engblogs

A practical framework for quantifying AI value in the AI age using a four-measure scorecard - Useful Intelligence per Dollar - balancing the cost per successful task, dependability, value at scale, and the work accomplished.

A practical framework for CFOs to quantify AI value by tracking Useful Intelligence per Dollar, cost per successful task, dependability, and value at scale across real workflows and compute.

Netflix details an in-house LLM serving stack that unifies deployment and inference across CPU/GPU, leveraging vLLM on Triton with a Model Scoring Service and an OpenAI-compatible API, plus constrained decoding and scalable rollout strategies for production-grade reliability.

Netflix details an in-house, end-to-end LLM serving stack built on vLLM and Triton, exposing a unified OpenAI-compatible API and employing careful deployment strategies for zero-downtime upgrades while implementing constrained decoding to ensure compliant output.

Netflix presents an end-to-end in-house LLM serving platform built on vLLM and Triton, unifying engine choice, packaging, and API surfaces (gRPC and OpenAI-compatible) to deploy, scale, and constrain generation in production.

A multidisciplinary journey of a researcher who chases questions across medicine, economics, and computing to design algorithms that enhance transparent, representative, direct democratic participation.

An interdisciplinary, curiosity-driven journey through medicine, economics, and computing that shows how cross-disciplinary research can produce algorithms and tools to improve democratic participation via citizen assemblies and enhanced public input.

Closing the Loop: a blueprint for turning recurring reviewer feedback into a self-improving AI system through an automated feedback loop that generates structured SKILL artifacts and is validated by a three-tier evaluation to converge on stable standards.

Explores engineering self-improving AI systems by turning reviewer wisdom into automated, structured feedback loops and a three-tier evaluation framework that continuously validates and refines generator outputs.

In the AI era, the cost of owning code changes has eclipsed the cost of writing it, so teams should use AI-generated patches as diagnostic probes to define scope, ownership, and risk before committing to delivery.

AI-assisted development shifts the cost from writing code to owning and validating changes, driving bounded patches, tests, and evidence-based reviews to decide what to ship.

Gemini 3.5 Flash Cyber introduces a lightweight, cost-efficient cybersecurity model built on 3.5 Flash and CodeMender to rapidly find, validate, and patch vulnerabilities at scale for defenders and trusted partners.