Peter Thomas

checking in @OpenAIDevs DevDay Exchange thx @mvkaran

anyone else seeing AI slop comments on open @github issues? mildly irritating passive aggressive nonsense @GitHubEng any suggestions?

Spent quality time at @APIdaysGlobal India last week and it was extremely valuable, exchanging notes with industry-experts and fellow-builders. Thank you Mehdi Medjaoui and Naresh Jain for the invite. I actively sought out opinions from API enthusiasts ranging from fresh graduates to veterans and asked many questions. What are teams seeing on the ground? How is adoption of AI going within the enterprise SDLC? Are you releasing to production without reading every single line of code? I also got asked a lot of questions. Does AI worry you? Won’t teams just create their own test-automation tools? I have certainly given a lot of thought to this question! My answer is that I’ve seen many teams decide to build their own API testing tool instead of choosing something open-source or off-the-shelf. Things look great for a while, the developers get a pat on the back, maybe even promoted. But guess what, after a while people tend to move to other projects or leave the company. And then the tool stagnates and becomes technical debt. It drags down the team and becomes one of those things that people don’t want to touch or work on. So yes, it is the same argument. You can vibe code a testing tool for your project, sure. But is that really the best use of your time? Are you sure you want to maintain this very important part of your pipeline? I am very thankful to be here in this moment where LLMs can amplify your unique skills, knowledge and experience. I have been shipping open-source software since the days of SourceForge - before Git was even a thing. In my opinion, the gap between an average developer and a good one will widen with AI. AI amplifies the good parts. It also magnifies the bad. Teams that are not mature at requirements and verification will be hit hard when coding stops being the bottleneck. When code-review is no-longer humanly feasible, teams will have their hands full keeping up with the velocity expectations of a gen-AI world. There will be no time to think about building testing tools. The only option is to focus on the core-business. Leave the business of testing tools to the experts! Here below is one slide from my talk. When you can’t review all the code, you have to rely on some visualization of what the LLM has done. I think we need to prepare ourselves for a future where you are looking at more dashboards and reports. But you are still accountable, you can’t shrug off an incident in production with the excuse that “it was written by AI”. That is the human approval step, auditable as a Git commit.

AI generated instrumental music but ...I actually like it. 😱 youtu.be/i0YZ-iB9s5Q?si=oyZQ…

yay, achievement unlocked, thx @vocastats youtu.be/XrHUHY0IwR8?si=z5FR…

engineers after getting claude to write an agent loop

maybe this will make your day - if you also come from the 80's and are a fan of the anime this references youtu.be/hjtNhzdxkHc?si=2F99…

instead of full blown MCP + json-schema approach, I suggest claude code (or eq) just writes pure python over HTTP, much faster, full control. sounds like you may have figured this out already, and your MCP tool is just an envelope over the blender python API. here's my initial version, already able to do a lot, and quite fast ! planning to automate the video sequence editor because I need it: github.com/ptrthomas/blender…