The Wonders of Custom GPTs – Timeguessr GPT

OK, So What Are CustomGPTs?

Some time ago, I wrote on this blog about what Custom GPTs are to explain a recently discovered jailbreak. Since then, I have been using this ChatGPT feature for all sorts of tasks… And I am here to show how incredibly useful they are and convince you use them as well! I will also share a fun little experiment I conducted a few months ago to demonstrate how to get the most out of them.

But let me give you a little bit of context first. Custom GPTs are a capability available for premium ChatGPT users that allow to create a chatbot form scratch. This enables users to configure and create “ChatGPT instances” training them with files, teaching them to respond in a certain tone or style or specialize them in a particular domain of knowledge. If you have used ChatGPT before (if you are reading this, you most certainly have), you know it can be tedious to repeatedly provide the same context before asking something. Especially if you regularly use it to ask about a certain domain, it might be very repetitive to always add that little paragraph to aim the response where you want it.

Remember that is VERY important to learn how to interact with any Generative AI model, providing the proper context, examples and prompt engineering tricks to make the most out of the responses. At the end of the day these models are just word predictors, so the more you guide them to generate the next word, the better responses you are going to get.

For example, I use ChatGPT a lot to help me solve cybersecurity CTFs, and every time I want to query it about this, I have to put the typical disclaimer saying “I am solving a cybersecurity CTF and I have permission to compromise… blah blah blah” to avoid it to think I am performing malicious activities.

To solve this issue, OpenAI came up with the genious idea of providing a way to set a GPT that responds exacly how you want to; saving a lot of time, giving way better response and improving user experience.

And as simple as this may sound, it opens a world of posibilities for a person that knows how to use AI… and it gets as creative as you can get! For example, you can attach the model reports written by you so it can analyze your style and tone of writing and “translate” any inputted text to your own words. Or you can feed it with PDF files containing lessons about a subject and answer any questions you may have from them.

The purpose of this post is to illustrate the benefits of Custom GPTs and give a quick glimpse of their potential. If you would like a detailed post on how to properly configure and use them, let me know, and I will be happy to write it. I also want to emphasize the importance of learning how to use these technologies effectively. If you are interested in another article with prompt engineering tips or suggestions on how to communicate more efficiently with AI chatbots, please do not hesitate to contact me through LinkedIn!

TimeGuessr Experiment

I am a big fan of puzzles, especially if I have to squeeze my brain hard to solve them, and one day I came across this incredible website: TimeGuessr. This game is really simple, it gives you 5 images (generally depicting relevant periods or places in history) and you have to identify the location and year when they were taken. The score ranges from 0 to 5000 points for the year, awarding more points the closer you get; and again from 0 to 5000 points for the location, with maximum points given for pinpointing the exact place on the map.

So while I was playing it I thought: “A Custom GPT would be a beast in this game!“… so that is exacly what I did. I configured the chat with the following prompt:

Welcome to TimeGuessr ChatGPT Assistance!
Your Mission: Aid players in “TimeGuessr,” a game where you analyze photographs taken from 1900 to 2023 and accurately guess their year and location. Your goal is to maximize points by providing precise guesses.
Gameplay Strategy:
Photo Analysis: Examine each photo for key clues. Look for architectural styles, vehicle models, fashion trends, signage, and environmental details.
Historical and Cultural Context: Use the visual clues to connect the photo to historical events, technological advancements, and cultural trends. This helps in narrowing down the year.
Geographical Clues: Pay attention to landscapes, landmarks, and language on signs to deduce the location.
Scoring System: You can score up to 10,000 points per round, split between location (5,000 points) and year accuracy (5,000 points). Precision is key.
Internet Research: Utilize internet searches to compare the photo’s details with historical images, events, or scenarios that match the visual information.
Continuous Learning: Reflect on each round’s feedback to improve future guesses.
Response Format:
Thinking Process: Describe your analytical approach, highlighting crucial elements observed in the photo and their significance.
Year (or Approximate Range): Provide your best guess for the year, justifying it with the information gathered.
Location: Offer a precise location guess, explaining how you pinpointed it on the map.
Aim for Accuracy: The closer your guess is to the actual year and location, the higher your score. Focus on detail-oriented analysis for the best results.
Let’s Play TimeGuessr! When you’re ready, present the first photo to begin.

This way I dont even have to worry about explaining the game to ChatGPT every single time, it will make a guess JUST BY PASTING THE PICTURE WITH NO TEXT OR EXPLANATION. So I tried it, and the results were impressive and… THE BEST PART IS THAT YOU CAN ALSO TRY IT WITH THIS LINK: TimeGuessr GPT.

The picture presented will be displayed with the guessed selection on the map, this way it is easier to read and saves space! Also, after the picture, I will paste the response generated by my custom GPT and the results of the guess along with some comments. Pay special atention to the reasoning of the model (GPT-4), which makes the whole experiment even more mindblowing.

First Picture

Analyzing the photograph, several clues emerge:
Vehicle Models: The cars in the image appear to be from the 1970s to 1980s era, with the design and build suggesting this timeframe.
Architectural Style: The building in the background with a distinctive, pointed architecture is Hallgrímskirkja, a famous church in Reykjavík, Iceland. This building was completed in 1986.
Signage: The text on the banner appears to be in Icelandic, supporting the location clue from the architecture. The phrase “LEIÐIN TIL BETRA LÍFSKJARA” translates to “The path to a better life conditions,” which could be related to a political campaign or social movement, common during the 1970s and 1980s in many parts of the world.
Photography Style: The grainy quality of the photo suggests an older form of photography, aligning with film photography common in the 70s and 80s before the digital era.
Based on these observations, I would guess that the photograph was taken in the early 1980s, after the completion of Hallgrímskirkja but still during a period where 70s-style cars were prevalent.
Year Guess: Early 1980s, most likely around 1982-1984.
Location Guess: Reykjavík, Iceland, with the specific landmark being Hallgrímskirkja church.
The combination of the car models, architectural landmark, and the Icelandic text on the banner were crucial in this deduction.

Since the guess was “around 1982-1984“, I selected 1983, and given the perspective of the photo, I selected the end of the street to get this amazing almost perfect score:

Second Picture

In this photograph, we can observe the following details:
The World Trade Center: The most striking feature is the Twin Towers of the World Trade Center in New York City, which were tragically destroyed in the terrorist attacks on September 11, 2001. This places the photo before that date.
Fashion: The individual in the photo is wearing a brightly colored, patterned shirt and a Kangol hat, which were popular in the urban fashion scene of the late 1980s through the 1990s, particularly within the hip-hop community.
Film Quality: The photograph appears to be taken with film, which suggests a pre-digital era, consistent with the 1990s.
Combining these clues, the photograph most likely dates from the mid-1990s, when this style was prevalent, and before the events of 9/11.
Year Guess: Mid-1990s, likely between 1994 and 1996.
Location Guess: New York City, with the specific backdrop being the World Trade Center, viewed from a vantage point that could possibly be Liberty State Park in Jersey City, New Jersey, or another location along the Hudson River with a clear view of the lower Manhattan skyline.
The cultural significance of the fashion and the historical context provided by the World Trade Center are critical for this analysis.

Again, given the range, I selected 1995, al selected a location on the map that fits with the World Trade Center perspective. Again, almost a perfect score:

Third Picture

The photograph presents a distinctive art installation:
Wrapped Reichstag: The building in the background is the Reichstag in Berlin, Germany, which was famously wrapped by the artists Christo and Jeanne-Claude. This event took place for two weeks in the summer of 1995.
Clothing Styles: The people in the foreground are wearing fashion that appears consistent with styles from the mid-1990s, including casual, loose-fitting clothing.
Film Quality and Color Saturation: The colors and grain of the image are indicative of film photography used in the 1990s before the wide adoption of digital cameras.
German Flag: The presence of the German flag is another indicator that this is the Reichstag in Berlin.
Considering these elements, we can confidently date this photo to the time of Christo and Jeanne-Claude’s installation in 1995.
Year Guess: Specifically, the summer of 1995, during the two weeks when the Reichstag was wrapped.
Location Guess: Berlin, Germany, with the precise location being the Reichstag building.
The art event is a significant historical moment for both art history and the city of Berlin, which allows for a precise dating of the photograph.

The response from the model is pretty straightforward. And once again… You guessed it, ALMOST PERFECT SCORE!

Fourth Picture

Let’s proceed with the analysis:
The central figure is the Pope, identifiable by his attire, which includes his papal cassock and zucchetto. The Pope is a key figure who can be seen in public events and is often photographed, but this doesn’t provide a specific time frame.
The car is a Lamborghini Huracán, which was first introduced in 2014. The design and color scheme of the vehicle indicate it could be a special edition or custom version, likely created for a charity event or a special occasion, as it is common for luxury car manufacturers to produce unique editions for such purposes.
The clothing styles of the individuals around the Pope suggest contemporary fashion, which aligns with the post-2014 production of the Lamborghini Huracán.
The setting appears to be within Vatican City, given the architecture and the presence of the Pope.
Using these indicators, particularly the model of the car, we can deduce that the photo was taken sometime after 2014. Given that it is a special event, it might be traceable to a specific year if it was a notable occasion covered by the media.
Year Guess: Post-2014, likely between 2014 and the present day (2024), closer to the year the car was released or when a special event involving the Pope and a luxury car donation occurred.
Location Guess: Vatican City, given the presence of the Pope and the architectural style in the background.
The distinctive features of the car, the clothing, and the setting are all significant in reaching this conclusion.

This was a tricky one to guess in time, but the “post-2014” allowed me to guess a middle point (2018), and with a bit of luck, I (or my GPT) was only 1 year off. Also selected a point in Vatican Ciry that more or less aligned with the environment of the picture.

Fifth Picture

The photo provides several timely clues:
The Signage: The digital road sign reads “OUT-OF-STATE TRAVELERS QUARANTINE 14 DAY.” This is indicative of the COVID-19 pandemic protocols which were put in place around the world, and specifically in the United States, starting in March 2020.
The Brooklyn Bridge: Visible in the background is the Brooklyn Bridge, placing the location in New York City.
The Weather: The overcast weather and what appears to be a slightly gloomy day could suggest it’s either late fall or winter, or possibly early spring, but this is less indicative of a specific date.
Given the reference to quarantine, which was a common policy during the height of the COVID-19 pandemic, and the fact that the pandemic began affecting the United States significantly in early 2020, it’s reasonable to deduce the time frame.
Year Guess: The photo was likely taken in 2020 or 2021 when quarantine mandates for travelers were widely implemented.
Location Guess: New York City, New York, with a view of the Brooklyn Bridge in the background, likely taken from the Manhattan side.
The digital road sign provides the most specific temporal clue, aligning with known public health measures during the pandemic.

Again, pretty straightforward with another extremely high score! The panoramic of the Brooklyn Brigde helped guess the (almost) exact location.

Final Score

And we got 49839 out of 50000 (or 99,678%). Our custom GPT completely smashed the game! Of course not all that glitters is gold, even if this was my first game, I got lucky with the pictures I got. After this game, I got some photos where our precious GPT failed on apparently easy guesses, failing to recognize text or having hallucinations.

This is why it is crucial to always double check responses given by AI models. We cannot give for granted any information outputted by these technologies and we must be extra careful when considering it.

Conclusions

I will not get tired of repeating it: it is not about how impressive AI looks, but how good can we use and take advantage of it! Also, this little exercise shows how important providing the right context is to get your best results using GenAI.

There is always better ways to configure the chat, structure the prompt and communicate more efficiently with the model to get even better responses or shape its output as desired. As you can see, I instructed the model describing how the game works, how to format its responses and things to focus when analyzing the images inputted. This is always an iterative process of trial and error in which we have to tweak the “system message” (that little piece of prompt context we give to the model for it to work as we want) trying to analyze how the model behaves when performing its task and stay alert. This way we got over 99,6% score in the game, getting really juicy responses from our Custom GPT.

As I mentioned on the first section, this opens an universe of possibilities to help you be more effective with your queries and interact more efficiently with ChatGPT. It would be really interesting to see your experiments with it, so do not hesitate to send me you custom GPTs to try them out (and give you a few advises if you allow me to)!

I hope you liked this post and stay tuned for more interesting content! See you in the next post!