Mistral AI Releases Groundbreaking Life Cycle Assessment of Large Language Model to Set New Industry Standards

Mistral AI has released what it describes as the first comprehensive life cycle assessment of a large language model, aimed at setting a new benchmark for transparency in the industry.

The report primarily focuses on Mistral’s flagship model, Mistral Large 2, detailing the environmental costs associated with training and 18 months of operation. By January, the model had generated 20.4 kilotons of CO₂ emissions, consumed 281,000 cubic meters of water, and used resources equivalent to 660 kilograms of antimony. (Antimony equivalence is a metric used to gauge the consumption of rare metals and minerals needed for producing equipment.)

The team calculated the environmental impact of a single response from Mistral’s Le Chat (400 tokens). It was discovered that each request results in emissions of 1.14 grams of CO₂, 45 milliliters of water, and 0.16 milligrams of antimony.

Recently, OpenAI’s CEO Sam Altman claimed that on average, ChatGPT uses just 0.32 milliliters of water, less than 1% of what Mistral consumes. However, as OpenAI has not provided comprehensive data on its environmental impact and Altman’s mention of this statistic was cursory, it remains unclear how comparable these figures truly are.

Mistral’s report emphasizes that the vast computational power required for running generative AI, often executed on GPU clusters in regions with high carbon emissions and sometimes facing water scarcity, has a significant negative environmental impact. This is not a new revelation, but the scale of the issue has escalated alongside the ongoing boom in AI technology.

The research also found a clear relationship between model size and its environmental footprint. Larger models have a proportionally greater impact: a model ten times larger has roughly ten times the environmental effect, assuming the number of generated tokens is constant. This highlights the necessity of selecting the appropriate model for each specific application.

Based on these findings, Mistral suggests that the industry adopt three key metrics: the total impact of model training, the impact per output (per query), and the ratio of output impact to overall life cycle impact.

Mistral argues that the first two metrics should be mandatory to provide the public with a clearer understanding of AI’s environmental impact. The third metric could serve as an internal indicator, revealing insights into the life cycle without needing to be public.

The company envisions two main approaches to mitigate AI’s environmental impact. Firstly, AI companies should disclose the environmental impact data of their models in accordance with internationally recognized standards, making it easier for users to compare and choose more eco-friendly options. Secondly, users can help by utilizing generative AI more effectively: selecting models suited to their needs, consolidating requests, and avoiding unnecessary computations.

Mistral acknowledges that this initial analysis is only a rough estimate. Accurate calculations are challenging due to a lack of established standards for assessing the life cycle of large language models and publicly available assessment metrics. Moreover, reliable data on the life cycle of GPUs is still lacking.

The company plans to regularly update its environmental impact reports and aims to contribute to the establishment of international industry standards. Results will be made available in the Base Empreinte database, a French reference platform with data on the environmental impact of products and services.

In the EU, the new Artificial Intelligence Act already mandates AI model providers to meticulously document their energy consumption. Developers are required to furnish technical documentation detailing the energy usage of their models. Energy consumption is one of the factors regulators consider when determining if a model poses a «systemic risk.»

*Would you like to stay updated on significant news from the AI world? Subscribe to our Telegram channel* [*BotHub AI News*](https://t.me/bothub).

For the original news source, click [here](https://the-decoder.com/mistral-ai-publishes-the-first-comprehensive-life-cycle-assessment-of-a-large-language-model/).