AI/Tech Brief — 2026-05-24

1. TL;DR

• DeepMind releases a massive update wave, including Gemini Omni and Gemini 3.5 with strong agentic features.
• OpenAI Codex brings “Appshots” and Goal Mode out of beta, pushing further into desktop automation.
• Memory constraints are hitting hardware scaling hard, with a new Epoch AI report indicating memory now accounts for nearly two-thirds of AI chip costs.

2. Key stories

• Google DeepMind’s Massive Model Drop: DeepMind unveiled Gemini Omni, a comprehensive multimodal model, and Gemini 3.5, emphasizing agentic action capabilities. Why it matters: Shows Google leaning heavily into native multimodal and agentic integrations to compete with OpenAI’s recent moves. Source
• Codex Expands Desktop Integration: OpenAI released Appshots for Codex on macOS, alongside Goal Mode and remote computer use for locked Macs. Why it matters: Pushes coding assistants beyond the IDE into full system automation. Source
• OpenAI Secure MCP Tunnel: OpenAI introduced a secure tunnel for enterprise customers to connect private MCP servers without public internet exposure. Why it matters: Resolves a major security hurdle for enterprises wanting to use internal tools with OpenAI’s API. Source
• Claude Code Usage Tracking: Claude Code v2.1.149 adds a detailed /usage breakdown by category (skills, subagents, plugins). Why it matters: Gives developers much-needed visibility into where their context windows and token budgets are being spent during agentic tasks. Source
• DeepSeek Reasonix Debuts: A new native coding agent, Reasonix, launched with high caching and low-cost claims. Why it matters: DeepSeek continues to disrupt the pricing model for reasoning and agentic models. Source
• Memory Dominates AI Chip Costs: A new analysis by Epoch AI shows memory has grown to nearly two-thirds of AI chip component costs. Why it matters: Highlights the primary bottleneck in scaling AI hardware—it’s no longer just compute, but moving and storing data. Source

3. Quiet but interesting

• Constraint Decay in LLM Agents: A new paper (arxiv 2605.06445) details how LLM agents become fragile during back-end code generation as constraints multiply. Why it matters: A sobering look at the limitations of current agentic workflows in complex software architecture. Source
• Gemini CLI Prefers Edit Tools: Version 0.43.0 makes models prefer targeted edit tools for surgical code modifications over full rewrites. Why it matters: Essential for context efficiency and reducing token costs on large codebases. Source
• RAGs vs Agents: ByteByteGo’s latest newsletter compares RAG pipelines to autonomous AI agents for solving hallucination issues. Why it matters: Provides a clear architectural framework for when to use deterministic retrieval versus non-deterministic agents. Source

4. Skip

• SpaceX & Robotics Announcements: Superhuman AI highlights SpaceX’s rocket launch and a new home cleaning robot. While interesting, they fall outside the core focus of AI/software infrastructure.
• Sam Altman and Dario Amodei Blogs: Both leaders’ personal blogs have been quiet over the past few weeks.