Researchers at Imperial College London and Thomson Reuters report a method that could help governments and other organisations run their own highly capable artificial intelligence (AI) systems. The goal is often called sovereign AI, meaning a system that an organisation owns and controls . Thomson Reuters used the method to build a model named Thomson. The company says the model can match much of the performance of the largest general models. Thomson’s reported development cost is 40 million dollars, far below the billions often spent to train those leading models from the beginning.
Thomson started from the open-source and open weigths AI model Qwen. The open question was whether anyone outside the largest laboratories could improve such a model on new work without ruining what it already knew. The report says that is possible.
How forgetting and values are handled
A central problem is catastrophic forgetting: when further training teaches new skills, older skills often fade. After testing several remedies, the researchers found a useful recipe. They first trained the model hard enough to get better at the new tasks. They then blended the new weights with the original weights, so the finished model kept both old and new abilities.
They also tried to steer the model toward safety and impartiality. They looked for patterns in the model’s internal activations, the internal signals that appear when it processes an input, that lined up with unwanted behaviour. An optimisation method then weakened those patterns while trying to keep useful skills. Desired answers that followed the constitution were reinforced with further training.
The researchers said that building a sovereign system used to mean training from scratch, and that even a billion dollars might only produce something competitive, and this alternative changes that. They added that useful AI needs domain knowledge and real-world context, not only larger models and more computing power. The work is described in a Thomson Reuters technical report.