5/5 Paradigm shift: shifted from managing training time matrix transformations to active stateful multi-agent consensus protocols during runtime. Base model scale karne se bada engineering alpha ab distributed evaluation systems build karne me hai
1/5 Spent Sunday reading 'Attention Is All You Need' (2017) aur Runyang You aur Hongru Cai ka 'Agent as a Judge' survey paper back to back. 2017 ke foundational architecture se lekar 2026 ke agentic expansion tak ka shift bilkul insane hai.
4/5 Agentic Judge inference time compute use karta hai. Single model feedback ke badle dynamic agent loops separate runtime env deploy karte hain, tool calling data points collect karte hain aur process based reward routing se logical breakpoints trace karte hain
@trySynara it defaults to codex, even though ive disabled it as a provider in setting and also i want split panel opencode, kilocode, gemini, do i have to use it via new terminal session for each of em
So I am not an Android dev, but I got sick of cloud audio tools harvesting data to train commercial models. So I went back to first principles.
Just shipped Signal which is a 100% on-device AI audio restoration workbench for Android. 100% offline close to the metal.
For the UI, just used nano banana for design inspiration and stitched the code.
Generates high-fidelity .m4a and raw uncompressed .wav fallback tracks via Scoped Storage.
Code is public and out there. Repo link below. π οΈ https://t.co/fEiexu3jl8
Broke my head fighting the Android Media Framework:
-> 99% EOS Bug: OEM hardware encoders swallow BUFFER_FLAG_END_OF_STREAM. Fixed via coroutine yielding + 4s timeout failsafe.
-> Corrupt MP4: MediaMuxer breaks if CSD data/negative PTS hit writeSampleData(). Filtered them out.