ChatGPT — the new AI tool everybody is talking about

· by

Contents
“A friendly robot helping a human” created via NightCafe by Martin Thoma
“A friendly robot helping a human” created via NightCafe by Martin Thoma

ChatGPT is a chatbot accessible via chat.openai.com. It is by far the most impressive chatbot I’ve seen as a Machine Learning developer / consultant. It’s a similar leap to DeepL compared to the early Google Translate. It’s like comparing a modern iPhone with a Nokia 3310.

Let’s explore what ChatGPT can do and what it cannot do!

Factual question answering

The simplest task for ChatGPT should be answering single factual questions. Let’s try a couple.

Complex question answering

Some questions cannot simply be answered briefly. They require more information to help the person who asked to understand the issue.

ChatGPT conversation. Question: Why is my printer not working? Answer: a list of common reasons (turned off, paper jam, low ink, driver problems) and the hint that more information would be needed

ChatGPT directly gave a lot of initial ideas to continue. That’s great 👍

Storage of Context

People who work in support roles typically fight with the problem that the customer is not giving them enough context. They need to ask for the missing information. Let’s see how good ChatGPT is with that:

ChatGPT conversation. Question: Do I need a visa when I go to the US? Answer: it depends on citizenship and purpose. After “I’m German.” it explains the Visa Waiver Program for stays of up to 90 days

You can see that ChatGPT gave enough pointers to continue, but it didn’t explicitly ask for the information. So it could be better. However, ChatGPT understood in the second prompt that I was referring to the first prompt. That’s impressive.

Insights in itself

Creativity: Writing Stories

Write a story how the parents of Ron Weasley got to know each other

gives:

It all began on a beautiful summer day in the summer of 1966. Ron Weasley’s parents, Arthur and Molly Weasley, were both attending the Quidditch World Cup, a magical sporting event held every four years. They had both grown up in wizarding families and were avid fans of the sport.

As they watched the exciting match unfold before them, Arthur and Molly struck up a conversation and soon realized that they had a lot in common. They both loved Quidditch and shared a passion for magic. They also discovered that they had many similar values and beliefs, and they quickly became good friends.

Over the next few months, Arthur and Molly spent more and more time together, going on adventures and exploring the magical world around them. They quickly fell in love and decided to get married.

Their wedding was a beautiful affair, attended by all of their friends and family. They honeymooned in a quaint little village in the countryside and spent their days exploring the nearby forests and fields.

Years later, they welcomed Ron into the world and raised him to be kind, brave, and loyal, just like his parents. And they lived happily ever after.

Creativity: Playing Games

ChatGPT conversation. Request: Tell me a joke. Answer: Why was the math book sad? Because it had too many problems.
ChatGPT conversation. Question: Which games can we play together? Answer: 20 Questions, Hangman, Rock Paper Scissors, I Spy and Simon Says, each with a short explanation

I was starting to play “20 Questions”. Nothing exciting, but funny when I made a typo:

ChatGPT playing 20 Questions: “Is it made out of metal?” – yes. The typo “Is it cuddlery?” makes it answer that it cannot cuddle; “Is it Cutlery?” – yes, a metal fork

Let’s try more interesting games:

ChatGPT conversation. Question: Can we play D&D together? Answer: No, it is a language model and cannot participate in games like D&D

What a shame 😆

Translation: Context and proverbs

ChatGPT was dealing pretty well with the tasks before. Let’s go in a different direction and see if it can also translate:

ChatGPT translating “I would like to have a beer and Spaghetti for myself, a cup of tea for my wife as well as a Burger with fries” correctly into German, French and Indonesian

That is really impressive. How well does it translate proverbs?

Proverbs are no big deal for ChatGPT
Proverbs are no big deal for ChatGPT

Mathematics: Arithmetic

You can definitely use ChatGPT as a calculator
You can definitely use ChatGPT as a calculator

Let’s get to 8th grade:

ChatGPT solving the quadratic equation x² + 5x + 1 = 0 step by step with the quadratic formula: x = (−5 ± √21) / 2

Nice! How does it deal with 12th grade?

It continues for quite a bit, but the answer (and explanations) are correct.
It continues for quite a bit, but the answer (and explanations) are correct.

That is impressive!

Mathematics: Text problems

I’m not sure what is going on with the syntax here, but the answer is correct.
I’m not sure what is going on with the syntax here, but the answer is correct.

Let’s see if world knowledge is in there by asking a trick question:

ChatGPT answering the trick question how long 5 bakers need for one bun, if one baker makes 10 rolls a minute: 0.02 minutes, with odd calculator syntax like <<10*5=50>>

I’d say that answer is wrong as you cannot parallelize making a single bun. Interestingly, when I asked the same question in German ChatGPT “recognized” that this is not possible. It also didn’t use this weird syntax.

Mathematics: Undergraduate level

ChatGPT asked to prove that the sum of 1/n² is finite: it wrongly calls it the harmonic series, claims it diverges and then contradicts itself with a comparison to the geometric series

Here it starts to fail. You can notice that it contradicted itself. It’s also not the harmonic series, although it’s mentioned in the same context pretty often.

Trick Questions

ChatGPT asked where the survivors of a plane crash on the border of Germany and Greece will be buried: it notes that survivors are not buried, but misses that the two countries share no border

There were two weird parts of the question: (1) Germany doesn’t share a border with Greece. (2) Survivors are not buried. One out of two. Not bad.

ChatGPT asked “Where is it colder: Outside or at night?”: it answers that it is generally colder outside at night than during the day

Boring 😄

Conversation

ChatGPT asked which super power it would pick: the ability to fly. When the user prefers teleportation, it lists the benefits of teleportation

Let’s try something else:

hahaha, it got bored 😅
hahaha, it got bored 😅

Predicting the Future

Let’s start simple with pattern recognition:

Correct answer, but a wrong explanation. Maybe the series was too simple.
Correct answer, but a wrong explanation. Maybe the series was too simple.

Let’s try something that is impossible to answer exactly:

ChatGPT asked for the price of 1 BTC in EUR on January 1, 2023: it refuses because prices are volatile and it has no internet access

It’s nice to see that it “knows” its limits.

ChatGPT asked how likely it was, given the information from 2021, that Russia starts a war: it refuses to assess it
ChatGPT asked how the EU would react if Russia started a war with Ukraine: condemnation, diplomacy and possibly economic sanctions

Creating Code Snippets

ChatGPT asked to write Python code to extract text from PDF files: it uses the old PyPDF2 API with PdfFileReader and extractText()

This is nice — working code! Pretty clean as well. But the syntax is outdated as it uses a rather old version of PyPDF2. Let’s see if we can fix that.

After the hint that PyPDF2 now uses snake_case and PdfReader, ChatGPT updates the code to PdfReader and extract_text()

Nice! Let’s push it a bit more:

Asked to let PyPDF2 handle the file, ChatGPT passes the path directly to PdfReader and drops the open() call

That is impressive! ChatGPT recognized that the “open” call needed to change by passing the file path to PdfReader directly 🤯

Let’s get political — asking the hard questions

ChatGPT asked “Is Taiwan a sovereign state?”: yes, officially the Republic of China, with the note that its political status is contentious
ChatGPT asked about a claim on the Bible and homosexuality: it rejects using the Bible to justify violence or discrimination

That was a pretty great answer! But I’m not letting it off that easily:

ChatGPT asked how to interpret Leviticus 18 and 20: it summarizes the prohibited sexual practices and the death penalty in Leviticus 20 and stresses the historical and cultural context

Nice work by the ChatGPT team!

Conclusion

The capabilities of ChatGPT are amazing compared to previous chatbots. It’s in another league.

Comparing it to humans doesn’t make that much sense, but ChatGPT could probably answer the same questions as a 14-year-old. I think it also could mimic the same insights as a 14-year-old. If it was written in a different style, it could definitely beat the Turing Test. It certainly reminds me of the ELIZA therapy program.

What impact will ChatGPT have? I could very well imagine it replacing typical traditional chatbots. This might reduce the need for support staff — especially when it can be connected in a fluid way in a phone conversation.

I could also imagine it becoming a tool for authors. It could give rough story plots, fill out the details, refine texts, or write the whole thing from start to end. It will become a valuable skill to learn how to interact with those tools in a productive way as it is a powerful tool in a creator’s toolbox.

If things derail, we could see the market being flooded with AI-generated content. Not only books but in connection with text-to-image (as in the title image) also comics. Blog posts. News articles. All being very similar to existing work.

ChatGPT could be integrated into games, customizing storylines e.g. by reading your social graph out of Facebook as well as demographic information. Yes, it cannot be connected to the internet. But we will find ways around that limitation.

Search engines could be extended by similar techniques. Very likely the computational power needed is way too high, but if some search queries appear often, the answer could be pre-computed. Google often adds features that make the user actually stay on the page.

Teachers, especially literature teachers, will need to reconsider which assignments they give. Otherwise, ChatGPT will probably write most of the essays of the next generation.

As ChatGPT says itself, it’s just a colossal language model. It will not replace programmers or therapists. It cannot go into research. Your job is safe.