Artificial intelligence has been declared dead and reborn several times in the past sixty years. Each cycle goes the same way: a breakthrough produces outsized optimism, funding floods in, the hard problems turn out harder than expected, and a winter sets in. Then the researchers who stayed produce something new.
I think the breakthrough we’re living through now is the biggest of them. To understand where we are, it helps to look at how we got here.
The first foundations (1950s to 1980s)
The field conventionally starts with Alan Turing’s 1950 paper asking whether machines can think. The early years were dominated by symbolic AI: explicit rules, hand-crafted knowledge bases, and logic engines. Expert systems like MYCIN and XCON showed that narrow, well-defined problems could be solved at expert level. Their brittleness was fatal, though. Add a new domain or change the rules slightly and the system fell apart.
The statistical turn (1990s and 2000s)
Statistical methods changed the approach. Instead of encoding knowledge by hand, you let the model learn patterns from data. Support vector machines, Bayesian networks, and early neural networks started beating symbolic approaches on perception tasks like speech recognition, handwriting recognition, and spam filtering.
This era produced reliable systems you could deploy. Recommendation engines, search ranking, and fraud detection, the invisible AI of the 2000s, ran on these methods. The ceiling was low, though: the models were good at narrow tasks and terrible at anything that needed open-ended reasoning.
Deep learning and perception (2012 to 2020)
AlexNet in 2012 marks the start of the modern era. A convolutional neural network trained on consumer GPUs won the ImageNet competition by a margin so large that the previous state of the art looked like it belonged to a different category.
Progress after that was fast: image recognition, speech synthesis, machine translation, game playing (AlphaGo, then AlphaZero), and protein folding (AlphaFold). Each application followed the same recipe of a large neural network, a large dataset, and enough compute.
The architecture that made this tractable was the transformer (Vaswani et al., 2017). Self-attention replaced recurrence for sequence modeling, and it scaled far better with more data and compute.
The generative era (2020 to now)
GPT-3 in 2020 was the first public demonstration that a single large language model could handle a startling range of tasks, including translation, summarization, code generation, and question answering, without being fine-tuned for any of them. It wasn’t perfect, but nothing before it had been that broad.
An arms race followed. GPT-4, Claude, Gemini, Llama, and Mistral improved faster than most researchers predicted. Generated text became coherent, then convincing, then genuinely useful for knowledge work. Image generation (Stable Diffusion, DALL-E, Midjourney) went from novelty to production tool in about eighteen months.
The mechanism behind all of it is surprisingly simple: predict the next token. With enough data and parameters, that objective produces models that have implicitly learned grammar, facts, reasoning patterns, and aesthetic judgment.
The agentic turn
The most recent shift is from models that respond to models that act. An agentic system uses a language model as its reasoning core and wraps it in a loop: observe the environment, choose an action, execute it, observe the result, and repeat.
That’s the architecture behind GitHub Copilot Workspace, Claude’s computer use, and a growing set of autonomous agents for coding, research, and other tasks. The underlying models are much the same. What has changed is the scaffolding around them and our understanding of how to prompt and chain model calls well.
The main challenges for agents are reliability (models hallucinate, and their mistakes compound over long loops), tool use (connecting the model to real APIs, file systems, and browsers), and safety (an agent that can act in the world needs constraints that a pure text generator doesn’t).
What comes next
A few threads seem likely to shape the next phase.
Multimodal reasoning is one. Current models handle text, images, audio, and video separately or in shallow combinations. The next generation should reason across them, for example watching a video and debugging the code shown in it.
Long-horizon planning is another. Today’s agents handle tasks that take dozens of steps. Tasks that need hundreds or thousands of interdependent decisions, like a multi-week software project or a scientific experiment, remain genuinely hard.
Then there’s compute efficiency. The environmental and economic cost of frontier models is real, and distillation, quantization, mixture-of-experts architectures, and better training curricula are all pushing toward more capability per watt.
Finally, alignment and interpretability. As models get more capable and more autonomous, we need to understand what they’re doing inside, beyond watching their outputs. Mechanistic interpretability is a young field, and I’d count it among the most important ones.
The distance from Turing’s question to today’s systems is remarkable. If you ask how far we’ve come, my answer is: far enough to be useful, and nowhere near finished.