The Long 2023
Periods in history are defined by events rather than calendars, so it is fair to stretch 2023 back a few weeks to the launch of ChatGPT.
Before that launch, generative AI was largely specialist territory. Capable language models existed, GPT-3 most notably, but reaching them generally required an API and some technical confidence. There was still real debate within the industry about whether large language models represented a fundamental breakthrough or merely an interesting research direction.
ChatGPT settled that debate by putting a conversational interface in front of anyone with a browser. Students wrote essays with it, programmers debugged with it, and businesses began working out where it fitted. The underlying technology had existed before November 2022. What changed was that ordinary people could use it.
The response from competitors was immediate. Google, which had been cautious about releasing conversational systems, launched Bard, later renamed Gemini. Microsoft, having invested heavily in OpenAI, pushed the technology into Bing and Office. Anthropic, founded by former OpenAI researchers, released Claude. Meta released its LLaMA models under relatively permissive terms, seeding an ecosystem of open-weight models whose parameters were public even though their training data and methods generally were not.
In March 2023 OpenAI released GPT-4, a substantial step up in capability. It handled more difficult tasks, accepted images as well as text, and was much more useful across a range of professional work. The lesson of the first few months after ChatGPT seemed straightforward: larger models, trained with more data and more computation, became more capable in surprisingly general ways.
The obvious response was to scale.
That response produced excitement and unease in roughly equal measure. Also in March 2023, the Future of Life Institute published an open letter calling for a six-month pause on training systems more powerful than GPT-4. Critics found the proposal impractical and questioned the motives of some of its signatories, but the letter registered a growing concern that capabilities were advancing faster than anyone understood their consequences.
That concern sharpened when Geoffrey Hinton left Google in May 2023 so that he could speak more freely about the risks of the technology he had done so much to develop. Yoshua Bengio, another of the three researchers awarded the Turing Award for the deep-learning revolution, began making similar arguments for stronger governance.
The legal challenges arrived in parallel. Authors, artists, programmers, publishers and other rights holders filed lawsuits alleging that models had been trained on their work without permission and thus violated copyright law. Despite clear analogies to other technologies courts have found to be fair use, these copyright suits became a funnel for people to channel their general anxieties about the AI revolution. That litigation is ongoing.
The Long 2024
If the Long 2023 was about scale revealing capability, the Long 2024 was about engineering and optimization. The question shifted from how large a model could be built to how much could be extracted from the models already available.
Systems rather than models. One lesson was that the model itself was only part of the product.
Retrieval-augmented generation, usually called RAG, connected models to external collections of documents. Instead of relying entirely on information absorbed during training, a system could search a database, retrieve relevant material, and give it to the model as part of the prompt. This made it possible to build systems around current information or proprietary corporate data without retraining the underlying model. It also provided some basis for attribution.
Contrary to early claims (by legal technology vendors in particular), RAG did not solve hallucination. A model could retrieve the wrong segment of a document, misunderstand the right ones, or simply say something its sources did not support. RAG made LLMs more useful for finding out about the world because a model no longer had to contain (or guess at) all of the information it might need. LLMs now find and interpret online information in real time.
The second development was agents: systems that carry out multi-step tasks with limited supervision.
The idea was more impressive than the reality in 2024. The obstacle was error accumulation. A model that performs each individual step reasonably well can still perform a long task badly. Break a task into twenty steps, give every step a small probability of failure, and the chance that the whole sequence completes correctly is not small.
Reasoning models. At the same time, evidence accumulated that the easiest gains from conventional scaling were becoming harder to obtain. Training frontier models was increasingly expensive, the supply of high-quality human-generated text was finite, and simply adding more parameters and more data no longer appeared guaranteed to produce the dramatic improvements seen in the previous generation. Everything else being equal, an LLM that is a hundred times larger costs roughly a hundred times more to run because each generated token has to pass through roughly a hundred times as many model weights, requiring correspondingly more computation.
And so, the focus on scaling shifted from data and model size to inference. Reasoning models spend more computation after a question has been asked, working through a problem before producing a final answer. The gains were especially visible in mathematics, programming and formal logic, where answers can often be checked objectively and where a system benefits from trying approaches, testing intermediate results and correcting itself. But the same additional computation could also improve performance on less easily verifiable tasks.
Small models and the DeepSeek moment. The other major engineering story ran in the opposite direction. While attention remained fixed on frontier models, smaller models improved rapidly. Distillation allowed a smaller system to learn from the outputs of a larger one. Quantization reduced the numerical precision required to represent a model. Pruning removed unnecessary parameters. Mixture-of-experts architectures allowed models with enormous total parameter counts to activate only part of the network for any particular input.
The result was that capability became cheaper.
The geopolitical importance of this became obvious in January 2025. The Chinese company DeepSeek released models that performed close to the American frontier despite operating under severe constraints on access to the most advanced chips.
2025–2026: deployment and the state
Our current period is harder to synthesize because we are living in it, but this is what seems significant about the last 18 months or so.
Agents started to work really well. Software development was the clearest early example. Coding is unusually well suited to agents because the environment provides feedback. Systems such as Anthropic’s Claude Code and OpenAI’s Codex moved beyond answering programming questions to writing code at scale. They could inspect a codebase, modify several files, write features, run tests, diagnose failures and continue working toward an objective. The role of the human began to shift from writing every line to specifying a task, reviewing what the system had done, and intervening when necessary.
Coding was the easiest case, not the only one. Agents could operate browsers and computers, call external tools, gather information, fill out forms and carry out sequences of actions on a user’s behalf.
AI also became a more physical presence. It had always been physical, of course. The cloud is somebody else’s computer, and somebody else’s computer sits in a building connected to an electrical grid.
But the scale of the AI buildout made that fact difficult to ignore. Every advance in the development of AI since the launch of ChatGPT, even (perhaps, especially) the ones that made AI more efficient, has fueled the demand for more and more processing capacity. This means more chips, more data centers, and more resources. Data centers require land, transmission lines, substations, cooling systems and enormous quantities of electrical power. New power generation requires capital, permits and time. Depending on their location and design, data centers can also create significant demands for water. There was something initially strange about the resulting politics. Search engines, streaming video, social media, cloud computing and online shopping had relied on enormous data centers for years without most people giving it much thought. However, by mid-2026, a lot people began to talk about data centers the way other people talked about masking and vaccines during the COVID Pandemic.
And the state arrived.
Not literally. The European Union had been grappling with AI for years and the American government was already involved through semiconductor export controls, the Biden administration’s 2023 executive order, the AI Safety Institute and other measures.
What changed in 2025 and 2026 was the breadth of the involvement. In the popular press, AI policy discussions expanded from a focus on employment, privacy, discrimination, and copyright to include industrial policy, energy policy, trade policy and national security. The U.S. federal government increasingly treated AI leadership as an economic and national-security objective, while state legislatures pursued their own approaches to everything from automated decision-making to deepfakes, employment, healthcare and the use of AI by government itself.
One of the dominant questions in American AI policy became: Who should make it? The federal government worried that a patchwork of state laws might constrain an industry it regarded as strategically important. States, in turn, were reluctant to wait for federal action.
← Previous: How Large Language Models Work · Contents · Next: Benefits, Harms, and Uncertainty →