Dom News

DeepMind's AI Control Roadmap argues that alignment is not enough and that advanced agents need defense-in-depth controls around access, monitoring, sandboxing, and permissions.

This is a useful engineering lens for agent systems: assume an agent can be wrong or misaligned, then design the runtime so sensitive actions require controlled, observable access.

• Official DeepMind post published with an AI Control Roadmap
• The post frames agent safety as system-level security
• It explicitly discusses permissions, monitoring, and defense in depth

• The roadmap is high-level and not an implementation guide
• Google's internal threat model may not map directly to small teams

Google DeepMind — Securing internal systems against increasingly capable and imperfectly aligned AI

The post turns a short PyCon lightning talk into annotated slides covering recent LLM movement, model behavior, tooling, and practical shifts for builders.

The value is compression. It gives enough context to recalibrate what to investigate next without reading every model provider announcement.

• The article links through annotated slides from a technical talk
• Simon Willison consistently tests AI tools through practical examples
• The format favors dense synthesis over launch marketing

• It is a synthesis piece rather than a single release
• Some conclusions reflect one practitioner's filter

Simon Willison — The last six months in LLMs in five minutes

Lore is a free, open-source version-control system announced by Epic for Unreal projects, aimed at teams dealing with large binary assets and game-production workflows.

This is interesting because game teams still live with Perforce-style constraints. If Lore handles large assets and editor workflows well, it could change how smaller Unreal teams collaborate.

• The article is tied to State of Unreal 2026 announcements
• The tool targets Unreal production workflows, not generic Git hosting
• Large binary assets are a real pain point in game development

• The announcement needs hands-on workflow validation
• Adoption depends on editor integration and migration cost

GameFromScratch — New Lore Version Control System from Epic Games

The article breaks MCP runtime infrastructure into the boring but necessary parts: OAuth token handling, user-scoped authorization, policy enforcement, tool execution, and auditability.

This is worth scanning because agent demos become real software only when tool calls are permissioned, logged, and revocable. The technical wedge is the action layer, not another chat UI.

• The piece focuses on runtime responsibilities rather than agent hype
• It covers token rotation, authorization, audit logging, and governance
• Arcade is directly building in the MCP infrastructure space

• It is vendor-authored and argues toward buying a runtime
• Claims should be compared with open-source MCP server patterns

Arcade.dev — Build vs. Buy MCP Runtime: 2026 Decision Guide

The article looks at Steam Next Fest discovery through AI disclosures, noting that the top-played demos include almost none of the hundreds of disclosed AI-content games.

This is a useful gaming signal because it separates a platform-level discovery problem from actual player pull. The question is not which game is biggest, but what players ignored.

• The article compares top-played demos with AI disclosure counts
• Steam Next Fest is a discovery surface for under-the-radar games
• It points to a real filtering problem for future game discovery

• Top-played demos can still skew toward already visible games
• AI disclosure data depends on developer self-reporting

PC Gamer — Steam Next Fest's top played games include only 1 of over 500 demos with an AI disclosure

The article filters a showcase down to a short list of smaller games worth wishlisting, including specific picks with mechanics and hooks instead of broad popularity rankings.

This is the gaming lane the site should track: curated discovery of games with unusual mechanics before they become obvious through mainstream word of mouth.

• The piece selects games from a niche indie showcase
• It names specific games and mechanics instead of ranking blockbusters
• Wishlist-worthy picks are closer to discovery signal than release churn

• Editorial taste still needs cross-checking against demos and Steam traction
• Some picks may already have publisher visibility

Game Informer — Eleven Games From The 2026 Six One Indie Showcase We Added To Our Steam Wishlists