Decoding the Puzzle: How Quartiles Highlights AI and Human Problem-Solving

Exploring the Intersection of Human Ingenuity and AI Limitations Through Apple’s Quartiles

Awhile ago Apple added Quartiles to its selection of puzzles.? If you’re not familiar with Quartiles, they’re a new puzzle available to Apple News+ subscribers on devices running iOS 17.5 or later. ?

Upon starting to play you are presented with a grid of 20 tiles containing two to four letters each.? You will need to combine the four of the tiles will form words.? Words can be made from one, two, three or four tiles.? Correctly forming words scores points ranging from 1 point for a single tile, 2 for two tiles, 4 for three tiles, and 8 for four tiles.? If you find all five four-tile words, you’ll score 40 bonus points.? The goal is to achieve the expert rank of 100 points.

When I first started working these puzzles, sometimes it took me a while to solve them – occasionally up to 30 minutes.? After more than 200 puzzles my elapsed time to achieve 100% is now 10 minutes or less.? Which only proves that the more you train on a task the better you get at that task.

The more you train on a task, the better you get at.? This encapsulates a fundamental principle of learning and skill development, highlighting the importance of repetition and practice when it comes to improving specific activities.? The concept applies to both human learning and machine learning.? Focused, repeated engagement strengthens neural pathways and algorithmic patterns.

One-tile and two-tile words are easy to spot.? Three-tile words are a bit harder.? And four-tile words can be darn right difficult.? Since you don’t reuse tiles for the file-tile words, the number of choices is reduced as you progress in solving the combinations.

Which you might think makes it easier to shuffle the tiles and solve for the final word or words.

But sometimes, you can run into a problem and just can’t solve the puzzle.? Friday was one of those days when I was really struggling to sort the last eight tiles into two words of four tiles each.? The screenshot below depicts the point at which I was stumped.

I found “milkshake”, “subscription”, and “persnickety” quickly.? But was stumped to make sense out of tiles “ity”, “ial”, “coll”, “ase”, “hr”, “par”, “ap”, and “eg”.? Spoiler alert, if you read ahead, you’ll learn the two words that solve the puzzle.

Having fussed with the solution for several minutes longer than I had patience for, I decided to see if ChatGPT could help me out.? So, I crafted a prompt in hopes the AI could solve for the remaining two words.? The results are in the screen shot below.

Perhaps ChatGPT misunderstood my instructions.? I knew what I meant, but ChatGPT may not have.? The AI responded by solving the problem as follows:

·???????? collateral: coll + ater + al + ity

·???????? phrasing: phr + as + ing + ap

Neither “ater” or “ing” are tiles in the puzzle. ?Nor did they appear in the list of the letter groups that I provided in the prompt.

But there was a hint in ChatGPT’s response that allowed me to solve the problem on my own.? “Phrasing” allowed me to realize that I missed “par” + “ap” + “hr” + “ase”, spelling “paraphrase”.? Bingo, four words down and one to go.

This made it easier to correctly sequence the remaining four tiles “coll” + “eg” + “ial” + “ity”, spelling “collegiality”.

I admonished ChatGPT calling its response wrong and informing it that the solution was two other words.? To which the AI responded that it was glad that we got it sorted out.

Later in the morning I tried Perplexity AI to see if that AI might solve the problem.? You can Perplexity’s response in the screen shot below.

This was equally disappointing.? I’m not sure why my admonishment was put in such large text.? Perhaps the interface lost formatting.? So, it seems like I was shouting at the AI about its error.? The AI thanked me for my patience and seemed to appreciate the clarification.

Puzzles, as I’ve discussed previously, often present significant challenges for conversational AI. ?This is particularly evident in problems requiring complex reasoning, such as solving word puzzles or mathematical concepts like Quartiles.

The difficulty lies in the way large language models (LLMs) like ChatGPT and Perplexity are designed. ?These models rely on algorithms that are really good at pattern recognition and predicting the next most likely word in the sequence. ?The predictive approach employed by these AIs works remarkably well for generating coherent text, answering factual questions, or simulating conversational flow.

However, these algorithms have limitations when it comes to abstract reasoning or tasks that involve breaking down and reconstructing parts of words, like letters in a Quartile puzzle.

Unlike humans, who often use intuition and logic and a bit of luck to solve problems, LLMs don't inherently "understand" the puzzles they solve. ?Instead, they draw on probabilities derived from vast amounts of text data they were trained on. ?If the specific type of puzzle or reasoning hasn’t appeared frequently in their training data or wasn’t part of the patterns the AI were trained on, they will struggle to generate a correct response. ?Limitations such as this are particularly pronounced in tasks that require logical deductions and creative leaps.

Additionally, puzzles often rely on a nuanced understanding of context, metaphor, or double meanings.? These are areas where our human brains excel.? Our innate cognitive flexibility and years of experience with language and problem-solving allow us to quickly find the solution. ?In contrast, AIs have not yet caught up in their ability to generalize from learned patterns to novel situations.

As AI models evolve, researchers and developers are finding new ways to enhance AI model reasoning capabilities.? Word problems and puzzles may not be high on their list of problems to solve for in their AI models, but improving their performance is. ?Techniques that integrate symbolic reasoning based on specialized datasets will likely help bridge any gaps.

For now, though, puzzles remain a fascinating area that highlights the divide between human and AI problem-solving strengths.

The solution and my score is revealed below.

In the meantime, with an AI’s help yesterday, I continued my winning streak (yeah me!) now in the 157th day, having achieved a 100% Expert rating for 223 puzzles that completed so far.

#HumanIngenuity #PuzzleSolving #ArtificialIntelligence

Postscript:

ChatGPT o1 solved the puzzle correctly while GPT-4o (above) did not.

Progress? I think so.