engblogs

An analysis of the text-speech understanding gap in LLMs, presenting SALAD (Sample-efficient Alignment with Learning through Active selection and cross-modal Distillation) as a data-efficient cross-modal distillation approach to improve alignment between speech and text in 3B/7B LLMs, with implications for multimodal decoding and streaming TTS as seen in Visatronic and SpeakStream.

depyf demystifies the PyTorch 2.x compiler by decompiling bytecode back into source code and exposing in-memory objects to on-disk sources for line-by-line debugging with a lightweight two-context-manager workflow.

An examination of Arvind KC's appointment as Chief People Officer and its implications for HR technology adoption, people analytics, and data-driven talent strategy.

RCCLX is the open-source AMD-optimized RCCL backend integrated with Torchcomms, introducing Direct Data Access (DDA) and Low-Precision Collectives to accelerate AI training and inference on AMD GPUs, with CTran integration enabling GPU-resident AllToAllvDynamic.

A one-week, AI-assisted reimplementation of Next.js as vinext on Vite, deploying to Cloudflare Workers with faster builds and smaller client bundles.

Explores AMUSE, a multimodal audio-visual benchmark for agentic multi-speaker understanding, and RAFT, a data-efficient alignment framework that enhances agentic reasoning via reward-based optimization and intrinsic self-evaluation in multimodal models.

A concise technical review of how Airbnb's engineering innovations intersect with academic publications to define the 2025 year-in-review in tech.

Build a scalable, serverless photo-search system that combines Amazon Rekognition for face and object detection, Amazon Neptune for relationship graphs, and Amazon Bedrock for contextual captioning to enable natural-language, semantic search across large image collections.

A technical overview of Piqama, an ecosystem for Pinterest quota management.

A concise deep-dive into reverse-engineering a handcrafted neural network puzzle using mechanistic interpretability and constraint solving (linear/integer programming and SAT) to expose an MD5-like computation encoded in its layers.

Examines HTML-to-Text extraction for LLM pretraining and shows that unioning multiple extractors increases token yield by up to 71% while enhancing coverage for structured content such as tables and code blocks without compromising benchmark performance.

A technical synthesis of Chain-of-Thought reasoning, introducing Trace Dynamics and a 'potential' metric to quantify how CoT steps influence the likelihood of correct completions, with insights on transferability across LLMs and implications for LRMs and VLMs.