NextFin News - On November 6, 2025, at the FT Future of AI Summit in London, Geoffrey Hinton joined a panel of Queen Elizabeth Prize for Engineering laureates to reflect on the milestones that shaped modern AI. The session, hosted by Financial Times AI editor Madhumita Murgia, brought together leading figures in machine learning to discuss the origins and implications of recent advances. During the conversation, Hinton described an early experiment from the mid-1980s that he now regards as a direct precursor to today’s large language models.
The panel was part of the FT Future of AI Summit and featured the laureates reflecting on decades of research and development in AI. The conversation has been published and archived by FT Live and is available in public video and transcript form.
Hinton’s 1984 experiment: backpropagation and next-word prediction
Hinton recounted a formative experiment from "about 1984," in which he applied backpropagation to the simple task of predicting the next word in a sequence. He summarized the setup and finding succinctly: In about 1984, I tried using back propagation to learn the next word in a sequence of words.
He presented this as an early instance of training a model on sequential text to forecast subsequent tokens.
How the tiny model learned word meaning
Describing the behavior that surprised him, Hinton reported that the network learned internal representations that captured semantics. As he said, it would learn interesting features for the meaning of the words. So just giving it a string of symbols, it just by trying to predict the next word in a string of symbols, it could learn how to convert words into sets of features that captured the meaning of the word
. He emphasized that the learned features interacted in ways that allowed the model to predict features of the following word.
Framing the experiment as a precursor to modern LLMs
Hinton was explicit about continuity between that small experiment and today’s models. He called the system a tiny language model from 1980 late 1984 that I think of as a precursor for these big language models
, stressing that the basic principles were the same. It was just tiny.
He underlined that the difference was scale rather than principle.
Practical limits kept the idea small for decades
Hinton described why the 1984 insight did not immediately lead to large-scale deployments: We had 100 training examples. It took 40 years to get us here though and it took 40 years to get here. And the reason it took 40 years was we didn't have the compute and we didn't have the data and we didn't know that at the time.
He framed those three shortages — compute, data, and awareness of their importance — as the main bottlenecks that delayed the realization of the idea.
Surprise at early limitations
Reflecting on his own reaction at the time, Hinton admitted puzzlement that the early successes did not immediately translate into broader breakthroughs: We couldn't understand why we weren't just solving everything with that
. The remark underscores the gap between theoretical or laboratory results and the engineering-scale resources later required to turn them into widely capable systems.
Panel context and closing notes
Hinton’s comments came during a broader discussion with other laureates about the historical sequence of breakthroughs that enabled contemporary AI — from algorithmic advances to the emergence of specialized hardware and massive datasets. The panel’s purpose was to surface the personal and technical turning points that, when combined over decades, produced the current wave of generative models.
References:
Transcript: The Minds of Modern AI — Singju Post (full transcript of the FT panel)
FT Future of AI Summit — FT Live (event page)
Explore more exclusive insights at nextfin.ai.
