Jabil’s plant is doing precision manufacturing of electronics components for telecom, railways, and other sectors.
Interacted with the talented team of young engineers and technicians contributing to Jabil’s success.
📍Ranjangaon, Pune
The result: Better supervision signals across exits, stronger and stable guidance leading to far better quality in 4 steps v/s other distillation methods.
Open weights and technical report will be hosted on AIKosh soon. Thanks to everyone for the encouraging curiosity!
We released https://t.co/n71TnZbcMR on Friday to enable fast, high quality video generation.
Varya generates 720p video in 45 seconds on a single H200. Against a 50-step Wan 2.2 baseline under the same cost model, Varya is 27× faster and cheaper, while remaining close on VBench quality.
@RajanAnandan@brettsingh@Avataar@shubhamjain_11@rakesh_vRocky@Re_tot@imGauravB
Varya uses the CA-DM decomposition at sampled exits of a bidirectional video rollout. CA is the prompt-direction signal: it uses the teacher's conditional-unconditional gap to improve semantic alignment and visual detail.
DM is the distribution-shape signal: it compares the fake score with the frozen conditional teacher score and stabilizes the student-induced distribution.
Delhi: On the MeitY Secretary Programme, CEO of https://t.co/tgyeY5kzuw Sravanth Aluru says, "The vision we had when we started this journey was that the IndiaAI Mission has made significant progress in modalities such as text and voice, as we all know. However, to achieve a real-world AGI level, there are other modalities that we need to focus on equally…"
India is betting big on video AI with a govt-backed model that's 27 times faster and cheaper than China's.
https://t.co/fOVwcHSrHs Founder Sravanth Alluru discusses the move with @Rish_Bhat.