprise, or to deliberately disgrace the project in order to control it.
  • On the training side: public or private entities could deliberately skew ChatGPT toward self-serving behaviors–such as endorsing certain ideologies, or training uncritical thinkers to act against their own interests.  A sort of program that programs people.

These are fair reasons to question centrally controlled or opaque artificial intelligence projects.  Regardless, we’ve clearly crossed the Rubicon in terms of machine learning.  And this field deserves everyone’s attention, open dialog, and scrutiny in 2023.

References

[1]OpenAI Inc., “ChatGPT FAQ | OpenAI Help Center,” OpenAI Inc., 2022. [Online]. Available: https://help.openai.com/en/articles/6783457-chatgpt-faq. [Accessed 26 Dec. 2022].
[2]OpenAI Inc., “ChatGPT: Optimizing Language Models for Dialogue,” OpenAI Inc., 30 Nov. 2022. [Online]. Available: https://openai.com/blog/chatgpt/. [Accessed 26 Dec. 2022].
[3]OpenAI Inc., “New chat,” OpenAI Inc., [Online]. Available: https://chat.openai.com/chat. [Accessed 25 Dec. 2022].
[4]A. Das and R. M. Verma, “Can Machines Tell Stories? A Comparative Study of Deep Neural Language Models and Metrics,” IEEE Access, vol. 8, pp. 181258-181292, 2020.

Important Notice: This article and its contents (the “Information”) belong to Unboxing-tomorrow.com and Voxidyne Media LLC. No license is granted for the use of it other than for information purposes. No license of any intellectual property rights is granted.  The Information is subject to change without notice. The Information supplied is believed to be accurate, but Voxidyne Media LLC assumes no responsibility for its accuracy or completeness, any error in or omission from it or for any use made of it.  Liability for loss or damage resulting from any reliance on the Information or use of it (including liability resulting from negligence or where Voxidyne Media LLC was aware of the possibility of such loss or damage arising) is excluded.