Shulou(Shulou.com)11/24 Report--

AI produces three orders of magnitude less carbon dioxide than human hands-on painting.

The conclusion comes from a paper recently shared by LeCun:

If you look at the number of comments forwarded, you can see that netizens can't sit still again.

Some netizens were a little surprised to see this result:

Who would have guessed that generative AI is more creative and environmentally friendly?

There are also a handful of netizens who have not read the paper and asked:

Do you take into account the carbon dioxide emissions during the training model?

In addition, many netizens questioned the calculation method used in the paper.

So how on earth is it calculated in this paper? Does the training model include it? Let's take a look.

The AI vs human study was carried out by researchers from the University of California at Irvine and the Sloan School of Management at the Massachusetts Institute of Technology.

In this comparison of man-machine carbon dioxide emissions, the AI team members sent by the researchers are ChatGPT (text), BLOOM (text), DALL-E2 (image) and Midjourney (image).

Let's first take a look at the comparison of the text.

Writing comparison

The first problem to be solved is to define the carbon dioxide emission source of AI.

The researchers believe that the two main components are the training emissions of the model and the emissions per query. Training emissions are calculated as an one-time cost and will be apportioned to each query.

The training emission data of the model referenced here are: training GPT-3 emits about 552 tons of carbon dioxide equivalent; training BLOOM emits 30 tons of carbon dioxide equivalent.

(carbon dioxide equivalent is a measure used to compare different greenhouse gas emissions)

In addition, the training model can not be trained once, and then it should be continuously trained and optimized. Here, researchers default to conduct a complete training of the model once a month.

The calculation of the emission of each query is based on the data that ChatGPT emits about 3.82t per day and replies to 10000000 queries, and it is estimated that each query will produce 0.382 grams of carbon dioxide equivalent. In the same way, BLOOM emits 1.5g per query.

After comprehensive training and inquiry, it is concluded that ChatGPT emits about 2.2g carbon dioxide equivalent per query, and BLOOM emits 1.6g carbon dioxide equivalent per query.

In terms of carbon dioxide emissions from human writing, an article in The Writer magazine pointed out that Mark Twain can produce about 300 words per hour, which can be regarded as the average writing speed of other writers.

Based on the above speed, it is estimated that it takes about 0.8 hours for a person to write 250 words (one page here).

Then the researchers' algorithm goes like this, aunt:

The average annual carbon emissions of Americans is about 15 tons of carbon dioxide equivalent, while Americans emit about 15 tons of carbon dioxide equivalent per hour / 8760 hours = 1.7kg carbon dioxide equivalent. As a result, the carbon emissions produced by Americans writing 250 words are about 0.8hr x 1.7kg / h = 1.4kg CO2 equivalent, or about 1400 grams.

By the same token, the average Indian emits about 1.9 tons of carbon dioxide equivalent a year. Indians emit about 0.22 kilograms of carbon dioxide equivalent per hour. As a result, the carbon emissions produced by Indians writing 250 words is about 0.18 kilograms of carbon dioxide equivalent, about 180 grams.

In addition, based on the average computer power and carbon emissions from power generation, the researchers calculated that laptops that support human writing produce about 27 grams of carbon dioxide in 0.8 hours and desktops produce 72 grams.

Let's take a look at the comparison results.

The carbon dioxide equivalent of each page of text generated by BLOOM is 1once 1500 of that of American writers and 190 of that of Indian writers.

The emission of ChatGPT is 1max 1100 written by American writers and 130 by Indian writers.

Considering the use of computers, AI writing also emits much less than humans plus computers.

Let's take a look at the image comparison.

Drawing comparison

DALL-E2 refers to ChatGPT's method and emits about 2.2g per query.

Midjourney CEO David Holz has said that each image requires tens of thousands of terabytes of computation.

So based on the calculations, the researchers calculated that electricity consumption was then converted into carbon emissions, and finally estimated that Midjourney emitted about 1.9g of carbon dioxide equivalent for each map generated.

In calculating the carbon dioxide emissions of human mapping, the researchers first calculated based on industry data that the average project cost of a human illustrator is about $200 and the average hourly wage is about $62.50 per hour. It is therefore estimated that it takes an average of 3.2 hours to complete a commercial drawing project.

The carbon emissions of illustrators in the United States and India are also assessed here.

The average American emits about 15 tons of carbon a year. It takes 3. 2 hours to create an image, and the corresponding carbon emissions are about 5500 grams. The average annual carbon emissions of Indians is about 1.9 tons, which corresponds to about 690 grams of carbon emissions in a map.

In addition, computer carbon emissions should also be taken into account. Laptops produce about 100 grams of carbon dioxide in 0.8 hours and desktops about 280 grams.

To sum up, the comparison results.

The carbon dioxide equivalent of DALL-E2 is about 2500 of that of American illustrators and about 310 of that of Indian illustrators.

The carbon dioxide equivalent of Midjourney is about 2900 of that of an American illustrator and about 370 of that of an Indian illustrator.

Also considering the use of computers, AI graphics are much less emitted than humans and computers.

Other factors should also be considered. Although AI emits much less carbon dioxide than human hands-on writing and drawing tasks, according to the researchers' calculations, they also point out their limitations:

The researchers believe that not all areas are suitable for AI intervention, and that some simple tasks are more efficient. And future technological changes may change the degree of impact of AI and human beings on the environment.

Secondly, AI also has potential social impact, which may lead to job loss and controversy over the legitimacy of training data.

In addition, with the improvement of AI technology, it is possible to lead to an increase in demand for goods and services produced by AI, resulting in a further increase in resource use and pollution through the rebound effect (rebound effects).

In short, the researchers believe that:

AI can play an important role in all areas of society and will not fall into the problem of carbon emissions at present. Although AI carbon emissions can not be ignored, AI carbon emissions are far lower than human carbon emissions in some writing and drawing tasks, and the advantages of AI over human carbon emissions should not be ignored.

The papers that caused heated discussion among netizens and were sent out by LeCun naturally attracted the attention of many people.

Netizens who have read the paper questioned the calculation method:

There is a big problem with the methodology of this article, and it is impossible to simply compare the emissions of a person with those of an AI model.

Other netizens have brought out the Givens paradox, thinking that when the use of AI increases, carbon dioxide emissions will bounce back:

I use Midjourney to generate far more images than I ever wanted to buy from human artists; when the price elasticity is less than 1, the reduction in unit cost will lead to an increase in total expenditure, although I admit that I may not generate 10,000 images.

In addition to complaining about flawed methods, some netizens think that such emissions are negligible compared with those from other activities:

Who is the biggest emitter of carbon dioxide? Shouldn't technicians pay more attention to those activities that really have a significant impact?

Although the researchers have more or less mentioned these issues in the final discussion of the paper, they have become a hot topic among netizens.

Of course, there are also netizens who affirm this comparison:

Despite the shortcomings of this approach, I think it opens the door to discussion and potential new approaches, promoting better comparisons between AI and "human" (traditional counterparts), which is a topic that must be opened.

There are also netizens who have reached doge with this wave of Q:

Don't tell Jensen.

Reference link:

[1] https://twitter.com/ylecun/status/1704098609535320365

[2] https://arxiv.org/abs/2303.06219

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.