Boltzbit, a London-based AI research company, has launched publicly with a new architecture it calls General Learning Intelligence (GLI), positioning the technology as an alternative to the pre-trained foundation models that dominate the current AI landscape.
The company argues that today's frontier models, while responsible for rapid progress in AI capability, are expensive to build and remain static once deployed, unable to adapt or improve without retraining. Boltzbit's live-learning approach is designed to let models continue learning at inference time, adapting to new data and conditions as they are used.
"It is undeniable that the current approach has enabled rapid progress and brought advanced AI capabilities into the mainstream," said Dr Yichuan Zhang, Boltzbit's chief executive and co-founder. "But if AI is going to be genuinely democratised, access isn't enough and the current trajectory is unsustainable. We need systems that perform and learn in the real world, and we need to democratise model ownership at user level."
AI-360
Co-founder and chief technology officer Dr Jinli Hu said training remains the central bottleneck in AI development, and that the company had chosen to widen access to its technology now, warning that adoption is concentrating around pre-trained foundation models in ways that carry implications beyond commercial competition.
Boltzbit says its technology is already running in production, including within financial services and other data-intensive sectors, where live learning has been used to reduce latency between insight and action and to support more context-aware AI agents. The company plans to release its first full product suite later this year, covering the creation, training and hosting of models and the applications built on top of them.
Founded in 2020 by Zhang, a former Google AI researcher, and Hu, a former Microsoft AI researcher, Boltzbit builds on academic research into Boltzmann machines. The firm describes its work as part of a broader shift it terms "AGI 2.0", moving beyond static, model-centric AI toward adaptive, context-driven systems.
AI-360