Barclays cutting ~5,000 jobs and calling AI the icing, not the cake
AspireAI community · BBC News
Barclays said 1,000 posts have already closed and it expects roughly 4,000 more by the end of 2026, around 5% of its global workforce. The reductions focus on UK retail banking, with some UK back-office work done in India, where a significant share of the cuts is expected to fall. Around 60% sit in support functions, including among developers, and the investment banking division is also affected. Group chief executive CS Venkatakrishnan was explicit about the role of AI: "We will employ the AI at scale. But AI is the icing on the cake not the cake. Because you need to do the infrastructure build to get the benefits of AI." The cuts close out a three-year cost-cutting programme, and Barclays expects its efficiency measures to deliver a further GBP1.1bn of savings in 2027. Q3 profit was GBP2.3bn for the three months to September, up year on year but slightly below forecast. Analyst reaction was mixed: Russ Mould of AJ Bell said Barclays was "reaching for the AI lever" but had failed to match expectations with its financial results, while Will Howlett of Quilter Cheviot said the cuts show how far AI and automation could reshape banking over the next few years. The honest limitation: Barclays frames this as an ongoing modernisation plan and attributes no specific share of the job reduction to AI, so "5,000 jobs because of AI" is headline framing rather than a measured causal claim. Source: BBC News, Oct 22 2026.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.