A UK AI Safety Institute report found that prior-generation AI models outperformed PhD students at troubleshooting laboratory experimental setups using smartphone photographs