AI acceleration of science – a personal reflection

There is much about the AI revolution that is unpredictable and possibly unknowable, but there are three things that we can be confident about:

• It is disruptive, in breadth and depth
• It is rapid
• It is uncertain

As a research community, we need to complement the optimism and hyperbole about AI potentially accelerating science 10-100x and “curing all diseases” with clear-eyed discussion about the real bottlenecks ahead. Only then will we really maximise the potential of the revolution in cognition that we are embarking on and maintain public trust in science and its institutions.

Some of these bottlenecks are physical in nature, potentially solvable with technological advances. Other bottlenecks are psychological, sociological and institutional – people, groups and organisations have their own responses, feelings, concerns, and appetites for change. Technological solutions alone will not be sufficient

Execution of scientific projects involves pipelines of different types of tasks. Different kinds of science have different pipelines, although many tasks are shared, including idea generation, accessing funding, data analysis, publication, etc.

On the other hand, tasks like physical experimentation, gaining regulatory approvals, and accessing biological resources are not shared between biology and maths, for example. Even within biology, there can be substantive differences between observational science (characterising samples/data from the natural world) and experimental science (self-contained laboratory work). Moreover, most scientific projects require coordination both within and between institutions.

Current, and presumably future, AI technologies are well recognised to be ‘jagged’, that is their performance on different tasks appears to vary hugely in competence, relative to external expectations of difficulty. Although one could also say that evolution has equipped humans with highly jagged capabilities across different domains of cognition.

Therefore, AI tools are likely to dramatically accelerate some tasks within a scientific pipeline much more than others. And because scientific pipelines are different, AI will accelerate some kinds of scientific projects much more than others. The AI revolution currently happening in mathematics is one example of the acceleration that is possible when many of the tasks are readily catalysed by AI tools.

We have at least an initial understanding of which tasks in scientific pipelines are most readily accelerated by current AI tools. This means that, in theory, we can predict the near-term bottlenecks with some confidence. Examples of institutions actively addressing particular bottlenecks include AI companies investing in lab automation and research funders adapting their grant funding processes.

However, I think we could be doing much better as a community in identifying and tackling these bottlenecks together. I have two linked concerns. The first is that we are not taking a holistic approach as a community. Individual institutions, or types of institutions, are addressing specific bottlenecks that are apparent to them, particularly those that relate to their mission (e.g. research funders) or business model (e.g. AI companies) and are within their competence to address (e.g. technology companies focusing on technological bottlenecks). This means we are likely to pay less attention to some bottlenecks than others.

The second is that our thinking about scientific pipelines is rather implicit, simplistic and theoretical, rather than explicit, comprehensive and empirical. For example, recent thought pieces on the acceleration of science by AI have often relied on Popperian notions of science, and in so doing prioritised generality ahead of much-needed real-world specificity.

More generally, scientists (and I very much include myself in this) tend to adopt a somewhat intuitive, ad-hoc approach to how they conduct their own scientific projects and processes rather than a rigorous approach, informed by data on past projects (I think there are many interesting reasons why this might be, but these lie beyond the scope of this piece).

If you are a scientist who disagrees with this notion, then consider the following thought experiment. Think about the last three substantive projects that you led. How many individual types of tasks made up each project? Were they the same between each project? How much time did each task take for each project? How has this changed compared to the three projects before that? I literally have no idea what my answers to these questions would be. I don’t think about my research group’s scientific projects in this way.

Sure, I aim to learn from one project to the next and optimise tasks that felt like they should have been completed more quickly, but this is very much an intuitive, ad hoc approach rather than a rigorous, comprehensive and data-driven analysis. I don’t think I’m very different from the average scientist in this regard.

Before I get to potential areas, I think we should be focusing on more to maximise the acceleration of science using AI, there is one additional point that is worth making. It is well accepted that all technology-driven revolutions are accompanied by accelerated change in institutions and in society.

I believe we can learn a lot about the nature of this institutional change, the types and cadence of institutional change, from previous technological revolutions. The advent of the printing press and the Industrial Revolution come most to mind. I also believe that we can learn from evolutionary principles. A dramatic change in environment induces new selective pressures that act on the existing diversity of institutions and societies. The AI revolution is acting on a much more globalised world than previous technological revolutions.

There is a very real risk that the existing diversity of scientific institutions we currently have is insufficient for society to adapt as rapidly and successfully as we would like. I worry more about the lack of diversity in organisational models of non-profit scientific institutions (e.g. Universities), where there is much more global consensus on what it means to be a university, than I do about for-profit institutions, where there is continued innovation in business models and ways of working. The research institutions and research ecosystems that succeed will be those most willing and able to adapt to new realities.

However, perhaps we need to think more broadly about institutional and societal change. Why do institutions exist in the first place? It is broadly recognised that institutions exist to coordinate individual efforts, to stimulate cooperation, to achieve objectives that are hard to achieve by an individual in isolation, and with greater efficiency. One difference with the AI revolution is that ‘individuals’, in this context, are not just other humans, but also non-human agents. Arguably, this difference compared to previous technology revolutions could lead to greater and more rapid institutional change. There will also be considerable challenges in coordinating individual efforts both within and between institutions.

Here are a few suggestions, based on the thinking above, about where we could be focusing more effort, both within institutions and as a biomedical research community. This is very much a preliminary list that should change over time.

1. We need to have a much more sophisticated empirical understanding of the scientific pipelines of sequential and concurrent tasks that underpin the end-to-end process of different kinds of science.
2. We need to spend time as a biomedical research community thinking holistically about the likely bottlenecks to science, gaining a shared understanding of the current and future bottlenecks and recognise our shared responsibility for addressing these bottlenecks, both technologically and through institutional change.
3. We need considerably greater diversity of research institutions, especially non-profits.
4. Existing institutions, especially those with a high proportion of core funding, will need to adapt their speed of decision-making and flexibility of resource allocation
5. Existing and new institutions should engage with behavioural scientists, psychologists and sociologists to help us think through how our communities can adapt to the AI revolution
6. Particular current and near-term bottlenecks to focus collective attention on include:

• • The speed and diversity of research funding models and processes
  • The nature and speed of regulatory approvals
  • The nature and speed of contracting between institutions (e.g. for sharing resources)
  • Advances in laboratory automation to accelerate physical experimentation and enable recursive interactions between AI models and laboratory automation
  • Addressing barriers to analysing sensitive data with cutting edge AI tools,  maintaining high standards of research participant confidentiality and privacy
  • Using AI-based approaches for scaling ‘peer review’ and assuring the quality of new scientific knowledge
  • New modalities for sharing, updating, storing and querying collective scientific knowledge that are more scalable and more accessible than the current system of scientific publication.
  • Adapting notions of achievement and recognition to support new career pathways for scientists.