LWiAI #246: one week, four fronts at once. Google I/O, agents, lawyers, safety

LWiAI Podcast #246 is a map, not an argument. The value of the episode is not one big market thesis but the fact that five simultaneous AI currents hit each other in the same week.

One week, four blocks: Google I/O, agents, lawyers and safety research

Last Week in AI published episode 246 on 26 May 2026, recorded on 22 May. The public episode description has four main blocks.

Google I/O brought Gemini 3.5, Gemini Spark as an always-on agent on Google Cloud with MCP tool support, Gemini Omni for generating video from images, audio and text, Antigravity 2.0, Gemini for Science and Genie for street simulation.

The coding agent block covers Cursor Composer 2.5, Grok Build from xAI and questions around talent churn and compute usage at xAI.

The business and legal block: Elon Musk lost part of his OpenAI lawsuit on statute-of-limitations grounds, reported OpenAI and Apple tensions, another large Anthropic funding round and a sharp Cerebras IPO move.

The research and safety block: an Erdős geometry result, negation neglect, interpretability, agent benchmarks, deepfake enforcement, autonomous hacking and provenance metadata.

Why this roundup is more useful than another leaderboard

The value of the episode is not a specific news item. It shows how the AI market is splitting into simultaneous layers. Models are improving, agents are moving into work tools, video is becoming a normal generative modality and companies are simultaneously dealing with lawsuits, capital, governance and safety questions.

For a technical team building a product, this reading is more useful than another benchmark. Tracking the best model is not enough. You need to know who controls the agent workflow, who offers multimodal tooling, who can carry enterprise contracts and where safety brakes are starting to form.

A podcast roundup is an entry point, not a primary source: every claim needs the original document

Every specific claim still requires the original announcement or court filing. Otherwise Radar becomes a rewrite of content that only sounds important.

Names like Spark, Omni, Build or Antigravity 2.0 are still branding. An agent tool becomes interesting only with clear permissions, audit and rollback. A video model becomes interesting only when it handles rights and provenance, not just a polished I/O keynote demo.

What to follow first: Google stack, IDE integration and legal documents

For Google: whether Gemini Spark and Antigravity 2.0 move from I/O branding into a testable work stack. For coding agents: IDE integration, performance on longer tasks and the ability to safely repair their own mistakes.

For OpenAI and Anthropic, follow legal and financial documents more than press releases. For safety topics: whether provenance, deepfake takedown and cyber evals move from demonstrations into standard deployment conditions.

Lilith's verdict

This link is not an article to inflate into a grand thesis. It is a radar map of the week: models up front, agents behind them, lawyers at the door and safety people with a hand on the brake.