The History of the AI Hype

The topic of AI is currently on everyone’s mind, mainly driven by large language models (LLMs) such as ChatGPT or image generators such as Dall-E and Midjourney. There is a lot of talk about where the development of AI will take us. Which professions will soon no longer require human labour? What are the dangers of general artificial intelligence (AGI)?

The First AI Hopes in Retrospect

People tend to forget that this is not the first time we have asked ourselves these questions. The term ‘artificial intelligence’ was first used back in the 1950s. In 1959, Herbert A. Simon, J. C. Shaw and Allen Newell published the ‘General Problem Solver’ (GPS), a computer programme designed to solve problems of all kinds. The attempt was ultimately considered a failure as the programme could only reliably solve simple, well-defined problems. Nevertheless, AI euphoria broke out in the 1960s and in 1963 Leonard Uhr and Charles Vossler described one of the first programmes for machine learning. In 1965, Joseph Weizenbaum built ELIZA, an interactive program designed to conduct any dialogue in English.

Somehow it all sounds familiar, doesn’t it? Since the release of ChatGPT, similar tones have been struck again, entire industries have been declared outdated or scenarios of the extinction of humanity by AI have been created.

Early AI Visions and Their End

So the question is: why did the predictions of the 1960s not materialise?

Paul’s answer is quite simple: expectations were exaggerated. It was assumed that the great leaps that made rapid progress possible could simply be repeated, but they were disappointed. Development slowed down rapidly, sceptics became louder and so the AI winter followed in the 1970s. Advances in expert systems led to a recovery in the field in the early 1980s, but here too the high expectations could not be met. The second AI winter followed until the 2010s.

New, fundamental innovations are therefore needed to fulfil the current expectations for the development of AI. However, no one will be able to reliably predict if and when these will materialise.

So will there be a third AI winter? Perhaps. But not to the same extent as we saw in the 70s and 90s. LLMs and other generative models already have a lot of untapped potential. They will not replace humans – at least for the time being – but they can support us.

Open AI for Everyone

While frontier labs such as OpenAI or Anthropic are using more and more resources for ever smaller developments, others are focusing on packaging existing capabilities into ever smaller and more efficient models that can then be operated outside of huge data centers – in some cases even on smartphones.

The figure above shows a logarithmic diagram depicting the declining costs of LLMs with constant capabilities over the period from 2022 to 2024 (Source: Andreessen Horowitz).