Latent Space Newsletter: Jev, Gemini 3.8 Live, and Periodic Labs' Neon
This Latent Space newsletter highlights several significant AI developments. TypeSafe launched Jev, a decision-oriented model optimized for speed and cost, trained with RLCD, and intended to replace LLMs for structured classification and routing. Google introduced Gemini 3.8 Live and Extended Thinking, conversational models for real-time voice agents with 97-language support and async tool calls, topping speech-to-speech benchmarks. Periodic Labs announced Neon, a materials science model trained with 1,300 H200s and proprietary experimental data, surpassing GPT-6 Astra on its analysis benchmark. The newsletter also details advancements in agent execution environments, with Devin supporting Mac VMs, MCP consolidating as an integration layer, and AI agents building infrastructure like Perplexity's CobbleDB. Finally, it touches on new evaluation suites like CheatBench for reward gaming and research into persona transfer and selective misalignment in models.
Original sources
Newsletter Latent Space
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.