Who invented the transformer, and why Google did not win with it.

Translation was the bottleneck.

Before 2017, the best translation models read a sentence one word at a time, in order. That made them slow to train and weak at connecting words far apart in a long sentence.

What the paper actually introduced.

Attention itself was older: researchers had used it since 2014 to let a translation model look back at the right words of the source sentence. The 2017 paper removed the word-by-word reading entirely and built a model out of attention alone.

Two parts were new and are still inside every large language model. Scaled dot-product attention lets every word weigh every other word at once. Multi-head attention runs many of those comparisons in parallel, each free to track a different kind of relation. Because the model no longer reads in order, the position of each word is added as a pattern of sine and cosine waves.

OpenAI, four months ahead.

In June 2018 OpenAI published GPT-1, a transformer trained to predict the next word on large amounts of text. Google's BERT followed in October 2018. OpenAI kept scaling the same idea through GPT-2 and GPT-3, and released ChatGPT in November 2022. Google announced Bard in February 2023, as the answer to a market OpenAI had already taken.

All eight authors of the paper left Google. One of them, Noam Shazeer, came back in 2024 to co-lead Gemini, and left again for OpenAI in June 2026.

One architecture took the money, the chips and the talent.

In 2019 Rich Sutton, who later won the Turing Award, published The Bitter Lesson: across 70 years of AI, general methods that make the most of computing power have beaten methods built around human cleverness. The transformer became the strongest case for it. It scales almost perfectly on GPUs, so it spread from text to images, audio, biology and robotics, and the industry put its money, chips and researchers into making the same design bigger.

That concentration is now criticized from inside. At TED AI in October 2025, Llion Jones, one of the eight authors of the 2017 paper, warned that the race built on that paper is stifling the next breakthrough, and has moved his own research away from transformers. Sutton himself argued in 2025 that large language models are a dead end, because they learn from human text instead of from experience, and so do not follow his lesson as well as people assume.

Faster transformers are still transformers.

The best-known alternatives, such as Mamba and RWKV, make a model read long text faster and with less memory. They still do the same job: predict the next word.

The limit is in that job itself. Predicting the next word from probabilities is extraordinary for language and was never built to check logic. Why LLMs cannot reason.

A different starting point.

PRISM does not start from predicting text. It starts from a principle of physics: systems settle into the state that takes the least energy. An answer that fits what the model knows has low energy, and a contradiction makes the energy rise, so PRISM corrects itself before it speaks. Our story.