Yesterday’s Magic
Two years ago, I wrote a Seed Scapes piece called Redefining Search. The premise was that artificial intelligence was beginning to change what the word “search” meant.
For most of the internet era, searching meant finding something that already existed. A webpage. A fact. A restaurant. A product. An answer someone had already written down. Google became extraordinarily good at helping us find the needle in an ever-expanding digital haystack.
LLMs introduced something different. Instead of merely retrieving an existing answer, they could assemble facts, perspectives and ideas into an answer that perhaps no one had ever written before.
I described the distinction this way: traditional search was about finding a needle in a haystack. AI might let us create a needle that never existed.
To test the idea, naturally, I consulted a six-pound toy poodle.
His name is Snowflake.
Snowflake presented a mystery. During the day, he follows the wife almost everywhere. The husband can call him from ten feet away and receive roughly the same response he would get from calling the refrigerator. Strangely, if the wife gives Snowflake permission, he will sometimes go visit the husband. Apparently, the husband is not merely second choice; he requires management approval.
But nighttime changes everything. When the wife goes to bed, Snowflake races into the bedroom, digs at the husband’s blanket until admitted, crawls underneath and settles into his armpit for the night.
In 2024, I gave that behavior to GPT and asked it to analyze Snowflake while doing an impression of Cesar Millan. Not because Cesar Millan had actually analyzed Snowflake, obviously, but because the fictional expert gave the experiment a useful lens. GPT talked about pack structure, leadership, security and the different roles Snowflake assigned to each human. It even prescribed homework: more walks for him, more cuddles for her.
At the time, the answer felt extraordinary. It was thoughtful, entertaining and unlike anything I could have gotten by typing keywords into Google.
I recently reread it.
It wasn’t nearly as good as I remembered.
The answer was pleasant, but generic. Ancestral pack instincts. The wife as Alpha. The husband as secondary leader receiving his nocturnal due. Swap in another dog, another couple and perhaps another species with a pulse, and much of the answer could have survived intact. It was a little like a horoscope with a Spanish accent.
That may be one of the more interesting measures of technological progress: yesterday’s magic becoming today’s mediocrity.
The original piece ended by comparing the AI of 2024 to Pong and noting that it took 31 years for Pong to evolve into Call of Duty. Imagine, I wondered, what AI might look like after a similar period of development.
Apparently, Snowflake wasn’t willing to wait that long.
He gave us more data.
And the machine changed.
The Natural Experiment
Since then, Snowflake’s household handed the universe a much better experiment.
The husband and wife are in the middle of a large interstate move — the kind involving containers, staging, storage, movers and that peculiar period when nearly everything they own seems to be either in the wrong house or the wrong state.
At one point, the husband drove a car down to the new house, planning to fly back. The wife and their son stayed behind with Snowflake.
Snowflake fell apart.
For days, he produced horrible, foul-smelling, soft messes on and around his wee-wee pad. Nothing obvious had changed with his food. There was no apparent illness. It became bad enough that “time for the vet” entered the conversation.
The obvious explanation was stress.
Hardly a Sherlock Holmes moment. His home was slowly disappearing into boxes and containers. Furniture was moving. Routines were changing. Dogs don’t receive closing documents, so presumably all Snowflake knew was that someone was dismantling his universe one room at a time.
Then the husband came home.
Snowflake’s digestive problems stopped almost immediately.
Which would have been a nice confirmation of the stress-from-chaos theory except for one inconvenient detail: the husband immediately made the house more chaotic. Furniture moved. Rooms changed. Staging accelerated. He spent his time dragging things around in preparation for the movers. By almost any objective measure, the physical environment became less stable after he returned.
The dog got better anyway.
Something about the explanation was wrong.
So the experiment was rerun.
Same dog. Same humans. Same basic mystery. The fictional Cesar Millan even returned, not because anyone suddenly needed celebrity dog-training advice, but because keeping the same character preserved something resembling a control.
But this time, instead of going back to GPT, the expanded story was handed to Claude Fable.
There was a reason for that.
The point wasn’t to stage a GPT-versus-Claude cage match or determine which company deserved a trophy. The idea was almost the opposite: go overpowered. Take a small domestic mystery that would seem almost comically undeserving of frontier-level AI, give a much more capable model far more information than a traditional search query would ever tolerate, and see what it did with the excess.
Don’t optimize the prompt.
Don’t strip out the details.
Give it the whole damn story.
The 2026 model began with the old observation but reached it differently. Snowflake wasn’t necessarily ranking the humans by affection, it suggested. He was assigning roles. The wife represented daytime security. The husband was the doorman, outside-trip coordinator and treat dispenser. Then, when her shift ended at bedtime, Snowflake switched jobs.
As fictional Cesar put it, the husband became a “warm, breathing cave wall.”
Finally, a professional designation worthy of putting on a business card.
But the interesting part wasn’t the joke or even the conclusion. It was how the model got there. Instead of placing a generic pack hierarchy over Snowflake’s behavior, it used the strange little details of his actual life: the permission structure, the treat economy, the timing of his allegiance shift, the husband’s absence, the digestive episode and, critically, what happened when he returned.
The humans had framed the problem as chaos.
The model noticed that their own evidence contradicted them.
If household chaos caused Snowflake’s gastrointestinal problems, then increasing the chaos should have made him worse. Instead, the husband returned, increased the chaos and Snowflake immediately improved.
The machine offered another hypothesis: perhaps the important variable wasn’t environmental stability at all. Perhaps it was pack integrity.
The husband’s absence had changed the social environment. His return restored it.
Or, as fictional Cesar rather memorably put it:
“The dog’s intestines just told you otherwise.”
Was that the answer?
Maybe.
But that “maybe” turns out to be one of the most important parts of the story.
Tell Me the Whole Story
Claude Fable didn’t diagnose Snowflake. Nor should it have.
It pointed out that stress-related gastrointestinal problems are real in dogs and that Snowflake’s symptoms and their timing were consistent with that possibility. But recurring or worsening episodes still warranted a veterinarian. Parasites, infection, diet and other medical causes don’t disappear merely because an AI has constructed an elegant behavioral theory. For a six-pound dog, dehydration isn’t something to philosophize about for very long.
In other words, the machine generated a hypothesis from evidence the humans hadn’t realized was evidence, showed why the hypothesis fit, and then identified where its confidence should end.
That distinction matters.
The remarkable thing wasn’t that AI had discovered what was wrong with Snowflake.
It was that AI noticed something the humans had mislabeled inside their own question.
And that’s when I realized this wasn’t really a story about how much AI had improved in two years. It was another chapter in Redefining Search.
For most of the digital era, computers have trained us to compress our questions.
Google rewarded keywords. Remove the story. Strip away irrelevant detail. Identify the important nouns. We became amateur database-query designers because the machine required us to know what mattered before we asked it to help us discover what mattered.
Even the early days of LLMs carried some of that legacy. “Prompt engineering” became an entire cottage industry. Assign the model a role. Carefully structure the request. Specify the desired output. Remove ambiguity. Feed the machine exactly the information it needs.
There is nothing inherently wrong with that. Clear questions still tend to produce clear answers.
But increasingly capable reasoning models introduce a fascinating inversion.
What if the details we were trained to remove are becoming the most valuable part of the question?
Imagine telling a friend you’re thinking about selling your company. You mention that you received an offer three years ago and turned it down. Revenue is higher now, but you’re enjoying the business less. Your son recently graduated from college and doesn’t want anything to do with it. Your wife keeps talking about traveling while you’re still young enough to enjoy it. A competitor recently sold at nine times EBITDA while another couldn’t find a buyer. And, strangely, the division now producing the most profit is the one you almost shut down five years ago.
Traditional search hears noise.
A reasoning model can hear variables.
The son’s lack of interest may matter. The wife’s desire to travel may matter. The failed competitor sale may matter. The unwanted division may matter. Or none of them may matter. The important capability isn’t that the model automatically knows which detail is decisive.
It’s that the model can afford to consider all of them.
That is fundamentally different from a search box.
If I already know which five facts matter, traditional search is extraordinary.
If I don’t know which five facts matter, I have a different problem.
And most of life’s interesting questions look a lot more like the second kind.
Business decisions, relationships, career changes, investments, negotiations and even mysterious little dogs rarely arrive as neatly structured prompts. They arrive as stories filled with contradictions, chronology, emotions, false assumptions, half-remembered events and details that seem irrelevant until suddenly they aren’t.
Wisdom often consists of knowing which supposedly irrelevant detail isn’t irrelevant.
For decades, computers forced humans to learn how machines wanted questions asked.
We may now be watching machines learn how humans naturally tell stories.
The Detail You Would Have Deleted
That makes me rethink my Pong analogy from 2024.
I imagined a fairly conventional technological progression. Pong becomes Call of Duty. Crude capability becomes sophisticated capability. AI gets faster, smarter and more knowledgeable until the primitive thing we were playing with becomes almost unrecognizable.
Two years later, I’m increasingly convinced I made the analogy too conservatively — not because we’ve already reached Call of Duty, but because I assumed I understood what game we were building.
I thought better AI would give us better answers.
Increasingly, I think its stranger contribution may be helping us discover that we were leaving half the question out.
The old search paradigm punished context. Every unnecessary word increased the distance between the query and the result. We learned to delete details until only the searchable skeleton remained.
The emerging paradigm can work in the opposite direction. Give it chronology. Give it contradictions. Tell it what happened before the thing happened. Tell it what you assumed. Tell it what changed afterward. Include the odd fact that doesn’t seem to belong. Let the machine decide whether it matters instead of making that decision before the conversation begins.
That doesn’t make AI omniscient. In some ways, it makes skepticism even more important. A machine capable of constructing a compelling explanation is also capable of constructing a compelling explanation that happens to be wrong. Elegant narratives are seductive whether they come from humans or silicon. Hypotheses still need evidence. Medical questions still need doctors and veterinarians. Financial questions still contain risk. Reality retains veto power.
But there is an enormous difference between outsourcing a decision and expanding the number of variables we can think about while making one.
That may be where the real transformation is happening.
The answer to Snowflake’s mystery may have been hiding inside details that once would have been considered clutter: who gives the treats, who opens the door, who needs permission from whom, who sleeps where, who disappeared, when the symptoms began, when they stopped, and the seemingly contradictory fact that the house became more chaotic precisely when the dog became healthier.
A traditional query would have stripped most of that away.
“Toy poodle diarrhea during move.”
Search.
Useful information would undoubtedly appear.
But the interesting question wasn’t really about diarrhea. It wasn’t even necessarily about moving.
The interesting question was which part of the story was actually the question.
Snowflake, for the record, may have known the answer all along. He’d been announcing his attachment nightly, from the husband’s side of the bed, for years. The humans just hadn’t considered that information relevant to what was happening on his wee-wee pad.
The machine did.
In 2024, I argued that AI might let us create needles that never existed.
Two years later, I think something even stranger is happening.
We can increasingly hand it the whole haystack.
And sometimes the needle turns out to be the detail we would have deleted.
🌱 Seed Thought: The future of search may not be finding the right answer faster. It may be discovering which part of the question you didn’t know mattered.
