engblogs

Cadmus is a small-scale autoregressive program-synthesis system with an integer VM and a DSL, enabling controlled experimentation to study inductive reasoning, training-distribution control, and affordable, transparent model analysis.

A Rust-based exploration of ecdysis, a graceful-restart library that enables zero-downtime upgrades for Cloudflare's high-traffic services by forking and execing a new process while preserving live connections.

Examines transferring optimal global hyperparameters across model sizes through a unified Complete Parameterisation that scopes width, depth, batch size, and training duration, and demonstrates per-module hyperparameter transfer to speed up training of large language models.

Faster convergence in federated stochastic variational inequalities is achieved by refining Local Extra SGD guarantees and introducing the Local Inexact Proximal Point Algorithm with Extra Step (LIPPAX) to mitigate client drift and extend guarantees to composite VIs.

A practical guide to validating and diagnosing AIPW-based doubly robust causal models in non-randomized settings, covering confounder management, propensity score corrections, overlap checks, and diagnostic scorecards to ensure credible treatment effect estimates.

Validation and diagnostics for doubly robust causal models (AIPW) in non-randomized settings, detailing confounder management, propensity-score diagnostics, sample balancing, and robustness checks to ensure trustworthy causal estimates.

A practical guide to unifying Databricks and cloud infrastructure costs across multi-cloud environments via the Cloud Infra Cost Field Solution, delivering a single TCO view for governance, optimization, and informed decision-making.

Examines MIT's J-PAL Project AI Evidence, a funding-backed initiative to rigorously test AI solutions, generate evidence on what works, and responsibly scale the most effective innovations to fight poverty.

Lambda appoints Jerry Hunter as Vice Chairman for Compute Delivery and Special Advisor to guide industrial-scale AI infrastructure and long-term compute strategy, drawing on his AWS, Sun Microsystems, and Snap experience to accelerate hyperscale AI delivery.

Trace length emerges as a simple, practical uncertainty signal for large reasoning models, revealing how post-training reshapes confidence estimates, how it complements verbal confidence, and its implications for robust uncertainty quantification across tasks.

A study-driven roadmap mapping the UX design space for LLM-powered computer use agents, presenting a taxonomy of design factors and empirical insights to align agent behavior with diverse user needs and scenarios.

A practical guide to building robust, long-running MCP servers on Amazon Bedrock AgentCore with Strands Agents, enabling cross-session execution via persistent memory and asynchronous task management for seamless results and reliability.