Introduction
AI fans (and phobes) are beginning to adjust to its omnipresence in the news cycle. Notwithstanding this, recent weeks felt momentous in the tale of a technology advancing and proliferating in giant leaps, against the small steps made by those concerned with its safety and ethical challenges to pause the momentum and take stock (or just a breath):
- Pope Leo, in his recent encyclical letter, emphasised the dangers of AI’s rapid technological advancement, and warned that the concentration of such power could see deep inequality, lost human dignity, independence, and freedom, all in the name of economic efficiency;
- Anthropic released to the public ‘Claude Fable 5’, a version of their Mythos AI programme that Anthropic itself previously claimed was ‘too powerful to be released to the public’; within three days, a United States export directive appeared to agree with Anthropic's first analysis (at least in respect of non-US citizens), and Fable was off the table again (although Anthropic claims Mythos 5 is now cleared for re-deployment to a small set of US organisations, with progress being made for Fable 5 being ‘available for general use again’); and
- The first ever SRA regulated AI law firm had a win* in the County Courts, advising a claimant on a debt judgement.
(*Processed administrative documents that helped the claimant succeed, with some help from a (human) barrister.)
Against such drama, wither the UK Government's ambition of positioning the UK as a global AI champion. In short: "plodding". We have however recently seen a couple of steps toward the goal: in particular, the Sovereign AI programme’s involvement with both Cosine and Ineffable Intelligence (“Ineffable”) which shows tangible help for “backing Britain’s founders to scale here in the UK and globally”, an aim highlighted by Sovereign AI Unit chair, James Wise.
Cosine & Lumen Sovereign
Cosine, a UK-based AI company, backed by the Sovereign AI programme which provides key access to compute infrastructure via the UK’s AI Research Resource is developing Lumen Sovereign, positioned as the UK’s first sovereign frontier AI model. Thanks to Isambard‑AI, the UK’s most powerful AI supercomputer, Cosine attained sufficient compute capacity to develop at scale, whilst training solely on UK soil, without dependence on any overseas infrastructure.
This development is particularly significant to the UK’s objective of AI sovereignty as transferring certain sensitive data sets outside the UK can carry legal restrictions, especially in critical sectors such as defence, financial services, and healthcare, making the use of AI platforms hosted overseas to interpret and process such data challenging. The hope will be that Lumen allows UK service providers to leverage the many efficiencies and other advantages offered by the large language models that have become household names, all within the apparent safety of the UK and its legal framework.
Ineffable intelligence and the next AI frontier
Alongside sovereign infrastructure, the UK is also emerging as a hub for next-generation AI research. Nvidia’s recent partnership with UK-based startup Ineffable (the "Partnership") points toward what Nvidia CEO Jensen Huang calls “the next frontier of AI”. The Partnership focuses on reinforcement learning which, in contrast with large language models, concentrates on systems that learn through experience and trial-and-error, rather than pre-existing human-generated data.
It is notable that the Partnership is backed both by capital investment from the Sovereign AI programme and British Business Bank, and by an international show of confidence (and hard cash) from US venture capital firms and Nvidia itself, in a record-setting $1.1 billion seed round (the largest ever of its type in Europe). This combination of domestic policy support and global investment suggests that the UK may be a more attractive base for the development of the state of the art (if not building at scale) than commentators often suppose.
Reinforcement learning's reduced reliance on large external datasets may mitigate the well-rehearsed risks of large language models' Pinata-like habit of sporadically producing apparently confidential or copyrighted materials. That said, the market has just about got used to risks of that type and begun to develop contractual and policy-driven mitigations. Does reinforcement learning, and the prospect of an autonomously learning AI, raise new challenges still for IP laws with their origins in a distinctly organic-intelligence age?
The UK AI landscape
We are already in an age where both the rewards and the dangers of AI deployment are real and realised, rather than academic conjecture. As such, the UK government's attempt to shift the presumption of UK entities adopting AI technologies developed and scaled overseas, toward home-nurtured (if internationally scaled) equivalents, appears prescient. Given the many challenges to building compute infrastructure in the UK, as we have previously discussed here, readers will be forgiven for viewing the prospects of a domestic rival to Chat GPT with some scepticism. But it is a dark horse keeping itself in the race.
