Employee salaries, which can account for half the cost, are not included
The cost of training AI models has skyrocketed in the past year, according to research from Epoch AI. This sharp rise highlights the increasing complexity and capabilities of AI models in a very short time.
Last year saw the release of ChatGPT-4 by OpenAI in March, which sparked a global AI frenzy. Google followed suit with its advanced AI model, Gemini, in December.
Training both systems was significantly more expensive than previous AI models, with development costs reportedly reaching hundreds of millions of dollars, according to Epoch AI.
The research firm estimated that training Gemini, a large language model capable of processing text, voice commands, and images, cost between $30 million and $191 million, excluding employee salaries. Epoch AI noted that salaries could account for 29% to 49% of the final cost.
The creation of ChatGPT-4, the latest version, technically ranged in cost from $41 million to $78 million. Sam Altman, CEO of OpenAI, previously stated that the model cost over $100 million to develop, supporting these estimates.
In contrast, previous AI models were much cheaper to train. For example, ChatGPT-3 cost only $2 million to $4 million in 2020, while Gemini's predecessor, known as PaLM, took between $3 million and $12 million to train in 2022, based on computing costs alone. Even at these price points, staying competitive with cutting-edge AI development has been challenging for academic institutions or public organizations traditionally active in AI research.
Although updated in late 2023 to support voice and image input, ChatGPT-4 was initially text-based. On the other hand, Gemini was designed as a multimodal language model from the start, which explains the lower initial training cost for ChatGPT. Gemini’s broader focus on applications—such as prompting users to take photos with their smartphones, identifying features, and analyzing them—justified the higher development costs.
Post a Comment
0Comments