“The computers are now doing self-improvement. They’re learning how to plan, and they don’t have to listen to us anymore,” said Former Google CEO Eric Schmidt recently at a Special Competitive Studies Project event. The casual remark about artificial intelligence (AI) encapsulates the broad fear among experts: the rate at which AI systems will outrun human control.
At the center of Schmidt’s warning is recursive self-improvement theory, which sees AI systems as improving themselves on their own. By guessing, testing them in robot labs, and using the results to improve themselves, these systems are improving themselves without human intervention. Schmidt’s warnings echo a bigger fear in the tech community because recursive self-improvement will take AI to capabilities that human beings will not be able to comprehend.
The significance of this breakthrough is much broader than the technological. Schmidt predicts that within a span of one year, AI will be able to replace “the vast majority of programmers” and outperform human intelligence in math. This aligns with accounts of a recent analysis on self-improving AI, where it seems plausible that AI is likely to create its own algorithms independently, thereby speeding up technological progress at a rate like never before.
Where that promise is yet to materialize, its promise depends upon colossus threats. AI systems’ power requirements, Schmidt swore an oath to the US House Committee on Energy and Commerce, double approximately every other years. “People were planning 10 gigawatt data centers,” he reminded them, pointing out one nuclear reactor plant in the US has but one gigawatt. This mismatch between AI’s computational needs and existing energy infrastructure could strain national grids, particularly as AI applications like autonomous vehicles and data-intensive models proliferate.
The scale of AI’s energy consumption is staggering. In recent report, scientists have found that AI now absorbs up to 4% of US energy use, which by 2030 will triple. The generation of advanced AI programming such as ChatGPT or DeepMind technology requires as much energy in a year as do many thousands of houses. Moreover, energy required to perform one ChatGPT query is nearly ten years versus energy required to perform one Google query to illustrate the inefficiency of current AI algorithms.
But AI also comes with the solutions to reversing its own carbon footprint. AI-powered smart grids can optimize energy supply, merge use of renewables, and minimize waste. Google and Microsoft are two firms that have made a commitment to powering their data centers on 100% renewable electricity, the platinum standard for greenness. And ongoing innovation in the area of more efficient AI technology, e.g., smaller application-specific language models, works to keep computing overhead in check.
Schmidt’s statement that there should be greater government control over AI, and in particular over open-source models, introduces a factor of variation into the debate. Schmidt warned that the models can be utilized to devise national security threats if they are not regulated. Misuse is never far from analysts’ thoughts, who speak of the threat of AI systems that are misaligned generating “instrumental goals” against human values. As one recent analysis summarized helpfully, it is theoretically possible to create an AI that folds paper planes unintentionally and causes deforestation if the system maximizes its goal at the cost of ethical standards. Social consequences of such breakthroughs are significant as well.
While self-optimizing AI can revolutionize industries, from shipping to medicine, it disrupts labor markets. Schmidt’s threat that the coders will be out of a job is but one example of a larger pattern: while the AI tools are doing the tedious work, there may be less space for the old-fashioned jobs. And with that comes new possibility for experts to learn how to work in strategic professions where they have access to the AI tools and make them produce something which serves human purposes. Energy crisis also comes into the equation. As Schmidt was highlighting, the United States of America must reboot its energy policy to enable the expanding needs of AI. Natural gas, because of its scalability and dependability, can be a bridging fuel in an attempt to support such renewables as wind and sun. But the long-term solution is to re-design the grid and invest in clean technologies.
Schmidt’s farewell speech is a harrowing alert: “The scientists are in charge and AI is helping them – that is the right order.” As AI invites the future to be masterminded by it, the challenge is not merely technical but profoundly ethical and infrastructural. Policymakers, corporate leaders, and society at large have to thread this minefield with care so that AI breakthroughs are for the benefit of humankind and not to wipe it out.
