ChatGPT may not be as power-hungry as once assumed

ChatziptOpenai’s chatbot platform may not be the once-energy-hungry. However, a new survey has shown, depending on how the hunger is mainly used in ChatzP and the question that the AI ​​models are answering.

Ay Recent analysis Epoche AI, a non -profit AI Research Institute, tried to calculate how much energy a simple chatter quarry costs. Ay Generally quoted status Is it necessary to answer a single question of ChatzPT about 3 watts of power or 10 times more than Google search.

Ipach believes that it is an excessive.

Using the latest default model of Openai for ChatzPT -for GPT -4OAs a reference, the apocha has been found that the average chatzPT Query consumes about 0.3 watts-hours-less than the tool of many households.

“Using ordinary equipment or heating or cooling your home or driving or driving is not a big deal,” analyzing the analysis of the apochet’s data analyst Joshua tells TechCrunch.

The use of AI’s energy – and its environmental impact, broadly speaking – AI companies are a matter of controversy to expand their infrastructure footprints quickly. Just last week, a team of more than 100 companies Publish an open letter AI calls and regulators to ensure that new AI data centers do not reduce natural resources and do not force utilities to rely on non-recurrent sources of energy.

Its analysis was encouraged by TechCrunch that his analysis was identified as the old of previous research. You have mentioned for example that the author of the 3-hour-hour-to-end report used to use old, less skilled chips to operate the OpenAI models of OpenAI.

“I have seen lots of public speeches that AI is about to consume a lot of energy in the coming years, but the power that was going on today did not describe the power that was going on today,” you said. “Also, some of my colleagues have noticed that the most wide estimates of 3 watts per Qur’an were based on fairly old research and some napkin seemed very much based on mathematics.”

Grant, an approximate, as well as an approximate, 0.3 watt-hour image of the apoc. The Open did not reveal the necessary details to make a specific calculation.

The analysis also does not even consider the extra energy consumption spent by ChatzPT features such as image generation or input processing. You have acknowledged that the “long input” chatzipt queries – for example long files with long files – for example – is probably used above electricity than a simple question.

You said that he was hoping to increase the cost of Baseline Chatzipt.

“[The] AI will be better, this II will probably require a lot more energy to train and this future AI can be used more intensely – people manage a lot more work and more complex tasks than how people use ChatzPT today, “you said.

Was there Egotistical In recent months, AI skills are expected to drive the extension of the AI, the AI ​​deployed, and the extension of the power-hungry infrastructure. In the next two years, AI Data Centers may need all the California 2022 power capacity (68 gigwatt), According to a rand reportThe By 2030, the report is predicted to claim electricity output equivalent to eight nuclear reactors (8 gigawatts) to train a border model.

The ChatzP is the only one – and the stretch – reaching the number of people, its server is similarly claimed. OpenAI, plans with several investment partners Spend billions of dollars on new AI Data Center projects Over the next few years.

Openai’s attention-baki is also transferred to the rest of the AI ​​industry-so-called rational models, which are usually more capable of the tasks they can perform, but more computing is needed. Against the GPT -1 and the models, which almost immediately answer the question, “Think” a few minutes before answering rational models, a process that makes more computing successful -and thus strength.

“Reasonable models will take up the tasks that old models cannot and produce more [data] To do this, and both need more data centers, “you said.

The OpenAI has begun to publish more power-rational models like OpenAI O3-miniThe However, it seems impossible, at least at this point, skill achievements will offset the “thoughts” process of rational models and the use of AI around the world will offset the demand for increasing power from increasing energy.

You have suggested that people are concerned about their AI energy footprints that applications use ChatzPT or select models that reduce the required computing – the amount that is realistic.

“You can try to use small AI models [OpenAI’s] GPT -4 -Minai, “You said,” and use them in a small amount so that one ton of data needs to be processed or produced “”