engblogs

Eight frontier models are analyzed for how they recover from errors in agentic coding tasks, showing that recovery, not perfection, is the differentiator and outlining actionable patterns and fixes to boost resilience in automated agents.

GPU-accelerated serving of two-tower models enables lightweight ads engagement prediction.

Guidance on configuring AgentCore Browser for AI agents with proxy routing, persistent browser profiles, and Chrome extensions to enable secure, stateful, enterprise web automation.

Analyzes error patterns and recovery dynamics across eight frontier models on the Agentic Coding benchmark to reveal how resilience, not perfection, separates success from failure.

A concise, technical guide to scaling social science research by applying scalable data collection, analysis, and workflow methods to large datasets.

GPT-5.2 derives a novel result in theoretical physics, highlighting AI-assisted approach and its implications for future research.

A practical guide to scaling access to Codex and Sora beyond rate limits, outlining high-throughput API patterns and resilient access strategies.

Technical overview of Lockdown Mode and Elevated Risk labels in ChatGPT, detailing the security controls and risk-aware behavioral changes.

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.