Google partners with EVE Online for AI testing

Google DeepMind tests AI models in EVE Online

Google DeepMind has partnered with CCP Games to test AI models in the virtual universe of EVE Online, following CCP’s $120M independence funding and rebrand to Fenris Creations. The collaboration leverages EVE’s persistent live environment to evaluate AI behavior under complex, player-driven conditions. Testing includes autonomous agent interactions and decision-making in a 16-year-old sandbox with 800,000 registered accounts.

EVE Online’s ecosystem provides a unique lab for AI because of its player economics, territorial conflicts, and emergent systems. The game’s physics engine and economic models create unpredictable challenges for AI systems. Fenris Creations CEO Hilmar Pétursson confirmed the partnership would focus on reinforcement learning techniques but declined to disclose specific technical benchmarks.

AI search integrates Reddit for expert advice

Google has updated its SGE AI search to pull “Expert Advice” from Reddit, expanding beyond existing sources like Wikipedia and YouTube. The feature surfaces relevant Reddit threads in response to technical queries about software development, hardware troubleshooting, and niche hobbyist topics. Users with Reddit Premium accounts see priority access to verified content.

The Reddit integration follows months of testing with Stack Overflow and specialized forums. Google’s algorithm maps query intent to subreddit domains using natural language processing. Early test results showed a 37% increase in accurate technical answers compared to generic web results. However, the system struggles with contextually sensitive topics where Reddit communities have conflicting information.

Open-source tool bridges Google Sheets and LLMs

An open-source extension called AISheeter now allows users to integrate any LLM with Google Sheets using their own API keys. The tool supports GPT-5.4, Claude Sonnet 4.5, Gemini 2.5, and Groq models with features like self-correcting formulas, progressive reasoning disclosure, and session memory learning. Developers can automate tasks like sentiment analysis, feature extraction, and urgent prioritization with natural language prompts.

The extension’s architecture includes an evaluator-optimizer pattern that catches errors in column references and logical parameters before execution. Performance tests show 10ms latency for cached operations using pgvector semantic search. The tool’s transparency features let users see the model’s thought process before final output. While the tool works well for structured data tasks, complex financial modeling still requires manual validation of AI-generated formulas.

Phoenix SaaS template gains traction

A modular Phoenix-based SaaS starter kit is gaining traction among indie developers, offering pre-built components for authentication, payments, and AI integration. The template includes production-ready Stripe and LemonSqueezy payment systems, Ecto-optimized PostgreSQL queries, and real-time analytics dashboards. Developers can deploy to Fly.io in under five minutes with automatic scaling and Docker support.

The template’s AI layer supports multiple providers with fault-tolerant processing and zero-shot prompt handling. Testing shows a 40% reduction in boilerplate code for common SaaS features. However, the framework’s Erlang dependency creates a learning curve for teams unfamiliar with functional programming. Early adopters report significant time savings in initial setup but note ongoing maintenance requires customizations beyond the template’s scope.

Industry context and technical limitations

The DeepMind-EVE collaboration reflects a broader trend of AI labs seeking controlled yet dynamic environments for training. Game companies like Ubisoft and Epic Games have similarly partnered with AI researchers, though CCP’s rebrand to Fenris Creations marks a strategic pivot toward AI infrastructure. The technical challenge lies in maintaining game balance while testing experimental AI systems.

Current AI search integrations remain limited by platform-specific biases. Reddit’s community-driven content model creates inherent risks of misinformation propagation, even with quality filters. The Sheets extension’s open-source nature allows for community improvements but lacks enterprise-grade security certifications. Meanwhile, the Phoenix SaaS template’s modular design requires developers to handle dependencies between components, which can introduce integration complexities.

What to watch

Fenris Creations plans to release beta results from the EVE Online AI tests by Q4 2024. Google’s SGE team will measure Reddit integration performance against traditional search results in upcoming benchmarks. AISheeter developers aim to add multi-model ensemble capabilities by mid-2024. The Phoenix SaaS template’s creators are exploring Elixir-specific AI optimizations through their Tidewave MCP runtime analysis tool.

Updates

• 2026-05-12 — iRacing arrives on Vision Pro with ‘immersion and fidelity never before seen in sim racing’ (source)