While visiting family over the Independence Day holiday, I got drawn into assembling a jigsaw puzzle—a delightful waste of time. As the pieces came together, the striking parallels to Machine Learning (ML) became clear. Because the way we solve puzzles is so universal, it serves as a powerful  metaphor for understanding modern Artificial Intelligence and its blue collar cousin–Machine Learning.

Design / Assemble the Frame

Most jigsaw puzzlers start by assembling the frame. The frame pieces all have the single common characteristic of one completely straight edge. The puzzle designer chose the picture and the frame. The puzzle assemblers, both human and AI, use the straight-edge feature to assemble the edges.

Just as in the puzzle, the human designer of the AI model designs the picture and the edges. For this metaphor, consider an AI that builds and manages a portfolio. The human designer must determine what features to use to train the specific AI model. Where are the edges and what features go inside the picture?  For stocks, technical features such as daily price and trading volume are likely candidates to be inside the frame. Fundamental stock data, such as earnings, revenues, and revenue growth, should also be in the frame.

Should the frame include commodity prices? Stock prices for oil companies, refiners, retailers and pipeline transportation companies are tightly linked to the price of oil. The prices of copper and other metals and minerals strongly affect the price of mining stocks. Likewise, food producers are strongly influenced by the weather and the prices of pork, beef, and chickens. Should the frame include these prices?

Flows of money strongly drive aggregate demand and prices for stocks. Interest rates, dates for quarterly tax payments, and differentials in currency exchange rates drive money flows in and out of the stock market. Maybe money flows should be included also.

The domestic economy and the global economy significantly affect stock prices. GDP growth, changes in employment, elections, wars and plagues (COVID)  affect the economy. Maybe they should be included inside the frame.

At some point, the designer of the AI model says, “Enough, already. The features set for this model is growing too big.” Project factors such as data cost, data quality, project development time, and model execution cost and time for training and inference all work to limit the size of the picture frame.

Find Strong Features

In  the jigsaw puzzle below, notice all the lines as well as the edges. These visual lines, called features or patterns, are taking shape, especially compared to all the empty spaces between the patterns. Notice the features growing down from the top edge, which is their anchor. These visual patterns have high information content compared to the areas in the middle, which are mostly single-color puzzle pieces. Puzzle pieces with only one color have very low information content, so humans tend to ignore them until the puzzle is almost finished. AI models ignore them because the computation cost to analyze them exceeds their value…most of the time.

ML will assemble a set of data points into a jigsaw puzzle just like the humans did with this one. It will find all the patterns with high information content, mostly lines and interior edges, and assemble them first because their strong patterns make the assembly easier. Then humans and AI models tend to fill in the spaces in between the patterns at the last, because these areas of solid colors have low information content. These “low information” puzzle pieces need the anchor points and the context provided by the strong features to find their correct places.

Find Complex Features

By focusing on patterns of lines and edges, both humans and ML first find bigger, more complex, more valuable patterns. The broom is an example of a complex feature. It contains horizontal lines, vertical lines, and diagonal lines. It contains three major colors. When fully assembled, it is recognizable as a distinct, physical object with high information content and associations.

The newspaper is also a complex feature with strong associations. The written words provide powerful clues for assembling it and, like the broom, it is a distinct, physical object with strong associations. Given the other elements near the newspaper, you can probably already describe what the puzzle will reveal when assembled.

Can you find four cats in this jigsaw puzzle? Look for cat eyes. Next, what is in the blue tub? Yep, it is the outline of a big cat — papa cat reading the newspaper. Let’s link him to the stock ticker NVDA. Mama cat represents AMD, brother cat connects with MU, and sister cat represents TSM. If you know which cat is in the litter box, you may have a good guess which stock is about to “explode” upward.

Note the humans and ML did not even bother to assemble the solid color areas at this stage of puzzle completion. The information content in those areas is simply too low to provide significant predictive value, except for rare moments.

Create a Video Story from the Static Puzzle Picture

We want to predict the future. Time sequence is part of our lives, embedded in weather, stock prices, sports, elections, and practically everything, including our bodies. To survive, we must predict the future. To extend the metaphor of a jigsaw puzzle picture into a time sequence, we must tell a story. The specific job of this AI model is to observe a time sequence of snapshots and predict the next few snapshots in the sequence.

Some Patterns Change over Time

We already suggested one pattern that changes over time — which cat is in the litter box. So the complete set of patterns is not just this single snapshot of the litterbox in a closet. It is a video and the patterns persist both in a single snapshot and change over time in different snapshots. When the outline of the cat in the litter box is small, it is probably brother cat (MU) or sister cat (TSM). When the outline is a big cat reading something, it is probably papa cat, (NVDA) or mama cat (AMD). If the reading material is the New York Examiner, the cat is probably papa cat. Mama cat usually reads Glamour.

Life and the stock market have patterns that occur predictably every day, every week, every month, quarter, and year. Can you name a few predictable patterns? However, even predictable patterns can have variations in timing, size, and shape. The cats come in usually twice a day. But the exact timing and the length of stay may vary. The TP roll shrinks only just before a cat leaves, and papa uses the most.

Disruptive Patterns Occur Rarely, often in Clusters   

Rarely, a person comes in, picks up the broom, and sweeps the floor. This is a completely different pattern. It may happen only once a week, usually on Saturday, but sometimes on Friday, Sunday, or Monday. What else might happen when a person picks up the broom? Normally the roll of TP gets smaller, pretty gradually over time. When the person comes to sweep, he also replaces the TP roll, so it “explodes” bigger. Perhaps the person entering is like a report of big GDP growth, and the broom sweeping represents significant funds flows into the market, from foreign flows and increases in margin borrowing. These features may drive sudden, big growth in the TP roll, er… S&P 500 index.

The Data Dimension: Scaling from Snapshots to Market Microstructure

This single jigsaw puzzle contains roughly 500 static pieces. If we capture our puzzle’s timeline using one-minute interval snapshots—specifically to isolate the 5 to 10 minutes when the sweeper enters the frame—the system generates approximately 10,000 snapshots weekly. This minor exercise expands a simple 500-piece puzzle into an optimization matrix of 5 million distinct features per week.

In a similar way, a living financial market behaves like a continuous high-definition video stream rather than a frozen snapshot. Each trading day transforms human and corporate behaviors into millions of granular signals. ML algorithms do not interpret markets through visual intuition; they ingest high-dimensional numerical tables. To map the sheer scale of financial microstructure, consider taking one picture of the the US equity market, every trading day market, after market close. Like the series of cat puzzle pictures, these end-of-day (EOD) snapshots tell a powerful story over longer time horizons.

An institutional AI model can parse, weigh, and optimize across this 10.8-billion-point matrix in mere minutes. Large Language Models (LLMs) scale this computational capability even further, processing tens of trillions of parameter weights simultaneously. While human cognition struggles to map or visualize interactions beyond three spatial dimensions, algorithmic models natively thrive within multi-billion-point geometric spaces.

Carbon Intelligence vs. Silicon Intelligence

Human intelligence is built on a carbon substrate. The human body and brain operate in the physical world, using proteins instead of electrons to communicate. The human brain consumes about 20 watts of electricity—less than a dim lightbulb, yet it is the highest power consumption of any organ in our body.

Temporally, our brain cells are shockingly slow. A neuron fires at a speed of about 200 cycles per second (200 Hz). This is an unbelievably glacial pace compared to a silicon AI chip running at 10 GHz—which is millions of times faster but gulps down kilowatts of industrial energy.

Despite this slow speed, the human brain has evolved brilliant architectural shortcuts that allow it to calculate at about one petaflop, matching modern supercomputers. It does this through two main advantages:

• Functional Specialization:Humans have a dedicated visual processing unit built right into our biology. It identifies lines, edges, and puzzle patterns almost instantly without needing heavy computational effort. The human brain also has other specialized areas to control language and speech, hearing, and semi-autonomous body functions, such as heart rate, breathing, and balance.
• Associative Memory:Silicon computers keep their memory separate from their processors, forcing data to travel back and forth. The human brain, however, bakes intelligence directly into the physical connections between brain cells. We process data simply by how our cells are linked, allowing us to jump from one memory to a related idea instantly.

Silicon intelligence is largely hard-wired. Because it lacks our natural visual intuition and flexible brain structure, it must overcome its limitations with blazing, raw processing speed.

Jigsaw Puzzles are a Metaphor. AI models are not Human.

Humans navigate the world through the lens of emotions, desires, and evolutionary survival metrics—often mapped out in structures like Maslow’s hierarchy. Because our brains are wired to understand other living beings this way, we naturally anthropomorphize AI, projecting human feelings and intent onto lines of code.

In contrast, AI has no sense of time, no needs, and no desires. It experiences no stress over a market crash, nor anticipation for a breakout. It possesses a completely different kind of intelligence—one built not on emotion, but on the hyper-efficient assembly of trillions of data points. AI is the ultimate “data-driven” intelligence.

Human intelligence excels at defining the frame and imagining the story. Machine learning excels at connecting the billions of scattered pieces within it. While the jigsaw puzzle is just a metaphor, the powerful synergy between human vision and machine calculation is entirely real.