“Securing Liberal Democratic Control of AGI through UK Leadership”, 2023-03-14 ():
[followup, commentary] This was originally a piece co-written to influence policy makers, and had input from a range of people including senior frontier industry figures, former senior government advisers, and others who share the concerns raised in this piece. This is being published openly now in updated form given recent developments in the AI space.
Within this decade, we may build Artificial General Intelligence (AGI)—AI capable of performing most cognitive labour a human can do. Such a development would have an unprecedented effect on our society; ‘agentic’ forms of AGI may also pose an existential threat to our security. The current development path towards AGI is inherently unsafe.
The UK is in a unique position to alter this path in alignment with our values and for our benefit. However, this advantage has been squandered for a decade, and is now rapidly evaporating under an unsafe ‘race to the bottom’ dynamic between private companies funded by US tech monopolies.
Ensuring that AGI is developed safely and in the interests of the British people and liberal democracies must be the highest priority of the British state over the next decade. We propose this should be done through pursuing a multilateral approach to advancing and controlling AGI in partnership with our companies and liberal democratic allies. This should begin with creating a commercially connected elite public AGI lab under leadership of a frontier tech industry expert.
There is a brief window over the next two years in which rapid action is required to provide any chance of success. Specifically, this requires that we:
Procure national AI supercomputing infrastructure comparable to leading US private labs.
Create an advisory group of frontier tech, not legacy academic, expertise to identify major AI research projects to run on this infrastructure.
Grow an elite public-sector research lab, led by a leader with the technical skills and entrepreneurial expertise, to build a research agenda at the frontier of AI.
We invest almost $25 billion per year in R&D—a modest fraction of this must immediately be diverted to a national effort toward frontier AGI leadership.
…Single models at the frontier, namely OpenAI’s GPT-4 and successors, are being trained on tens of thousands of the highest specification GPUs (AI training chips) for months on end, roughly equivalent to using what is called an ‘exaflop’ supercomputer continuously for months. Unfortunately, the UK public-sector currently has less than 1,000 such top-spec GPUs, shared across all scientific fields. This means that one private lab in California is now using at least 25× the total compute capacity available through the entire UK state, just to train a single model. Our lack of compute undermines our ability to attract the best global talent in this technology; our business ability to commercialize and deploy it; and perhaps most critically, our state soft power over international use and control of it.
…The Review’s first key recommendation is to purchase one single exaflop supercomputer, roughly equivalent to 30,000 GPUs, for shared use by all UK research communities (not exclusive to AI) by 2026. This leaves the entire nations’ compute capacity in 2026 behind one relatively small frontier US lab in 2022. We emphasise that whatever we do procure will be very diluted versus OpenAI using it to train a single model—and it’ll arrive 4 years after OpenAI trained its model.
Leaders such as OpenAI will only continue to increase their compute usage. Another leading lab, Anthropic, has said that a state would require 100,000 top-spec GPUs within 3 years to be competitive in this space. This is a major upscaling of ambition merely to keep pace with these organizations that are still relatively small startups. Google, Facebook, Microsoft and others are all using even more, and the US will likely start building an AI supercomputing resource of 75,000 top-spec GPUs soon. As competition grows, it is a necessity that we have a sovereign supercomputing resource to enable our objectives in this space. As suggested by experts from leading AGI labs, we need to procure 100,000 top-spec GPUs for sovereign supercomputing capability dedicated to AI, for delivery ASAP.
The fastest any supercomputer could physically be procured is likely late-2024, leaving a vulnerable window of almost 2 years amidst intense and growing competition with other companies and nations.
…We should aim to rent 30,000 GPUs as soon as possible, and also build out dedicated engineering resources for using these. While this is lower than needed to match what startups such as OpenAI have access to today, the scarcity of available GPUs to rent already makes this ambitious to attempt. This could be done on a 2-year contract using cloud, or in partnership with domestic firms that are already substantial users of GPUs via cloud.