Summary

• London-based Boltzbit has launched a new brand around General Learning Intelligence, or GLI.
• The company says its live learning models can adapt in production rather than relying only on costly pre-training and retraining.
• The proposition touches model ownership, training cost, enterprise data control, and concentration in frontier AI markets.

Boltzbit has launched a new brand and public push around General Learning Intelligence, arguing that enterprise AI needs live learning systems that can adapt in production rather than remain dependent on static pre-trained models.

The London AI research company, founded in 2020 by Dr Yichuan Zhang and Dr Jinli Hu, is building on research into Boltzmann machines to develop generative models that learn from new data and environments at inference time. Boltzbit says its technology is already in production across financial services, data, and high growth technology sectors, with its first product suite planned later this year.

The company is targeting one of the most important structural issues in AI adoption. The current frontier model market is dominated by a small number of companies that can afford massive pre-training runs, large data infrastructure, and continuous model improvement cycles. Enterprise customers can access those capabilities through APIs, cloud platforms, and software products, but they rarely own the model layer or the economics of training.

Boltzbit says General Learning Intelligence is intended to change that model. Its live learning approach is designed to allow AI systems to improve through use, adapt to context, and support user level model ownership. The company says this can reduce the latency between insight and action, support more context aware AI agents, and address the cost bottleneck created by traditional training.

Dr Yichuan Zhang, Boltzbit’s CEO and co-founder, said the current approach has brought advanced AI into the mainstream, but warned that access alone is not enough if the trajectory remains expensive and centralised. Dr Jinli Hu, CTO and co-founder, said training is the real bottleneck and that live learning models have already shown production value through work with selected corporations.

The commercial significance sits in the tension between AI adoption and AI dependency. Many organisations are moving quickly to use foundation models, but they are doing so through systems they do not control and often cannot adapt deeply without vendor support, retraining, fine tuning, retrieval architectures, or layered applications. That can work for many use cases, while creating limits in environments where data changes quickly, decisions are context specific, and intellectual property or regulatory constraints require tighter control.

Financial services is an obvious test case. Markets, risk signals, and client behaviour shift constantly, while regulated institutions are wary of black box systems that move data or decision logic outside their governance boundaries. A model that can adapt live while maintaining data control would be attractive only if it can prove reliability, auditability, and safe behaviour under changing conditions.

First principles AI claims need evidence. The market is crowded with companies arguing that transformers, large language models, or centralised foundation models are insufficient. Some of those arguments are technically plausible, but buyers will ask whether alternatives can match performance, integrate with existing workflows, and avoid introducing new risks around drift, safety, security, and validation.

Live learning also creates governance questions. A model that changes through use may become more useful, but it also requires monitoring. Organisations need to know what has changed, why performance has improved or deteriorated, whether unintended behaviours have emerged, and how the system can be rolled back or constrained. In regulated sectors, continuous adaptation has to be paired with evidence, controls, and clear accountability.

Boltzbit’s launch does not overturn the foundation model economy. It does, however, ask whether the next phase of enterprise AI will be built mainly on rented intelligence, or whether more organisations will seek AI systems they can train, adapt, and own more directly.