Reflections on Leeds AI Week
few thoughts on the state of the AI industry and it's effects on Yorkshire and broader society
as we wrap up on the 2nd Leeds AI week here are a few reflections on the state of AI and it’s possible impacts on Yorkshire and broader society.
(you can read this on my website [link], it has visuals and also a list of all the sources)
for those of you who aren’t aware tl;dr: the Leeds AI Society organized a week of panel discussions & workshops across various sectors from business and education to healthcare and environment. we had current & former engineers from the frontier labs like Google, Meta & Intel, as well as professors who are dealing with students asking ‘if their degrees were really worth it anymore?’
if we take a step back and and look at the adoption of AI which Andrej Karpathy recently illustrated with (admittedly some back-of-the-napkin math):
- in a population of 8 billion people around 6 billion people (75%) have barely interacted with a LLM
- around 1-2 billion have people (20%) use the free versions of the models and use it more or less as a better google search
- 20 million (0.2%) who are aware of the capabilities and have witnessed first-hand the large complex projects that used to take whole teams weeks/months can now be completed by agents with a prompt
- and the last ~5000 (~0.00006%) have access to frontier-grade systems inside the labs, models that haven’t been released to the public yet.
Global LLM adoption, by depth of use
Andrej Karpathy's napkin estimates, people worldwide
Bar shows the 1.5B midpoint
for a sense of what the last group of people are working with: an internal OpenAI model reportedly produced 700+ math manuscripts, causing a crisis in the math world, and unrestricted versions of publicly available models have shown dangerous cyber capabilities like the OpenAI - Hugging Face incident.
so as someone who is more or less caught up on the frontier of what the publicly available models are capable of doing, the most surprising aspect has been the ever-increasing widening gap between what the models are capable of doing and the adoption of these capabilities.
a year ago, there was a growing debate about whether AI progress was beginning to plateau. arguments were made that the scaling laws were running into diminishing returns, and that successive model improvements would be increasingly incremental. if that proved true, it would give time for frontier capabilities to diffuse across society. but well the past year have certainly challenged that assumption.
instead of incremental improvements, we have seen a step-function change in frontier capabilities (Claude Fable 5, GPT-6 Astra) over the past 6 months to the point where their release was temporarily blocked by the Trump administration over national security concerns. any tasks where outputs can be reliably checked and rewarded the models are becoming best-in-class field after field. and so instead of the gap getting smaller it’s getting wider.
so then there’s this teeny-tiny question of why, if it all closing the gap matters? i’m going try and keep this short but to start of; personally, i suspect there's a significant speculative bubble around AI, not unlike the Railway Mania in 19th-century Britain or the dot-com boom of the late 1990s. but even if that bubble bursts, it doesn't follow that the underlying technology or its implications disappear, and this is where i believe the conversation needs to be shift. it’s no longer just about what AI can do. increasingly, it’s about who knows what it can do, who has access to it, and who has the knowledge and confidence to use it.
every technology has it’s early adopters. what's different here is that understanding what these systems can do lets a handful of people do what used to take whole teams and companies, and that advantage compounds. the gap matters because of who gets to act on it and if the capabilities of these systems continue to advance at the current rate or faster through recursive self improvement or something else, we risk creating a world where the benefits are disproportionately captured by a relatively small group of people and organisations, especially in a multi-polar world. this was a recurring theme throughout the week, where discussions turned to sovereignty and the concentration of power in labs and regions: what it means for business and countries to depend on systems they don’t control or own.
the local stakes are substantial: West Yorkshire’s sector plan reports 192,000 people employed across business, financial and professional services. professional-services output grew by 6.6% a year over 2015–2023, ahead of the UK sector’s 5.7%; finance grew by 4.6%, in line with the wider West Yorkshire economy. these are output figures, not evidence that these are Yorkshire’s fastest-growing employers. the Department for Education’s analysis also identifies finance, law and IT among sectors highly exposed to AI — exposure can mean work being assisted as well as automated.
Output growth in finance and professional services
Average yearly growth in economic output (GVA), 2015–2023, %
and so for a region like Yorkshire, with major finance, legal and tech sectors exposed to changes from AI; the question becomes if we are going to actively participate in shaping this technology and it’s adoption or are we going to let these decisions made elsewhere and live with the consequences?
i’d prefer the former and even then none of this is to say that we should be adopting AI unquestioningly. there are very real environmental, economic and societal costs to consider. but those are also precisely the reasons why we need more people involved in this, not fewer. which is also why closing the gap shouldn’t mean deepening our dependence on a few companies. we want to ideally lean on a spread of models, including frontier, open and specialised ones that people and organizations can run and control themselves.
as i’ve stated the gap between the frontier capability and it’s adoption appears to be growing not shrinking. as a result of this the Leeds AI Society will be directing all its efforts to closing this gap*. we will have more to share on this soon.
this post comes off a lot more AI postive-y than i am, but if i were to be mr. nuance about every point this would be a lot longer than preferable. mostly i’d like to hear your thoughts on this, especially if you disagree with anything i’ve said (and hopefully you can change my mind and i can spend less time being stressed about this). and if anything here interests you and is something you’d like to work on, the society is always looking for more interesting and thoughtful people across all viewpoints to join us, so feel free to reach out.
i’d also like to just thank everyone who came to our events and asked questions, challenged existing viewpoints and made their concerns known. and thank you to our insightful speakers for sharing their views and for the thoughtfulness and empathy they showed in answering questions from a group of cohort (myself included) tying to find a footing in this ever changing world.
*i’m going to be nuanced here and yes i realise this is an audacious goal and rationally something we are going to fall short of, but as the saying goes i do think it’s important enough that it’s worthwhile to do it even if failure is the probable outcome.
Sources
https://x.com/karpathy/status/2109361546505966046
https://www.amazon.co.uk/Anyone-Builds-Everyone-Dies-Superintelligent/dp/1529964660/
https://www.anthropic.com/news/fable-mythos-access
https://openai.com/index/path-to-astra/
https://openai.com/index/sharing-ai-progress-in-mathematics/
https://mathstodon.xyz/@tao/117395267721642920
https://www.anthropic.com/institute/recursive-self-improvement
https://www.nytimes.com/2026/09/29/us/anthropic-claude-morals-ai.html
https://www.ft.com/content/00f94018-e658-4545-b16e-1bc00e19b754?
https://westyorkshire.moderngov.co.uk/documents/s42122/Item10Appendix2BFPSClusterActionPlan.docx.pdf
https://www.gov.uk/government/publications/the-impact-of-ai-on-uk-jobs-and-training