Reinforcement learning from human feedback (RLHF) is a popular technique for training high-quality AI assistants. However, RLHF may also encourage model responses that match user beliefs over truthful responses, a behavior known as sycophancy. We investigate the prevalence of sycophancy in RLHF-trained models and whether human preference judgments are responsible. We first demonstrate that five state-of-the-art AI assistants consistently exhibit sycophancy behavior across four varied free-form text-generation tasks. To understand if human preferences drive this broadly observed behavior of RLHF models, we analyze existing human preference data. We find that when a response matches a user's views, it is more likely to be preferred. Moreover, both humans and preference models (PMs) prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time. Optimizing model outputs against PMs also sometimes sacrifices truthfulness in favor of sycophancy. Overall, our results indicate that sycophancy is a general behavior of RLHF models, likely driven in part by human preference judgments favoring sycophantic responses.

Although Cloud computing emerged for business applications in industry, public Cloud services have been widely accepted and encouraged for scientific computing in academia. The recently available Google Compute Engine (GCE) is claimed to support high-performance and computationally intensive tasks, while little evaluation studies can be found to reveal GCE's scientific capabilities. Considering that fundamental performance benchmarking is the strategy of early-stage evaluation of new Cloud services, we followed the Cloud Evaluation Experiment Methodology (CEEM) to benchmark GCE and also compare it with Amazon EC2, to help understand the elementary capability of GCE for dealing with scientific problems. The experimental results and analyses show both potential advantages of, and possible threats to applying GCE to scientific computing. For example, compared to Amazon's EC2 service, GCE may better suit applications that require frequent disk operations, while it may not be ready yet for single VM-based parallel computing. Following the same evaluation methodology, different evaluators can replicate and/or supplement this fundamental evaluation of GCE. Based on the fundamental evaluation results, suitable GCE environments can be further established for case studies of solving real science problems.

The compass of Cloud infrastructure services advances steadily leaving users in the agony of choice. To be able to select the best mix of service offering from an abundance of possibilities, users must consider complex dependencies and heterogeneous sets of criteria. Therefore, we present a PhD thesis proposal on investigating an intelligent decision support system for selecting Cloud-based infrastructure services (e.g. storage, network, CPU). The outcomes of this will be decision support tools and techniques, which will automate and map users' specified application requirements to Cloud service configurations.

The Cloud infrastructure services landscape advances steadily leaving users in the agony of choice. As a result, Cloud service dentification and discovery remains a hard problem due to different service descriptions, nonstandardised naming conventions and heterogeneous types and features of Cloud