AI progress is advancing very quickly, at an exponential rate.

Judge as of 2024-03-29. True: credible measures of AI progress up to early 2024 (e.g. training compute, benchmark performance, price-performance of compute, model capability trends from sources such as Epoch AI, Stanford AI Index) show sustained rapid growth that is roughly exponential, meaning a roughly constant doubling time or multiplicative rate over years. Mostly true: progress was very rapid and key inputs (e.g. compute) grew exponentially, but capability measures grew in a way that is only loosely exponential or uneven. Mixed: some major measures show exponential growth while others of similar weight show linear, saturating or slowing trends, so 'exponential' holds for about half of the relevant evidence. Mostly false: progress was fast, but the evidence shows it was not exponential in any meaningful measure, or only in a narrow input stated as if it applied to AI progress overall. False: evidence shows AI progress was not rapid, or was stagnating or slowing, as of early 2024.

Set before the research began.

1. E1 Compute trends across three eras of machine learning Epoch AI · 16 February 2022

   “Furthermore, we find that the training compute has grown by a factor of 10 billion since 2010, with a doubling rate of around 5-6 months.”

2. E2 Compute trends across three eras of machine learning Epoch AI · 16 February 2022

   “Interestingly, the growth of compute in these Large-Scale models seems slower, with a doubling time of about 10 months.”

3. E3 Technical Performance Stanford Institute for Human-Centered Artificial Intelligence (HAI), 2024 AI Index Report · 1 January 2024

   “AI models have reached performance saturation on established benchmarks such as ImageNet, SQuAD, and SuperGLUE, prompting researchers to develop more challenging ones.”

4. E4 The 2024 AI Index Report Stanford Institute for Human-Centered Artificial Intelligence (HAI)

   “According to AI Index estimates, the training costs of state-of-the-art AI models have reached unprecedented levels. For example, OpenAI’s GPT-4 used an estimated $78 million worth of compute to train, while Google’s Gemini Ultra cost $191 million for compute.”

Epoch AI found that training compute for milestone ML systems grew about 10 billion-fold from 2010, doubling roughly every 5-6 months [E1]. Even the slower large-scale trend still doubled about every 10 months [E2]. So the main input grew at a roughly constant exponential rate. The 2024 AI Index says models have saturated established benchmarks like ImageNet, SQuAD and SuperGLUE [E3], and that frontier training costs reached $78M for GPT-4 and $191M for Gemini Ultra [E4]. That points to very fast progress, but it does not show capability itself growing at a constant exponential rate. The criteria call this case mostly true: inputs grew exponentially, while the capability evidence is only loosely exponential.