tl;dr:Abi discusses the following: (1) Avoid single-metric models. (2) Measure all outcomes you care about, and capture multiple metrics for each outcome. (3) Be mindful of incentives created by measurement. (4) Measure different facets of productivity. (5) Use system-based and self-reported data together.
tl;dr:Google‘s Engineering Productivity Research team sharing how they define Software Quality. For engineering leaders focused on quality, this paper provides a helpful framework for articulating what it means and where improvements need to be made. The paper posits there are four types of quality that influence each other: process quality, code quality, system quality, and product quality. Abi discusses each in detail.
tl;dr:Google‘s Engineering Productivity Research team sharing how they define Software Quality. For engineering leaders focused on quality, this paper provides a helpful framework for articulating what it means and where improvements need to be made. The paper posits there are four types of quality that influence each other: process quality, code quality, system quality, and product quality. Abi discusses each in detail.
tl;dr:“To provide better insights, Google researchers identified the key goals developers are trying to achieve in their work and developed measurements for each goal. In this paper, they explain their process and share an example of how this new approach has benefited their teams.”
tl;dr:“This study explored what causes cognitive load in an international corporation. While every organization is different, the cognitive load drivers identified in the study provide a helpful starting point for leaders who aim to make it easier for developers to ship software.”
tl;dr:“To evaluate whether Copilot should be used org-wide, the authors of this paper conducted an experiment for six weeks, and compared the tool’s impact on a test group versus a control group. They based their evaluation of the tool’s impact using measures for productivity, quality, and security.“
tl;dr:(1) Human memory is complex, with recall activating a network of neurons that can lead to unexpected insights. Stepping away from problems can facilitate innovative solutions. (2) Long-term memory, as opposed to working memory distinguishes experts. Hence, cognitive load becomes a factor. This can be reduced by simplifying tasks or improving information presentation. (3) Experts recognize patterns quickly, while beginners reason line-by-line. Reading more code helps beginners become experts faster. And more.
tl;dr:(1) Human memory is complex, with recall activating a network of neurons that can lead to unexpected insights. Stepping away from problems can facilitate innovative solutions. (2) Long-term memory, as opposed to working memory distinguishes experts. Hence, cognitive load becomes a factor. This can be reduced by simplifying tasks or improving information presentation. (3) Experts recognize patterns quickly, while beginners reason line-by-line. Reading more code helps beginners become experts faster. And more.
tl;dr:In 2020, the code review team at Meta discovered that 85% of developers were satisfied with the code review process in general. They were less satisfied with the speed with which their code was reviewed. This inspired a core hypothesis that the NudgeBot could decrease code review time in 3 ways: (1) The time a diff waits in the ‘needs review’ status. (2) The number of diffs that take over 3 days to close, this timeframe was chosen because they were only nudging diffs after 24 hours. (3) The time to first action.
tl;dr:The report focusses on LLMs e.g. GitHub Copilot and its impact on software development, suggesting it has the potential to improve productivity and reduce cognitive load. However, its benefits are distributed unevenly across users and it introduces new challenges. Key takeaways: (1) Benefits of LLMs in software engineering depend on the specific task e.g. easier to start a project with an LLM but difficult to change generated code. (2) Issues arise with writing prompts and overreliance e.g. burdensome to inspect code, accepting incorrect code. (3) LLMs help the least experienced the most. (4) Adoption is influenced by how well AI tools fit within workflows. (5) Analyzing and integrating information become more important than generating code.