What role will open source play in the future of AI | Pulling the thread
You're invited to "Pulling the thread" - the next in a series of research dinners designed to facilitate an exchange of ideas across research and industry in frontier tech.
Hosted by a research collective including Fabric Ventures, Cambridge University, and AWS this evening is designed for academics, researchers, industry experts, and startup founders shaping the future of the compute stack.
What role will open source play in the future of AI
The modern technology economy was built on top of open source software. Linux, Kubernetes, PyTorch and countless other projects became shared infrastructure upon which some of the world’s most valuable companies were built.
AI is bringing this question back to the frontier. Despite the conjecture that the cost of training models would drive a meaningful performance gap between closed and open models over time, that frontier has not pulled away (if anything, it has narrowed). Some key stakeholders such as NVIDIA and Meta argue that open models are essential to innovation, sovereignty and Western AI leadership. Others, most notably frontier labs, warn that releasing increasingly powerful weights is dangerous and irreversible: the models can be modified such that safeguards are removed and their use can no longer be monitored. Furthermore, what began as an industry debate around software-development approach has become a question of geopolitical industrial strategy and security.
Key questions:
Is open-weight AI genuinely open source? Should we demand more insight into training data, training processes?
How does the evolution of the technological frontier change the balance between open and closed source; is there a stage at which customers are ok with open source capabilities behind the frontier as they’re “good enough?”. Do potential advances in harnesses, post-training, and/or test-time training shift the desirability of customisable open weights models?
Does openness create a competitive market for intelligence, or merely shift market power from model laboratories to chip and cloud companies?
Where should openness end? At what level of capability does diffusion become an unacceptable risk? Can open models strengthen defenders more than attackers, or does irreversibility fundamentally change the equation?
Could open source AI become Europe and the UK’s route to technological sovereignty, or would it simply exchange dependence on foreign APIs for dependence on foreign chips, data and model roadmaps?