The Evolution of AI: Why We Keep Reinventing Intelligence
“Forests are not silent collections of individuals but living networks, where trees exchange nutrients, warnings, and memory through underground fungal systems. This “Wood Wide Web” challenges our habit of equating intelligence with brains, suggesting instead that awareness can emerge from relationship, cooperation, and slow time.
In 1950, a British mathematician sat down to write a paper that would haunt us for the next 75 years.
Alan Turing didn't ask "can machines think?", he thought that question was meaningless. Instead, he proposed something simpler and more devious: a test where a human judge converses with both a human and a machine, unable to see either, and tries to identify which is which.
If the judge can't tell the difference reliably, does it matter whether the machine "really" thinks?
This wasn't just a technical proposal. It was a philosophical trap. Because whether you answer yes or no reveals what you believe intelligence actually is.
And we're still arguing about it.
But what Turing couldn't have predicted is that the path from his imitation game to ChatGPT wouldn't be a straight line of steady progress. It would be a strange journey of abandoned approaches, forgotten winters, unexpected revivals, and solutions that worked for reasons nobody fully understood.
This is the story of how we kept trying to build intelligence and kept discovering we didn't know what we were building.
🜏
The Symbolic Dream (1956-1974): If We Can Think It, We Can Code It
In the summer of 1956, a small group of researchers gathered at Dartmouth College for a workshop. They had an ambitious goal: spend eight weeks figuring out how to make machines intelligent.
They called their field "artificial intelligence." They thought they'd solve it in a generation.
Their approach seemed logical: human thinking uses language, logic, and symbols. Mathematics is symbolic manipulation. Computers manipulate symbols. Therefore, if we just formalize the rules of thinking, computers should be able to think.
This was called "symbolic AI" or "Good Old-Fashioned AI" (GOFAI).
Early results were intoxicating. In 1959, Arthur Samuel created a checkers program that learned to beat its creator. In 1961, James Slagle's SAINT program could solve calculus problems at college freshman level. By 1965, Joseph Weizenbaum's ELIZA could conduct conversations that fooled people into thinking they were talking to a therapist.
The researchers were so optimistic that Herbert Simon predicted in 1965: "Machines will be capable, within twenty years, of doing any work a man can do."
That was 1965. He was off by... well, we're still waiting.
What went wrong?
The real world turned out to be vastly more complicated than formal logic could capture. Sure, you could program a computer to prove mathematical theorems or play chess by encoding explicit rules. But how do you encode "recognizing a cat"? Or "understanding a joke"? Or "knowing when someone's being sarcastic"?
Symbolic AI could handle well-defined problems with clear rules. But most of intelligence isn't like that. Most of intelligence is pattern recognition, context sensitivity, dealing with ambiguity and incomplete information.
And nobody knew how to formalize that.
By the mid-1970s, funding dried up. Governments and corporations realized the grand promises weren't being delivered. The field entered what's now called the "first AI winter", a period where even mentioning artificial intelligence could end your research career.
But the dream didn't die. It went underground and changed shape.
🜏
The Expert Systems Boom (1980s): Intelligence as Database
In the 1980s, AI resurged with a different approach: if we can't make machines think like humans, maybe we can make them think like experts.
Expert systems were essentially sophisticated if-then rules encoded from human specialists. A doctor's diagnostic process could be broken down into decision trees: if symptom A and symptom B but not symptom C, then probably diagnosis X.
MYCIN, developed at Stanford, could diagnose blood infections as accurately as human specialists. XCON, created for Digital Equipment Corporation, configured computer systems so well it saved the company millions annually.
This looked like practical AI. Not philosophical speculation about machine consciousness, but useful tools that augmented human expertise.
For a brief moment, it seemed we'd found the right path: don't simulate general intelligence, just capture specific expertise in specific domains.
Then reality intruded again.
Expert systems were brittle. They only worked within narrow domains. Expand the scope even slightly, and they broke. They couldn't learn from experience or adapt to new situations. Every edge case required explicit programming.
Worse, extracting knowledge from experts turned out to be extraordinarily difficult. Experts often don't know how they know things. Their expertise is intuitive, accumulated through thousands of cases, impossible to fully articulate as rules.
By the late 1980s, the second AI winter arrived. Companies that had invested heavily in expert systems abandoned them. AI became, again, a term associated with overpromise and underdelivery.
But something else was brewing, something that would eventually change everything.
🜏
The Neural Network Renaissance: Learning Instead of Programming
The strange part of AI history is, that neural networks were invented in the 1940s and 1950s, right alongside symbolic AI.
In 1943, Warren McCulloch and Walter Pitts created a mathematical model of how neurons might compute. In 1958, Frank Rosenblatt built the Perceptron, a machine that could learn to recognize simple patterns.
The Perceptron was supposed to revolutionize AI. Instead, it nearly killed neural networks for decades.
In 1969, Marvin Minsky and Seymour Papert published a book showing the Perceptron's fundamental limitations. It could only learn linearly separable patterns, essentially, it was too simple to solve most interesting problems.
Neural networks fell out of favor just as symbolic AI was ascending.
But a few researchers kept working on them quietly. In the 1980s, several breakthroughs emerged:
Backpropagation: An algorithm for training multi-layer neural networks by calculating how much each connection contributed to errors. This solved the problem Minsky identified, deeper networks could learn non-linear patterns.
Recurrent networks: Architectures where information could flow in cycles, enabling networks to process sequences and maintain something like short-term memory.
Convolutional networks: Structures inspired by the visual cortex, designed specifically for processing images.
Yet even with these advances, neural networks remained mostly academic curiosities through the 1990s and early 2000s. They required enormous computational power and massive datasets, neither of which were readily available.
And most researchers had been burned by AI hype before. Why invest in an approach that might hit the same walls symbolic AI and expert systems had hit?
What changed wasn't the ideas. What changed was the infrastructure.
🜏
The Perfect Storm (2010s): When Everything Converged
Around 2012, several things happened simultaneously that nobody predicted would align:
Moore's Law delivered. Computers had become exponentially more powerful. GPUs, originally designed for rendering video game graphics, turned out to be perfect for the parallel computations neural networks required.
The internet generated data. Billions of images, texts, videos, and interactions created training datasets of unprecedented scale. Neural networks are data-hungry; suddenly there was an all-you-can-eat buffet.
Algorithms improved. Researchers figured out how to train deeper networks without them collapsing. Dropout, batch normalization, better activation functions, technical advances that individually seemed minor but collectively enabled step-change improvements.
In 2012, a neural network called AlexNet won an image recognition competition by a margin that shocked the computer vision community. It cut error rates nearly in half compared to traditional approaches.
That was the moment. Suddenly, everyone was paying attention again.
Within a few years:
Computer vision systems surpassed human-level performance on specific tasks
Speech recognition became reliable enough for commercial deployment
Machine translation improved from barely usable to genuinely helpful
Neural networks started beating human champions at increasingly complex games
But this wasn't symbolic AI finally working or expert systems scaled up. This was something genuinely different: systems that learned patterns from data rather than following programmed rules.
And nobody fully understood why they worked so well.
🜏
The Transformer Revolution: Attention Is All You Need
In 2017, researchers at Google published a paper with an audacious title: "Attention Is All You Need."
They proposed a neural network architecture called the Transformer that abandoned the sequential processing of earlier models. Instead of reading text word by word like humans do, Transformers could process entire sequences simultaneously using something called "attention mechanisms", essentially, learning which parts of input are most relevant to which other parts.
At the time, it seemed like a technical improvement, not a revolution.
Within five years, Transformers became the foundation for essentially every major AI breakthrough:
GPT (Generative Pre-trained Transformer) could write coherent long-form text that often seemed indistinguishable from human writing.
BERT revolutionized natural language understanding tasks.
DALL-E and Stable Diffusion could generate images from text descriptions.
AlphaFold solved protein folding, a biology problem that had stumped researchers for decades.
What made Transformers different?
Scale. They could be trained on vastly larger datasets than previous architectures. GPT-3 was trained on hundreds of billions of words. GPT-4's training details are secret, but the scale is presumably even larger.
And with scale came something unexpected: emergent capabilities. Properties that weren't explicitly programmed but appeared spontaneously as the models grew larger and saw more data.
GPT-3 could perform arithmetic despite never being explicitly taught math. It could translate languages it wasn't trained to translate. It could write code, compose poetry, explain jokes, engage in logical reasoning.
Nobody predicted this. The researchers themselves were surprised.
And that surprise is important, because it means we've built something we don't fully understand.
🜏
What Actually Evolved: The Question Nobody's Answering
Here's the uncomfortable truth about AI evolution: we still don't know why our current approaches work as well as they do.
We can describe what neural networks do, adjust billions of numerical parameters to minimize prediction errors. We can measure their performance, often superhuman on specific benchmarks.
But we can't fully explain how they generate the outputs they generate.
This is radically different from traditional software. If a calculator gives you the wrong answer, you can trace through the code and find the bug. If GPT-4 generates a plausible-sounding but completely false statement, you can't trace through 100+ billion parameters to find where the "hallucination" originated.
The system is working, but it's not working the way we designed it to work. It's working the way it learned to work.
And that raises questions we're spectacularly unprepared to answer:
When GPT-4 writes a poem about loss and grief, did it "understand" those emotions? Or did it just identify statistical patterns in text written by humans who felt those emotions?
When AlphaGo discovered novel strategies that human Go players hadn't conceived in 3,000 years of playing the game, was that creativity? Or sophisticated pattern matching?
When do we stop calling it "artificial" intelligence and start calling it just... intelligence?
🜏
The Evolution Nobody Planned
Looking back at 75 years of AI development, the pattern is clear: we keep succeeding for reasons we didn't predict and failing for reasons we didn't understand.
Symbolic AI failed because intelligence isn't primarily logical reasoning.
Expert systems failed because expertise isn't primarily explicit rules.
Neural networks succeeded not because we understood intelligence better, but because we stopped trying to program intelligence directly and instead built systems that could learn patterns from examples.
But this raises a disturbing possibility: maybe we've succeeded by accident.
Maybe our current AI systems work well not because we've discovered fundamental principles of intelligence, but because we've stumbled onto statistical methods that approximate intelligent behavior in domains with enough data.
And maybe, this is the part that keeps me up at night, we're building increasingly powerful systems without understanding what we're building or where it leads.
🜏
Evolution Toward What?
What we know about the trajectory:
AI systems are becoming more general. Early systems solved one narrow task. Modern systems can handle thousands of tasks with the same underlying model.
AI systems are becoming more autonomous. Early systems required explicit instructions for every step. Modern systems can break down complex goals and pursue multi-step strategies.
AI systems are becoming more opaque. Early systems' decision-making was traceable. Modern systems operate as "black boxes" where even their creators can't fully explain specific outputs.
AI systems are becoming more surprising. Early systems did exactly what they were programmed to do. Modern systems exhibit emergent capabilities nobody predicted.
Where does this trajectory lead?
Some researchers think we're approaching artificial general intelligence, systems that can match or exceed human cognitive abilities across all domains. They predict this within decades.
Others think we're hitting fundamental limitations we haven't recognized yet. That our current approaches will plateau, and we'll need entirely new paradigms.
The honest answer is: we don't know.
Because the evolution of AI hasn't followed a master plan. It's been a series of surprising discoveries, abandoned dead ends, and unexpected revivals of old ideas.
We're not steering this evolution so much as surfing it.
And surfing only works if you're paying very close attention to the wave beneath you.
🜏
The Evolution We Should Be Having
What troubles me about AI evolution discourse is, that we focus almost entirely on capabilities and almost not at all on purpose.
The conversation is: Can AI write better than humans? Can it diagnose diseases? Can it drive cars? Can it generate art? Can it code? Can it reason?
The conversation should be: What do we want AI for? What problems should it solve? What should it empower humans to do? What shouldn't it do at all?
Because capabilities are trajectory. But purpose is direction.
And right now, AI evolution is being directed primarily by:
Whatever problems generate the most profit
Whatever capabilities are easiest to achieve with current methods
Whatever impresses investors and attracts funding
Whatever militaries think provides strategic advantage
These aren't bad criteria necessarily. But they're not sufficient.
Because evolution, biological evolution, doesn't have purpose. It has selection pressure. Whatever survives and reproduces, continues. Whatever doesn't, disappears.
We're creating selection pressure for AI evolution through our choices about funding, deployment, regulation, and adoption.
And mostly, we're selecting for: impressive demos, economic efficiency, and competitive advantage.
Are those the right selection pressures for technology that might eventually exceed human capabilities in most domains?
I don't know. But I know we should be asking that question much more seriously than we are.
🜏
What History Suggests Happens Next
If the pattern holds (and patterns in technological evolution often do) we should expect the following:
1. The current approaches will hit limitations we don't currently recognize. Just like symbolic AI and expert systems before them, neural networks will eventually encounter problems they can't solve, barriers they can't cross with just more data and more compute.
2. Something currently dismissed as impractical will become the foundation of the next wave. Just like neural networks were dismissed for decades before becoming dominant, there's probably some approach currently languishing in obscurity that will revolutionize the field in 15 years.
3. We'll be surprised by what works and what doesn't. The history of AI is a history of confident predictions proven wrong. The things experts said would be easy turned out to be hard. The things they thought were impossible turned out to be straightforward.
4. The next breakthrough won't come from the center of the field. It never does. Backpropagation came from psychology research. Transformers came from machine translation work. The next paradigm shift will probably come from some unexpected corner.
5. We'll overestimate short-term progress and underestimate long-term impact. This has been true for every technological revolution. The hype cycle overshoots early, crashes during the "trough of disillusionment," and then the real transformation happens slowly, quietly, over decades.
But here's what's different this time:
The stakes are higher. Previous AI failures inconvenienced researchers and disappointed investors. If we fail to align advanced AI systems with human values, the consequences could be irreversible.
The pace is faster. It took symbolic AI two decades to reach its limits. Neural networks went from breakthrough to dominance in less than a decade. The next paradigm might transform the field in years.
The systems are more opaque. We understood why symbolic AI and expert systems failed. We don't fully understand why current systems succeed, which means we might not recognize failure until it's catastrophic.
This isn't inevitable progress toward some predetermined destination. It's evolution. Unpredictable, contingent, and potentially dangerous if we're not careful about the selection pressures we create.
🜏
Where This Leaves Me
I started exploring AI evolution thinking I'd find a clear narrative: from Turing's imitation game to modern language models, a story of steady progress toward machine intelligence.
What I found instead was something messier and more interesting: a history of brilliant ideas failing, discarded approaches being revived, and accidental discoveries changing everything.
We didn't engineer our way to modern AI through careful planning. We stumbled our way here through trial and error, with each generation building on the unexpected successes and forgotten failures of previous attempts.
And now we're at an inflection point. The systems we've built are powerful enough to transform society, opaque enough that we don't fully understand them, and improving fast enough that predictions become obsolete within months.
What happens next depends on choices we're making right now:
How we fund research
What we regulate
What we deploy
What questions we ask
What values we embed in development
Evolution doesn't have foresight. It doesn't plan. It just selects for whatever works in the current environment.
But we're not evolution. We can think about where this is going. We can choose what to select for.
The question is whether we'll use that capacity.
Or whether we'll just keep building more powerful systems because we can, without asking whether we should.
The evolution of AI isn't written yet. But we're writing it, one decision at a time.
And the draft we're producing right now will determine whether the next chapter is triumph or tragedy.
— N.H.
Further Reading:
Stuart Russell & Peter Norvig - Artificial Intelligence: A Modern Approach (comprehensive technical history)
Pamela McCorduck - Machines Who Think (historical narrative)
The Dartmouth Conference proposal (1956) - founding document of AI
"Attention Is All You Need" paper (Vaswani et al., 2017)
Yann LeCun, Yoshua Bengio, Geoffrey Hinton - "Deep Learning" (Nature, 2015)