Yatishara Blog
Paper digest: Video diffusion and world models — DiT, Sora-era, Open-Sora
From spacetime patches and Diffusion Transformers to open STDiT stacks. Research lineage only: what ‘world simulator’ claims mean, what Open-Sora actually ships, and where eval still lies.
Paper digest: RAG for cultural LLMs
Lewis et al. Retrieval-Augmented Generation and the dense-retrieval lineage. When “Cultural LLM” should mean corpus + retrieve, not only weights.
Paper digest: Preference alignment — RLHF to DPO in plain English
From human preference models and PPO to Direct Preference Optimization. What operators actually buy when a vendor says ‘aligned,’ and which failure modes the papers already named.
Paper digest: Omni-Diffusion
Nanjing University / Tencent Youtu / CASIA on masked discrete diffusion any-to-any. Text, speech, and images without an AR MLLM backbone.
Paper digest: Mixture-of-Experts
From sparsely-gated MoE to Switch Transformer and Mixtral. How sparse experts scale parameters without paying dense compute on every token.
Paper digest: LaViDa-R1
Adobe/UCLA/Georgia Tech on multimodal diffusion reasoning. Unified SFT+RL, answer-forcing, tree search, and the numbers that moved.
Paper digest: FlashAttention
Dao et al. on IO-aware exact attention. Why Transformers got memory- and bandwidth-efficient without approximating the math.
Paper digest: Creole & low-resource NLP
Tokenization, benchmarks, and data realities for Creoles and other low-resource languages — with JamPatoisNLI, CreoleVal, and NLLB as anchors for JA/TT speech-text work.
Paper digest: Caribbean speech ASR — Whisper, Creole gaps, and eval honesty
Whisper’s weak-supervision lineage, Jamaican Patois fine-tuning numbers, Caribbean emergency triage work, and why LibriSpeech wins do not transfer to WhatsApp voice notes.
Paper digest: Attention Is All You Need
Vaswani et al. 2017 in plain terms. What the Transformer claimed, what changed after it, and who should still care in 2026.