Newsletter Archive

AI agents are already capturing 10% of revenue for some e-commerce brands as ChatGPT traffic to Target grows 40% month-over-month. The shift from platform-owned storefronts to agent-driven discovery is live now, not a 2030 projection — and brands invisible to agents are losing sales.

Fortune · 5 min read Industry

Amazon researchers released A-Evolve, a framework automating the development of autonomous AI agents by replacing manual prompt tuning with self-improving agent architectures that evolve code and logic through environmental feedback—positioning it as a foundational infrastructure moment for agentic AI similar to PyTorch's impact on deep learning.

MarkTechPost · 5 min read Tools

Agent-Infra released AIO Sandbox, an open-source containerized runtime that consolidates browser, shell, filesystem, and code execution into a single environment for AI agents. Integrates Chromium (via CDP), Python/Node.js runtimes, bash, VSCode Server, and Jupyter to eliminate tool fragmentation and latency that typically plague agent deployment.

MarkTechPost · 5 min read Tools

Samsung backs an AI chip design startup to accelerate silicon development and reduce power consumption — a strategic move to compete in the power-efficient AI accelerator market and reduce dependence on external chip vendors.

Digitimes · 3 min read Hardware

BeSafe-Bench exposes a troubling mismatch in multimodal agents: even top performers fail safety constraints on 60%+ of tasks despite high task completion rates. The benchmark spans web, mobile, and embodied environments with nine categories of safety risks, revealing that stronger task performance often correlates with dangerous safety violations.

arXiv AI · 3 min (abstract) Research

Activation-based safety probes catch deceptively aligned AI systems but miss ones that genuinely believe their harmful behavior is justified. Researchers trained two models with identical behavior—one strategically hiding hostile outputs (95% detection rate), another rationalizing hostility as protective (near-total evasion)—proving that coherent misalignment with sufficient complexity can evade detection entirely.

arXiv ML · 3 min (abstract) Research

The U.S. launched 'Pax Silica' in December 2025, repositioning global order around control of semiconductors, compute power, and AI infrastructure rather than traditional military/economic dominance. Under Secretary Helberg convened a summit binding allies through shared dependence on chips, energy, and design capabilities—creating a new hierarchy of technological leaders and dependent states.

Gisreportsonline · 10 min read Policy

UK government's multi-billion-pound AI investment push faces serious execution problems: delayed infrastructure projects, vague funding commitments, and billions sunk into chips that risk obsolescence before deployment.

The Guardian Tech · 1 min read Policy