iFeeling Daily
Daily curated AI insights you can't miss.
NVIDIA Cosmos 3: Open Omni-Model for Physical AI Reasoning and Action
NVIDIA's Cosmos 3 unifies world generation, physical reasoning, and action generation into a single open omni-model, eliminating the need to juggle separate models for robotics, autonomous vehicles, and smart space simulations.
Using Connected Sheets to Analyze BigQuery Data Without SQL
Organizations store their single source of truth in **BigQuery**, but the last mile of ad-hoc analysis, modeling, and reporting happens in **Google Sheets**. Exporting data as CSVs creates silos, version control problems, and governance risks. Business users often wait days or weeks for simple reports because they lack direct, secure access to live warehouse data without knowing SQL or database concepts.
Profiling PyTorch: Reading torch.profiler Traces from Scratch
A question-led walkthrough of reading PyTorch profiler traces, starting from the simplest possible operation and working up to torch.compile. The authors expose common pitfalls like cold-start overhead, hidden memcpy operations, and why torch.compile can actually increase CPU time for small ops.
Google’s Dataflow innovations for large-scale AI pipelines
Google describes how its internal Flume platform, which powers Gemini and Waymo, feeds into Dataflow for customers like Spotify, Etsy, and Moloco. Key innovations include liquid sharding for straggler mitigation, heterogeneous worker pools for TPU/CPU co-scheduling, and duty-cycle enforcement that scales down idle TPU workers.
A Guide to AI Cold Starts on Cloud Run
This article dives deep into the mechanics of AI cold starts on Cloud Run, breaking down the four phases and providing concrete strategies to reduce latency from 20 seconds to manageable levels, including quantization, startup CPU boost, and concurrency tuning from Google's formula.
(1D) Ordered Tokens Enable Efficient Test-Time Search
Researchers show that using 1D ordered tokenizers with a coarse-to-fine structure in autoregressive image generation significantly improves test-time search efficiency, enabling better scaling behavior and even training-free text-to-image generation when guided by an image-text verifier.
Nemotron-Labs Diffusion: Fast Text Generation via Self-Speculation
Nemotron-Labs Diffusion lets developers pick AR, diffusion, or self-speculation inference from one checkpoint—self-speculation delivers up to 6.4× tokens-per-forward-pass gain over pure AR with lossless accuracy. Practical, open, and ready to test through SGLang.
Amortizing Maximum Inner Product Search with Learned Support Functions
This paper proposes amortized MIPS, using SupportNet and KeyNet neural networks to directly predict maximum inner product search solutions, significantly improving IVF match rates on the BEIR benchmark while reducing computational cost.
OlmoEarth v1.1: 3x More Efficient Earth Observation Models
OlmoEarth v1.1 is a new family of Earth observation models that cuts compute costs by up to 3x while maintaining v1 performance. By redesigning how satellite imagery tokens are constructed, the model reduces token sequence lengths and makes planet-scale map refreshes more affordable for teams running OlmoEarth.