yeswondwerr
2K posts
ok admittedly the movie was pretty meh but the sound system carried
English
Saw cursor wipe half the supabase migrations for a candidate’s on-site and then it ran the reset-db command without even asking lol
Bro didn’t even notice until I asked him where are the tables
GIF
English
Drawing my Moots: Final Update
This was fun! If I missed you there's always next time!
English
@stochasticchasm one for arxiv paper threads too that can be indexable and if u post slop u get ban
English
We gotta have a place to aggregate random anecdotal tips and tricks like this
Liliang Ren@liliang_ren
💡 Million-dollar hard lessons we’ve learned from pre-training at 1K GPUs with Standard Parameterization (SP): a. Use fp32 for unembedding weights when training with large vocabulary (200K), if you're using fused linear cross-entropy loss. Otherwise, Grad Norm will keep growing and blow up the loss eventually. (2/4)
English
@EsotericCofe i legitimately find comfy to be 10x harder to use than python, i dont get how people await comfy releases
English
@kalomaze data prep tools for audio is the bottleneck, at least in the open
English
@cloneofsimo >i have a problem
>looking for inspiration on ode relaxation shenanigans
>how are they doing it here arxiv.org/abs/2410.19814
>it resolves small scale physics
English
flamingo, metaquery, eomt show the underlying model is much more expressive/capable and were not bringing its full potential out
English