(I'll join in on the fun!) I've been experimentin...
# devlog-together
s
(I'll join in on the fun!) I've been experimenting with using generative AI inside Github Actions lately and have been really intrigued by this async/event driven interaction model with code generation. In this repo we have a Github Action that runs every morning and has Gemini, Claude, and GPT-4 generate a PR adding additional functionality. Then on PR merge a release message is sent to our discord with example use cases and each of them riff on what was merged and try to come up with additional related functionality. I've already merged a few PRs and some of the ideas they've come up with are pretty interesting, especially the riffs - I merged a PR about querying a notion database, and Gemini came up with taking a row in a Notion database and creating an Asana task. Some of the things they've come up with so far:
k
Do the prompts aim in the direction of some rough goal or domain, or are they fairly untargeted, getting the AIs to add anything (that compiles?) to the codebase?
s
For the daily run its pretty untargeted...just "here's what Sublayer Actions are used for, here's a repository of ones we've found useful in the past, come up with some new ideas for things we might find useful and generate them"
but once something is merged it becomes a bit more targeted: "we just added an action to do X on this service, what other things might people want to do on that service?"
for this example its kind of anticipatory code gen
k
I see. So the act of humans merging something is guiding the computer to push more in that direction? This is cool. It reminds me of Doug Lenat's "Automated Mathematician" which would go off and search for interesting relationships between numbers and try to generalize from them. Once it defined some predicate or function over numbers, a human could go in and rename it (say if it was a well-known concept, like 'odd'). One of the heuristics it used was to pay more attention to definitions if a human went in and renamed them.
s
Yeah exactly! The main thing keeping it from growing more is that since these are external services I need to do a lot of work to verify them by hand, test API keys and stuff....working to come up with something with a little less manual review... But that Automated Mathematician looks really interesting, I hadn't heard of that before! It seems like I've been pulling on similar threads to that for a while now though with programming concepts instead of mathematical ones...