Delivering the AI Action Plan
Putting out the Bat Signal for people to help on the AI for Science and UK Sovereign AI agenda
In 2023, AI discourse went up a notch or two. Open letters, company earnings calls, public affairs events: you name your ’vibey deliverable’, and AI may well have been mentioned.
In 2024, the discourse around AI died down a little bit. People began to familiarise themselves with the free (and lower quality) versions of the generative version of the technology. Fears that deepfakes would seriously undermine election integrity never really came to pass. Reporting on the future of the technology suggested that there was a plateau in progress, implying that the hype was petering out.
Many criticisms of AI hype are true.
But those that call AI mostly hype, those that label it a bubble: I believe that they are far more wrong than they are right.
The following data points from just the month of December illustrate the magnitude of change underway.
150,000 trips a week are taken in Waymo’s autonomous cars. Compared with human drivers, Waymo trips have seen an 88% reduction in property damage claims and a 92% reduction in body injury claims. There is even some evidence to suggest that in the city of San Francisco, Waymo has a bigger market share than Lyft.
Previously impossible benchmarks in mathematics and general reasoning are beginning to be eaten by the state of the art, relegating discussion of LLMs being a ‘stochastic parrot’ to a relic of a bygone era.
An AI agent has been developed which is capable of reasoning across genomic contexts to predict functions of proteins. Anthropic has also released a feature on Claude which enables AI to dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.
Text to video model Veo2 now has a pretty impressive ‘world model’ of physics, with the way that items fall, reflections change, and shadows move, starting to look realistic. All other leading labs are working on similar tools.
While the media has been reporting on compute improvements ‘hitting a wall’, new dimensions of scaling have been discovered that we are only in the foothills of.
Each of these developments, in their own right, warrants societal focus. The fact that they are all taking place at the same time should demand societal prioritisation.
In 2025, I believe AI will return to being at the forefront of the discourse again. But this requires concerted effort from across the ecosystem (the s&t sector, civil society, politics, enabling sectors such as energy) to get involved and grip the matters at hand.
One policy programme I am excited to play a role in supporting is the DSIT AI Action Plan, led by Matt Clifford, co-founder of Entrepreneur First and chair of ARIA. It recognises the scale and pace of progress that continues to take place in the field, and lays out some clear policy priority areas to help the UK make the most of the promise of the technology.
I recently began a secondment into DSIT to work on the delivery of the plan. In particular, I am focusing on AI for science, as well as helping to set up the Sovereign AI team, which aims to maximise the UK’s stake in the frontier of AI.
Google DeepMind put out a fantastic report recently that outlines the role that AI will play in transforming the way scientific knowledge is distilled, data is generated, experiments/simulations are carried out, and novel solutions are identified.
The AI Action Plan is a great document, produced by very talented folks in DSIT. It lists AI for science several times as an area where the UK can build a meaningful stake in the frontier of AI.
The awarding of this year’s Nobel Prizes in Chemistry and Physics to AI boffins means that for science, AI is not merely some cute toy, but fundamentally re-ordering the nature of research. The recent o3 preview suggests that the sciences could be one of the early domains totally transformed by AI. But for this major opportunity to be fully realised, the following areas need to be improved:
Data infrastructure, both improving pre-existing data assets, and creating the next PDB!
New tooling for scientists to adopt AI effectively
AI-enabled physical infrastructure (wet labs, self-driving labs, etc), as well as compute
Skills to equip scientists and workers for a future of AI-mediated research and development
I want to hear from and be connected to people who can help me and the AI Policy Directorate (in the UK Department for Science Innovation and Technology):
Develop the AI for Science policy agenda.
Identify, enhance, and build the next generation of strategic AI assets in the UK
Broker partnerships to maximise the UK’s stake in the frontier of the technology
These sorts of step changes take place gradually, then suddenly. 2025 will feel sudden in many, many ways. If you’re keen to make sure such suddenness ends up positively impacting UK science and research, message me on here, Twitter, or Linkedin. If you want to help more broadly, get in touch with aiactionplan@dsit.gov.uk.