AI agents are inorganic systems that are specifically designed to mimic human intelligence. These systems are increasingly becoming smarter and more important to the extent that two of the Nobel Prizes awarded in the year 2024 were related to AI. The Nobel Prize in Chemistry was related to a crucial problem in biomedical research that aims to come up with better drugs for tackling diseases.
We are interacting with such platforms on a day-to-day basis. It could be a web search for information; or a movie, or product suggested on our screens; or a command to our voice assistant. With the rise of generative AI apps like ChatGPT users are creating content for all sorts of things. Most of us who have either used or heard about the potential of AI are in awe of it. However, not all are aware of the requirements related to deploying such an agent.
To understand end, let us take the example of BigScience Large Open-science Open-access Multilingual language model (BLOOM). It is a large language model (LLM) that is designed for natural language processing. Such models form the basis of generative AI chatbots like ChatGPT. Now, these models, before their deployment, go through extensive data exposure and training. For example, BLOOM is a 176 billion parameter model that was trained on 1.6 terabytes of data in 46 natural languages and 13 programming languages. The total training time was estimated to be almost 118 days. The total amount of energy consumed for training purposes alone was 433,196 kWh.
Let us put these numbers into perspective. At the rate of 100 kilobytes per page, 1.6 terabytes of data would amount to almost 18 million pages; which is almost equivalent to 56,667 books@300 pages per book. Concerning energy consumption, if we take a household in India that consumes 200 kWh of electricity per month, then 433,196 kWh would be equivalent to 180 years of electricity consumption. To this, if one adds the energy cost associated with: (i) data such as its sourcing, collection and processing; (ii) architecture engineering and model evaluation; and (iii) computing systems manufacturing including chips, material manufacturing and raw material extraction; then the total energy cost associated with a single LLM, even before deployment, is significant. The energy consumption post-deployment may be estimated by the fact that a simple keyword search on Google consumes around 0.3 kWh, whereas an LLM-based interaction can consume 10 times or even more.
It is no surprise that Sam Altman, the chief executive at OpenAI (ChatGPT fame), has described the cost of running the services as "eye-watering." Sundar Pichai, CEO of Google, commented that "one of the constraints for AI could be the infrastructure, including energy."
Other than energy requirement, the successful deployment of an AI agent is mainly limited by two important factors: computing power and data. One way to compare the computing power of different countries is to look at the number of supercomputers each country has. If we go by the world’s top five economies, then the US has the largest 34.6% of supercomputers, followed by China (12.6%), Germany (8.0%), Japan (6.8%), and India (1.2%). Regarding the availability of data in India, MeitY's 2019 report states, “A large amount of data exists today in a plethora of sectors. However, it mostly resides in stand-alone mode without it being used effectively. Even when it is used, it gets used only in silos. Many times, data integration across sectors produces spectacular benefits, which are missed.” In these times, data is seen as a new form of wealth, like oil and gas. Most of it, belonging to Indian internet users, is stored on foreign servers. Despite being in the top five economies, we are far behind other countries in terms of computing power and data. Bridging this gap has huge financial implications. However, these implications are not just financial in nature.
As we have seen earlier, the energy costs associated with an AI agent are significant. As per IEA, in the year 2022, energy-related global CO2 emissions rose by 1.1% to a level which is more than 37 billion metric tons. This increase in emissions comes at a time when the world’s population is feeling the effect of climate change in one form or another. In the context of Assam, the effect of floods is steadily increasing the number of people affected, and tea growers are increasingly facing uncertainties related to rainfall. The growing thrust toward expanding the reach and bounds of AI-related technology may be substantial in the near future. More so, in countries like ours that are lagging behind for the time being.
Apart from that with the increase in productivity of labour owing to AI, the world is already seeing the replacement of working masses with that of AI. Alphabet CEO, Sundar Pichai has publiclyaccepted that 25% of the code in Google is being written by AI. It is no surprise that as a result, many tech giants including Google, and Amazon have paused hiring in India. The British Telecommunication giant BT announced in 2023 that it is “going to replace 10,000 workers with AI as part of a wider cull of up to 55,000 staff in a bid to slash costs”. All this would certainly affect each one of us in this connected world.
Let us see what Geoffrey Hinton, one of the godfathers of AI, has to say about the future. For his pathbreaking work concerning AI systems, he was awarded the Nobel Prize in Physics in 2024. He put forth his concerns as “There is enormous uncertainty about what is going to happen next!... It may be that we look back and see this as a turning point when humanity had to make the decision about whether to develop these things further and what to do to protect themselves if they did.”
One thing that should be clear to any concerned citizen of the world is that, at the core of it, the present AI race is fuelled by the lack of trust among, and the insecurities of, the powerful countries of the world, along with large corporations’ ever-expanding desire for market share and profits. And, with these emotions and interest in the driving seat, can we dream of a better world?
Last Monday Jan 6, 2025, an edited version of the same got published in the Assam Tribune.
