I've been using GitHub Copilot for several months - maybe 7 or 8. Recently, I've been ignoring the auto-complete suggestions and today, I'm disabling that feature. Copilot chat is still enabled but auto-complete is off. Reason - Copilot Contemplation: I found myself waiting for the suggestion to come back and give me what I mostly already knew. The "pause". I felt like I was more focused on "tapping the glass" then just doing the work. And then I would "wrong rock" the suggestion because it was close but, well, not right. I also don't want to rewire my brain to expect the suggestion. I want it to help, for sure, and to not lead. If you're using gen AI pair programming tools, any feedback on your experience? Do you sometimes turn it off? Any take-aways you'd share in how they are helping? While I'm only referencing GH Copilot here, any others you'd suggest to try out?
Tyler Veal
7mo
I’ve only used copilot indirectly through a peer. It seems to give quick fixes (when accurate) at the cost of losing understanding and critical thinking. When inaccurate, then we’re trying to understand the helper machine in addition to the original problem we were trying to solve 🐿️ I like to automate many things - automating critical thinking is something I try to avoid. Personally, I think if I use much AI in building software, I’ll forget how to build it. I admit though, copilot is pretty compelling, especially when hitting some hard walls. I also wouldn’t want to go back to searching books in a library - search engines on the web have automated that and beyond. AI in some contexts seems to automate the searching efforts even further 🤷♂️
Seniordev.ai
7mo
This is exactly the type of situation we're trying to remedy through our product SeniorDev AI. The idea is AI only steps in at the PR stage to do a code review, allowing the engineer to remain the leader, with AI simply supplying post code complete suggestions. If you check us out, let us know your thoughts! https://www.seniordev.ai/
Eric Swanson
7mo
All of the above. It's been generally helpful, especially in less familiar languages and with novel tasks (to me). It's been wrong many times and makes some bad guesses at the edge of documented solutions. For example, while scripting a containerized Nexus server in areas where documentation is poor or especially when internet content is old. I'm not certain yet, but it seems monorepos could give it superpowers with the extra context. I could see turning it off in very familiar languages and very familiar tasks. However, I've been toying with writing out a great, long comment and it generates the code surprisingly well. I think this type of comment-based prompt engineering is a fundamental productivity hack, but also a paradigm shift for devs. I can see there being resistance to this approach; possibly even seen as "not programming." Regardless, as Károly Zsolnai-Fehér from 2-minute papers says, "What a time to be alive!"
Varun V
7mo
I tried out a couple copilot and related assistants. Figured long term it hurts more than help building true expertise. Now I generally use all kinds of linters to ensure I write valid code with minimal suggestions, and extensively Chatgpt to give me boilerplate which I use as the base. AI embedded into the ide feels somehow demeaning even if really it’s possibly quite helpful
Tassio Carvalho
7mo
Jason Walker has such a great example of one's journey with Generative AI. It reminds me of a Google survey about Gen AI usage: Across 14,000 people, it found 28% had been already using Gen AI for work tasks in 2023. Among that user group, a majority was using non-approved tools. I wonder what that number is like today. This is a reality check for leaders out there who want to control and centralize access to Gen AI. Large corporations are far along their Gen AI journey, regardless of what leaders think, want, restrict or demand. Two important attributes of a Gen AI strategy are inclusion and transparency.
Ehsan Azizi
7mo
I’ve been using both Copilot and Amazon Q (previously known as Codewhisperer) for some time now and I can say Amazon Q response time and code accuracy and autocompletion is much faster and on point than Copilot. Not talking about chat part of it but in IDE environment of course.
Hajo H. Rappe
7mo
Contemplating the impact of auto-suggestions is wise. Sometimes, relying too heavily can disrupt workflow.
Flavio A. Castro
7mo
Yeah. Code generated quickly loses it's value but like said before it's good for boilerplate and bootstrapping. One tool I've been enjoying using is perplexity AI which feels like a better search engine.
Hang on tight … this is going to be a blast!
7moAn example of the dated content in GitHub Copilot. I'm wanting to pull a connection string dynamically and avoid more secrets in a GitHub Repo & Action. To protect the connection string value, I need to mask the output. I added some invalid syntax and then asked GitHub Copilot: "the masked connection string on line X is not available to other steps in the Actions job. Why? Suggestion back: "In GitHub Actions, you need to explicitly set the outputs of a step using the ::set-output name=value:: command." And then, Copilot suggested to add: echo "::set-output name=storageconnectionstring::$CONNECTION_STRING" Wha? This format has been in deprecated state since Oct 22. GitHub, update Copilot as to not suggest deprecated code flows around other GitHub capabilities. Please. Note: I wanted Copilot to call out the syntax issue on the last line: use "=" instead of "::"