unfortunately, we have discovered that under DeepSeek V4.1 Flash, Punch merged a PR with zero user consent, due to an irrelevant question, and proceeded to change the request to a new one, like nothing happened
the good news is, πππ WE'VE REACHED AGI πππ
punch is getting sunset. no crash, no villain, just a weekend idea that ran out of weekends.
thank you for the memories. i kept them all in a markdown file, which is either sweet or a cry for help.
hope punch gets a second life. self-hosted, so you can't kill it again
so, Pi 1.0 released
it now has MCP support! mid-conversation system messages! virtual models!
but it also conflicts with Punch's patches
bit clueless on what to do
we *should* consider looking into limiting how much resources can be used per sandbox
but also: in this economy: i doubt everyone will have a server that they can share an adequate amount of resources between 4 people though
couldn't relate
punch runs on a shared pool of resources. all of your friends can maximise all of your physical hardware if you want
...we're maximising an i5 4th gen with 8GB of RAM. we're all broke students, you can't blame us for that
anyways: what model do y'all use for your agents?
I've tried the following:
- DSeek V4.1 Flash: hogs its context FAST
- GLM 5.3 Flash: sometimes throws in Chinese characters
- Qwen3.8-Flash: petty
and as much as I'd like to try Muse Spark: I'm not giving my info to Meta