The most interesting recent phenomenon in tech isn’t how ‘smart’ machines have become. It’s why almost nothing happened for so long, and then so much happened so quickly.

by Tim Leogrande, BSIT, MSCP, Ed.S.

🗓  JULY 30 2026 • 10 MIN READ

My friend Anthony has an uncle Sal who owned a TV repair shop in Brooklyn for thirty-one years. The business barely saw any change for the majority of that time. If a tube went out, Sal swapped out the tube. If a capacitor was bulging, he replaced the capacitor. Sal could have fallen asleep in 1978 and woken up in 1998, and he would have still been able to repair about ninety percent of what came through the door.

But in the mid-2000s, flat panel monitors and smart cable boxes became affordable and widely available. Because these technologies don’t have the same repairable components, the television repair business model of swapping out parts vanished almost overnight. The expertise that kept Sal and his family fed and sheltered for over three decades was now about as useful as the scrap metal in the junk TVs stacked in the back storage room. Anthony says “Sal always said the shop didn't die, it fell off a cliff that he didn't even know was there.”

I think about that shop often, because it's a superb, but sadly unfortunate, example of a naturally occurring phenomenon characterized by long lulls followed by an occasional sharp upheaval. In 1972, two paleontologists, Niles Eldredge and Stephen Jay Gould, developed a thesis asserting that Charles Darwin’s work was misleading people about how species changed. Darwin’s theory was gradualism, the idea that over multiple generations small changes would accumulate over time, slowly gaining in magnitude and complexity, forming a steady evolutionary slope. An elegant and easy-to-comprehend idea, to be sure. Too bad the fossil record completely contradicted it.

Fossils were discovered that depicted species in a certain way, and they often stayed that way for millions of years with basically no variation from one generation to the next, and then a species died off or suddenly morphed into an entirely different species with nothing resembling a transition anywhere in the boneyard. For a century paleontologists had been trying to explain the missing data that represented these transition periods as the result of not having enough fossils, or the inability to find enough fossils.

Eldredge and Gould argued that the missing data was actually an important part of the data, asserting “stasis is data.” So they came up with a new concept, punctuated equilibrium. It was the idea that species would sit in stasis for most of their existence and, if they did evolve, they would do so in brief spurts. In geological terms, brief didn't mean a second; brief could mean several hundred or several thousand years. But regardless of how long the evolution period was, the general idea caught on.

<aside> 💡

Suddenly, evolution wasn't a ramp. It was a flat plain, where species spent possibly millions of years walking across a low plateau only to suddenly take a leap to a higher plain. The missing transitions were no longer the result of a lack of knowledge. They were a key part of the terrain itself.

</aside>

Now, I'm no biologist. I'm an IT security professional and educator. But I've lived through enough media hype cycles to know that scientists are often quoted by people who are trying to make a point they have already decided to support. Punctuated equilibrium was no different, so from day one it was surrounded by controversy. Several leading evolutionary thinkers, including Richard Dawkins as the most outspoken, argued that Eldredge and Gould were merely putting a flashy new outfit on old-style gradualism. Their idea of “sudden” change was composed of the same small-scale genetic alterations, just taking place over shorter intervals. To some extent, these critics were right. But the debate isn’t over, and anyone who tells you it’s settled is overselling it.

Yet the concept of punctuated equilibrium crossed over from the fossil record into the social sciences, because it described a situation that many people believed, at least in the abstract, must be true of real-world organizations and technologies. Connie Gersick, a scholar who studies how teams and businesses change, discovered this pattern repeatedly:

“Groups spend a considerable amount of time in 'stable' periods of development (or 'deep structure') that tend to resist change, but they eventually restructure themselves in a brief and usually relatively dramatic ('revolutionary') change event, in which the accumulation of tensions eventually destroys the existing structure of the group.”

Similarly, Frank Baumgartner and Bryan Jones built a model of US policymaking on what they call “a simple observation: Political processes are generally characterized by stability and incrementalism, but occasionally they produce large-scale departures from the past.” Tobacco policy was the same, was the same, was the same, and then it jumped forward about a decade in one go. And if you’ve ever sat in on six policy meetings in which a change was rejected and the seventh in which the same change was accepted unanimously, you already know about punctuated equilibrium.

Artificial intelligence was Sal's repair shop for a long time. It was at rest. In a class I took as an undergrad many years ago, my professor explained in great detail why a computer would never be able to distinguish a photo of a wolf from one of a husky. This was, to be fair, a reasonable claim at the time. The field was aiming for general human-level intelligence after the Dartmouth Conference of 1956, and it had spent the intervening decades disappointing everyone, jumping from massive funding to periods of zero scientific progress that were colloquially known as “the AI winters.” In a word, stasis. A long stasis, well-documented enough to ruin careers.

And then in 2020, lightning struck. OpenAI’s GPT-3 surprised researchers. Instead of needing retraining for each task, it could learn from just a few examples in its prompt, tackling poetry, code, or conversation. This demonstrated that scaling models up — more size, more data — unlocked new abilities on its own, reshaping how the entire field pursued innovation.

The people building AI models talk about these giant technological leaps in pretty much the same nomenclature that Eldredge and Gould have used to talk about trilobites. The technical term for this is “emergent capabilities.” The models slowly scale up with no new capability appearing, their performance on a particular hard task hovering near zero. They then reach some specific level of scale, and the ability appears, in the researchers' words, “abruptly,” like a bolt out of the blue.

If you plot the development of AI on a timeline, it reveals this dynamic. For the longest time, computers couldn’t beat chess masters. Then Kasparov, arguably the greatest player of his time, suddenly fell to Deep Blue. Image generation had been plodding along at a snail’s pace for ages, then in the span of a year it went from disappointing streaks of color to family photos your grandmother would believe. LLMs wrote gibberish autocomplete for years, and then from one version to the next were writing legal memoranda.

Nothing, nothing, nothing, then boom. All at once.