In the summer of 1956, a small group of mathematicians and engineers gathered at Dartmouth College for a workshop whose proposal contained one of the most consequential sentences in the history of technology: that “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” John McCarthy, who coined the term “artificial intelligence” for that gathering, expected meaningful progress in a matter of months. The optimism was not naive so much as premature. It set a pattern that has repeated, with remarkable fidelity, for nearly seventy years: a bold claim, a rush of progress, a collision with reality, and then a quieter period of consolidation from which the field always emerged stronger than before.
The symbolic era and the first winters
The decades following Dartmouth belonged to symbolic AI, the conviction that intelligence could be captured in explicit rules and logical manipulation of symbols. The approach produced genuine landmarks. Programs proved mathematical theorems, played respectable chess, and parsed constrained slices of language. By the 1970s and 1980s it matured into expert systems: hand-built knowledge bases that encoded the judgment of physicians, geologists, and engineers into thousands of if-then rules. For a while these systems were a commercial industry unto themselves.
But the cracks were structural. Knowledge had to be elicited and entered by hand, rule by rule, and the systems were brittle at the edges of their narrow domains. They could not learn, could not generalize, and grew impossible to maintain as they scaled. When funding bodies measured the results against the promises, the gap was stark. Two distinct downturns followed — one in the mid-1970s after critical government reviews, another in the late 1980s as the expert-systems market collapsed — periods now remembered as the “AI winters.” The honest lesson of those winters was not that the ideas were worthless. It was that confident rhetoric had outrun demonstrable capability, and that the field paid for the difference in lost credibility and funding.
The statistical turn
What pulled AI out of the cold was a shift in philosophy. Rather than telling machines the rules, researchers began letting them infer patterns from data. Through the 1990s and 2000s, statistical machine learning quietly took over the practical frontier — probabilistic models, support vector machines, decision-tree ensembles, and the steady accumulation of digitized data to train them on. This era was less photogenic than the grand promises of the symbolic age, and that was precisely its strength. It traded sweeping claims for measurable accuracy on concrete tasks: spam filtering, fraud detection, recommendation, search. The capability floor rose not because anyone declared victory, but because the methods worked and could be evaluated.
The deep learning breakthrough
The next inflection arrived with force in 2012, when a neural network called AlexNet won the ImageNet image-recognition competition by a margin that startled the field. Neural networks were not new — the underlying ideas dated to the 1980s — but three things had finally converged: very large labeled datasets, the parallel computing power of graphics processors, and refinements in how deep networks were trained. Suddenly, learned representations outperformed decades of hand-engineered features. Within a few years, deep learning had reshaped computer vision, speech recognition, and machine translation. The pattern of the field held: a long, unglamorous build-up, then a sharp leap that looked sudden only to those who hadn’t been watching.
Transformers and the language model age
In 2017, a paper titled “Attention Is All You Need” introduced the transformer architecture, which dispensed with the sequential bottlenecks of earlier models and proved extraordinarily good at scaling. The consequences unfolded across the following years through successive generations of large language models, which exhibited an uncomfortable but undeniable property: as they grew in parameters and training data, they acquired capabilities — translation, summarization, reasoning, code generation — that no one had explicitly programmed. By the early 2020s, these systems had moved from research artifacts into the hands of millions of ordinary users.
Here the recurring lesson asserted itself again, in a new key. The hype around language models has at times been every bit as overheated as the promises of 1956, and a sober observer should expect a correction wherever expectation has outrun evidence. Yet the floor, once more, has risen and held. A model that can draft a coherent argument, write working software, or interpret an unfamiliar document is now ordinary infrastructure, not a demonstration.
From tools to collaborators
The present moment is defined by a subtler shift than raw capability. For most of its history, AI was something you operated: you posed a query, ran a classifier, called a model. What is changing now is the degree of autonomy. Agentic systems can hold a goal across many steps, decide which actions to take, call tools and other models, observe the results, and adjust — closing the loop between intention and execution with diminishing human intervention. The most striking application of this is reflexive. AI systems are beginning to participate in their own construction:
- writing, testing, and debugging the code that other systems run on;
- generating and curating training data, and evaluating model outputs at scale;
- searching the space of architectures and configurations rather than relying solely on human intuition;
- orchestrating fleets of specialized models toward a single objective.
This is not the arrival of general intelligence, and it should not be sold as such — that would simply be the Dartmouth error in modern dress. It is something more grounded: a change in what the human does. The work moves up a level of abstraction, from operating the machine to directing a collaborator that handles much of the building itself.
The next phase
If the long arc from Dartmouth teaches anything, it is to be measured about the claims and serious about the trajectory. Each era over-promised; each one also left the capability floor permanently higher than it found it. At Artificial Heights, our thesis is simply the next step along that line — not artificial general intelligence, not a revolution announced in advance, but AI that builds AI: systems that help design, construct, and orchestrate the systems that come after them. We believe that is where the next durable gains will come from, and we intend to earn the claim by delivering it rather than declaring it.