The Origin of the AI Era: How One 2017 Paper Launched an Industry

Claude, ChatGPT, Gemini, and Llama come from different companies with different philosophies, yet they share a common ancestor. To understand why the AI market looks the way it does, and why so many companies arrived at the same moment with strikingly similar products, you have to go back to a single research paper published in June 2017.

The Paper: "Attention Is All You Need"

Eight researchers at Google published Attention Is All You Need in June 2017. Its stated goal was modest: a better approach to machine translation. What it introduced was the Transformer, a neural network architecture that now sits at the core of nearly every major large language model (LLM). It is the "T" in GPT.

Before the Transformer, leading language models used recurrent neural networks (RNNs), which read text one word at a time while carrying a running summary forward. Imagine reading a novel through a keyhole that shows a single word. By the last chapter, the details of chapter one have faded. The Transformer replaced that with attention, a mechanism that lets the model look at every word in a passage at once and learn which words relate to which.

Attention itself was not new. Researchers had been using it alongside RNNs since 2014. The breakthrough was the claim in the title: attention alone was enough.

The second advantage turned out to matter even more. Because the Transformer processes a whole sequence in parallel rather than step by step, it fits naturally on graphics processing units (GPUs), chips built to run thousands of calculations at once. An architecture that could absorb more hardware could absorb more data, and that set up everything that followed.

Eight Authors, Many Companies

If you want a literal answer to which event gave birth to many AI companies, look at the author list. By mid-2023, all eight authors had left Google, and most had started companies of their own:

• Aidan Gomez co-founded Cohere, which builds language models for enterprises.
• Noam Shazeer co-founded Character.AI before returning to Google through a 2024 licensing deal.
• Ashish Vaswani and Niki Parmar co-founded Adept, then Essential AI.
• Jakob Uszkoreit co-founded Inceptive, applying similar ideas to RNA design.
• Llion Jones co-founded Sakana AI in Tokyo.
• Illia Polosukhin co-founded NEAR.
• Łukasz Kaiser joined OpenAI.

From a systems architecture perspective, this is the pattern worth noticing. Google invented the core idea, published it openly, and then watched the people who understood it best carry it into new organizations, including direct competitors.

The Spark Needed Fuel

A paper alone does not start an industry. Three other ingredients had to be in place.

Compute

In 2012, a model called AlexNet won the ImageNet image recognition competition by a wide margin, trained on two consumer gaming GPUs from NVIDIA. That result convinced the field that deep learning on GPUs worked. By the time the Transformer arrived, cloud providers were renting GPU capacity at scale, so any well-funded lab could get the hardware.

Data

The internet had produced an enormous body of human writing: web pages, books, forums, documentation, and code. Language models learn by predicting the next word, a task that needs no human labeling, so all of that raw text could serve directly as training material.

Scaling Laws

In January 2020, a team at OpenAI published Scaling Laws for Neural Language Models. They found that model performance improved along a smooth, predictable curve as three things grew together: model size, training data, and compute. The trend held across many orders of magnitude.

That finding turned AI progress from a research gamble into something closer to a capital plan. If you could fund more compute, you could forecast a better model. Several of the paper's authors, including Jared Kaplan and Dario Amodei, went on to co-found Anthropic in 2021.

From Text Predictor to Assistant

A model trained only on internet text is a powerful autocomplete, not a helpful assistant. Ask it a question and it might respond with three more questions, because that is what much of the text it learned from did.

The fix was reinforcement learning from human feedback (RLHF), described in a 2017 paper by researchers at OpenAI and DeepMind. People rank a model's responses, and the model is trained toward the responses people prefer. In 2022, OpenAI applied the technique to language models with InstructGPT. Human evaluators preferred the outputs of a 1.3 billion parameter InstructGPT model over those of the 175 billion parameter GPT-3. Alignment with human intent beat raw size.

This is where the human-centered part of the story lives. The capability came from scale, but the usefulness came from people showing the system what good looks like.

Why Everyone Arrived With the Same Technology

The striking thing about the market since 2023 is how similar the products look. That convergence was not a coincidence. Three forces made it close to inevitable.

The research was public. The Transformer, the scaling laws, and RLHF were all published openly, and any lab could read them the week they appeared. In 2018, Google and OpenAI each released influential Transformer models within months of each other (BERT and the first GPT).

The talent pool was small and mobile. A relatively small group of researchers at Google Brain, DeepMind, OpenAI, and Meta drove most of the early progress, and they moved between those organizations, carrying ideas with them.

Once the recipe was known, money became the moat. After the formula of Transformer plus scale plus human feedback was understood, the main barrier was no longer secret knowledge. It was the cost of compute and the ability to hire people who knew how to use it.

When everyone starts from the same architecture, trains on similar data, and follows the same scaling curves, they arrive at similar products. The differences show up in training choices, safety approaches, and product design, but the foundation is shared.

November 2022: The Moment It Became Visible

The technology had been compounding quietly for five years when OpenAI released ChatGPT on November 30, 2022. A UBS analysis estimated that it reached 100 million monthly active users in January 2023, about two months after launch, making it the fastest-growing consumer application on record at the time. TikTok had taken roughly nine months to reach the same mark.

That moment changed the funding environment almost overnight. Microsoft announced a multiyear, multibillion-dollar investment in OpenAI in January 2023, and Meta began releasing its Llama models to researchers the following month. More capital bought more compute, more compute produced better models, and better models attracted more capital. That feedback loop is a large part of why the last few years have felt so fast.

What This History Means for Your Content Stack

If the leading models share a common origin, so do many of the AI features being sold to you. Most AI capabilities inside CMS platforms are built on foundation models from a small number of providers, accessed through an API. The model is rarely what separates one vendor's offering from another.

That shifts where the advantage lives. What separates useful AI from generic AI is the context around the model: well-structured content, clean data, clear governance, and workflows that put human judgment in the right places. We explored how to tell real capability from relabeled features in Agentic CMS: What It Actually Means, and why shared models produce shared output in Why AI Content Still Sounds Like AI Content.

It is the same lesson RLHF taught the industry. A model becomes useful when people shape what it should do.

Practical Takeaways

• Ask what model a feature uses and what context it receives. The model is probably shared with your competitors. The context is where the value is.
• Invest in structured content. A model can only reason well over content it can read cleanly.
• Expect convergence to continue. Capabilities that feel unique today tend to appear across vendors within months, because the underlying research is shared.
• Keep humans in the loop. The breakthrough that made LLMs usable was human feedback, and that principle still applies inside your own workflows.