The Presage SmartSpectra SDK Challenge is live!
🔵 $3,000 prize pool across 10 winners
🔵 Build with real-time human signals like pulse, breathing rate & HRV
🔵 1 prompt + 2 prize categories = more ways to win!
dev.to/devteam/join-t…
Selective MoE offloading looked like a sure win for running big models on a 16GB GPU. Across five models, one ran more than four times faster with the old layer-splitting approach. 🧵
{ author: @sizzlebopzz }
"Ask anything" only works if the replies make people glad they asked. @sizzlebopzz looks at how a fifth of developers worry about being judged in communities, and what it takes to be welcoming in practice.
Neon Realtime kept filter, join and top-10 queries up to date in 40-46 ms, about one network round trip. @kubeden measured it against Supabase Realtime, LISTEN/NOTIFY and polling, and found the aggregate case where it sends more data.
Selective MoE offloading lost to traditional layer splitting on Qwen, LFM and Gemma in the first benchmark. Rerun at one quantization with interleaved runs, the picture changed. 🧵
{ author: @sizzlebopzz }
