As far back as in 1958, computing visionary and architect John von Neumann had already foreseen that at some point the power of computations would enter a regime where the computers themselves are able to boost their own efficiency and applicability. This view has been echoed by a number of experts in the last several decades, and this point seems to be fast approaching.

The tipping point would be what is called Artificial General Intelligence (AGI) – the point at which AI is capable of learning by itself any arbitrary skill that a human or animal could learn. Four polls have been conducted in 2012 and 2013 among AI experts, with 50% were confident AGI would arrive between 2040 to 2050. Perhaps the newest results from OpenAI would lead some of these experts to estimate an even sooner date for this. For instance, early experimenting with the newest version of the algorithm behind ChatGPT, called GPT4, indicate human-level performance in a large number of tasks.

At this stage, a positive feedback loop is formed: the more compute power is created, the more computers can improve themselves. Digital solutions will become increasingly demanding and economic pressures for datacentres and supercomputers will become insurmountable.

GPT3 is estimated to have consumed 500 tonnes of CO2 equivalent to be trained. GPT4 could have a comparable carbon footprint – it contains more parameters but was trained with less inefficiencies. For now, a single training of a new AI model is costly but not -yet- catastrophic.

The red flags start to be raised when we think about the boom in new AI tools, the expanding size of the data sets in which they are trained and the indiscriminate use of compute power. Governments around the world have started regulating footprints for data centers in order to control their tabs in carbon emissions. Otherwise, hitting net-zero emission pledges will be impossible.

This will be the ceiling at which point the diminishing returns of AI will limit its applicability and potential to fulfill its potential.

Economically and environmentally sustainable AI will create an unavoidable push for alternative technologies. There is only very limited room for improvements to the efficiency of the software itself, and ultimately the computers will need to become more efficient machines. 

Some immediate solutions are being devised. AMD’s CEO Lisa Su recently announced a new accelerated processing unit (MI300) that can reduce the time and energy consumption for training large language models. Less orthodox tricks may involve alternative cooling methods or developing neuromorphic purpose-specific chips.

All these solutions, however, suffer the same issue – they offer a fixed increase in energy and cost efficiency, pushing the horizon for a hardware bottleneck a couple of years further. Ultimately, the exponential growth of demand for digital resources will catch up with these solutions and AI will again conflict with environmental goals.

The ultimate endgame for the power gluttony of AI is a hardware with compute capabilities that also grow exponentially. Only quantum computing offers this potential. With the availability of quantum computers, AI will finally be able to indefinitely expand its positive impact in the world. 

The author, Dr. Andre Saraiva, is Head of Solid State Theory at Diraq. Andre is a world leader in quantum chemistry and solid state physics.