The Arc of AI

The modern AI era did not begin with ChatGPT. It began with a summer research workshop in New Hampshire in 1956, and the sixty years between are a story of hype, winter, and revival. Knowing the arc helps you place today's AI coding tools, and it is the context interviewers increasingly expect candidates to have.

The Dartmouth workshop

In the summer of 1956, John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon organized a two-month workshop at Dartmouth College on the working hypothesis 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." The term artificial intelligence was coined there. Almost nothing concrete was built that summer, but the field, its founders, and its founding promise were all established.

The symbolic era and the first winter

For the next two decades, most AI research was symbolic: logic, search, and rules. Programs played checkers, proved theorems, and parsed language. Expert systems like MYCIN (medical diagnosis) and DENDRAL (chemistry) encoded human knowledge as rules, and by the late 1970s the hype was enormous.

Then the money ran out. The Lighthill Report in the UK (1973) and the ALPAC report on machine translation convinced funders that the promises had outpaced results. Government and corporate funding collapsed, and the period from roughly 1974 to 1980 became known as the first AI winter.

A second winter followed in the late 1980s: the expert-system boom busted when rule-based systems proved brittle and impossible to maintain at scale, and funding collapsed again around 1987 to 1993.

The statistical revival

The revival was not symbolic. It was statistical machine learning: instead of hand-writing rules, researchers learned them from data. Neural networks, invented decades earlier (the perceptron in 1958, backpropagation in 1986), returned when computing power and data finally caught up.

The inflection point was 2012, when AlexNet won the ImageNet image-classification contest by a huge margin using deep neural networks trained on GPUs. Deep learning exploded across vision, speech, and language. Speech recognition, which had been built on hidden Markov models for decades, was replaced by deep networks within a few years.

The transformer era and LLMs

The decisive architectural step came in 2017 with the paper Attention Is All You Need, which introduced the transformer. Transformers dropped the sequential processing of earlier networks in favor of self-attention, making it possible to train vastly larger models in parallel. Every major language model since, from GPT to Claude to Gemini, is a transformer.

Scaling transformers on internet-scale text produced a surprising emergent capability: models that could complete and generate fluent text, reason, follow instructions, and write code. RLHF (reinforcement learning from human feedback) aligned those raw models into helpful assistants, and the public release of ChatGPT in November 2022 triggered the current era.

Today, in one paragraph

In 2026 the frontier is defined by reasoning-capable models and the harnesses around them: OpenAI's GPT-5.6 family (Sol, Terra, and Luna), Anthropic's Claude Fable 5 and Mythos 5, plus agentic coding tools like Claude Code, Codex, and OpenCode that turn a model into a worker. The rest of this section walks each thread, from the types of AI, to the history of GPT and Claude, to agents, harnesses, and why learning AI coding matters now.

Sources and further reading

• Dartmouth: the 1956 Summer Research Project
• Attention Is All You Need