Our research

AI that can accelerate research could drive a century of technological progress over just a few years. During such a period, new technological or political developments will raise consequential and hard-to-reverse decisions, in rapid succession. We call these developments grand challenges.

These challenges include new weapons of mass destruction, AI-enabled autocracies, races to grab offworld resources, and digital beings worthy of moral consideration, as well as opportunities to dramatically improve quality of life and collective decision-making.

We argue that these challenges cannot always be delegated to future AI systems, and suggest things we can do today to meaningfully improve our prospects. AGI preparedness is therefore not just about ensuring that advanced AI systems are aligned: we should be preparing, now, for the disorienting range of developments an intelligence explosion would bring.

Suppose we want the future to go better. What should we do?

One approach is to avoid near-term catastrophes, like human extinction. This essay series explores a different, complementary, approach: improving on futures where we survive, to achieve a truly great future.

The development of AI that is more broadly capable than humans will create a new and serious threat: AI-enabled coups. An AI-enabled coup could be staged by a very small group, or just a single person, and could occur even in established democracies. Sufficiently advanced AI will introduce three novel dynamics that significantly increase coup risk. Firstly, military and government leaders could fully replace human personnel with AI systems that are singularly loyal to them, eliminating the need to gain human supporters for a coup. Secondly, leaders of AI projects could deliberately build AI systems that are secretly loyal to them, for example fully autonomous military robots that pass security tests but later execute a coup when deployed in military settings. Thirdly, senior officials within AI projects or the government could gain exclusive access to superhuman capabilities in weapons development, strategic planning, persuasion, and cyber offense, and use these to increase their power until they can stage a coup. To address these risks, AI projects should design and enforce rules against AI misuse, audit systems for secret loyalties, and share frontier AI systems with multiple stakeholders. Governments should establish principles for government use of advanced AI, increase oversight of frontier AI projects, and procure AI for critical systems from multiple independent providers.

AI companies are increasingly using AI systems to accelerate AI research and development. Today’s AI systems help researchers write code, analyze research papers, and generate training data. Future systems could be significantly more capable – potentially automating the entire AI development cycle from formulating research questions and designing experiments to implementing, testing, and refining new AI systems. We argue that such systems could trigger a runaway feedback loop in which they quickly develop more advanced AI, which itself speeds up the development of even more advanced AI, resulting in extremely fast AI progress, even without the need for additional computer chips. Empirical evidence on the rate at which AI research efforts improve AI algorithms suggests that this positive feedback loop could overcome diminishing returns to continued AI research efforts. We evaluate two additional bottlenecks to rapid progress: training AI systems from scratch takes months, and improving AI algorithms often requires computationally expensive experiments. However, we find that there are possible workarounds that could enable a runaway feedback loop nonetheless.

Humanity is not prepared for the AI-driven challenges we face. But the right AI tools could help us to anticipate and work together to meet these challenges — if they’re available in time. We can and should accelerate these tools.

Key applications include (1) epistemic tools, which improve human judgement; (2) coordination tools, which help diverse groups work identify and work towards shared goals; (3) risk-targeted tools to address specific challenges.

We can accelerate important tools by investing in task-relevant data, lowering adoption barriers, and securing compute for key R&D. While background AI progress limits potential gains, even small speedups could be decisive.

This is a priority area. There is lots to do already, and there will quickly be more. We should get started, and we should plan for a world with abundant cognition.