Roundup - Sick AI Tools For Developers

Episode Summary

Executive Summary: The episode surveys the rapidly expanding landscape of AI tools for developers, focusing on VS Code extensions, code search assistants, code review helpers, commit-message generators, and browser-based IDE models. Scott and Wes compare tools like Cody, Copilot Labs/Chat, CodeWhisperer, OpenCommit, Replit Ghostwriter, StarCoder, CodeGeex, Codium, and Tabnine, emphasizing practical use cases, cost, context handling, and the importance of verifying AI output rather than trusting it blindly.

Main Topics: AI tools beyond GitHub Copilot: The hosts frame the episode around the explosion of developer AI tools and note that the space is moving quickly beyond Copilot into many specialized products and workflows. Context-aware coding assistants: Tools like Sourcegraph Cody, Copilot Chat/Labs, and Codium are discussed for their ability to use repository or file context to improve completions, explanations, refactors, and test generation. AI for code review and linting: Lintrule is highlighted as a command-line AI reviewer that can catch issues traditional linters miss, such as privacy/security problems in logs or queries. AI for Git workflows and commit messages: OpenCommit is presented as a way to generate better commit messages and improve developer workflow by summarizing diffs and producing conventional-style commits. Alternative models and providers: The episode compares Amazon CodeWhisperer, Replit Ghostwriter, StarCoder, CodeGeex, Codium, and Tabnine, noting tradeoffs in quality, openness, latency, and enterprise privacy needs. Practical caution and verification: A major theme is that AI tools save time but are not trustworthy by default; users must still understand the code and validate output before shipping.

Key Arguments: AI development tools are becoming a broad ecosystem, not just a single product like Copilot. Context is the key differentiator: tools that understand the whole repo or surrounding files tend to produce more useful results. Copilot Labs and Cody are useful for refactoring, documentation, code smells, and test generation, but still experimental. CodeWhisperer offers a free Copilot-like alternative and is described as close in quality in the hosts' experience. Lintrule fills a gap by detecting issues that static lint rules cannot infer, such as leaking personal or sensitive data. OpenCommit is a strong AI use case because summarizing diffs and crafting commit messages is tedious but important. Browser-based and open-model options like Replit and StarCoder show the ecosystem moving toward openness and local/self-hosted possibilities. Even the best AI coding tools can produce bad or incorrect code; developers must review outputs carefully.

Data Points: Copilot Labs availability: VS Code Nightly - Copilot Labs chat/brush features were described as only available in Nightly at the time of recording. Copilot Labs usage frequency: 2-3 times a day - Wes said he uses the custom brush in Copilot Labs at least twice or three times daily. CodeWhisperer comparison: ~90% as good as Copilot - Scott estimated Amazon CodeWhisperer performed about 90% as well as GitHub Copilot in his testing. Cody AI query limit: 250 GPT-3.5 queries/month - The hosts noted Cody's free tier includes a monthly query limit. Lintrule pricing: $1 per 1,000 lines changed - The tool was discussed as potentially expensive for large code review workloads. Speed test ping: 10 milliseconds - Scott mentioned his wired connection ping as a baseline while discussing latency impacts on AI tools. Starlink ping: 100-150 milliseconds - Wes contrasted his cottage/Starlink latency with his city connection when talking about AI responsiveness. Token capacity example: An entire hour of transcript - They referenced newer ChatGPT token limits being large enough to accept an uncompressed hour of transcript text. Model breadth: 80+ programming languages - StarCoder was described as being trained on more than 80 programming languages.

Pivotal Quotes: "There's a whole world beyond GitHub co-pilot." — Wes: Opening framing for the episode’s main topic: the growing AI developer tool ecosystem. "This is a perfect use case for AI because once you're done doing all the code, now you got to spend a whole bunch of time figuring out what did you actually do." — Scott: Explaining why OpenCommit-style commit message generation is valuable. "These tools will help you accomplish tasks, but they're not necessarily going to do the task for you." — Scott: The episode’s main caution about relying on AI without human review.

Implications: AI tools are becoming standard dev infrastructure, especially for repetitive tasks, but the winners will be the tools that combine context, privacy, and reliability. Developers should expect faster workflows, more automation, and more need for judgment.