OpenAI’s new LifeSciBench is 750 tasks written by 173 PhD scientists across biotech and pharma. The top model passed just 36.1%. Buy “AI scientist” pitches against that number, not the demo. Read our take.

In a sworn June 16 filing, the Pentagon’s AI chief said Grok Gov helped deploy over 2,000 munitions to 2,000 targets in 96 hours during Operation Epic Fury. The most explicit on-record admission that a commercial model has directed live munitions. The Hill.

Qwen-RobotNav, Qwen-RobotManip, and Qwen-RobotWorld cover navigation, arm control, and world modeling, now in pilots with Alibaba Cloud customers in manufacturing and logistics. The China robotics stack is consolidating around Qwen fast. Read our take.

Deployment Simulation strips the original reply from de-identified production conversations, re-runs them on a candidate model, and checks for known failure modes, with a median 1.5x multiplicative error. Pre-release eval is moving from benchmarks to traffic. Read our take.

As of June 18, Pro, Ultra, and free individual accounts no longer get Gemini CLI requests served. The replacement is multi-agent with a new plugin system; enterprise Code Assist licenses are untouched. If you script Gemini, your install command changed: npm install -g @google/antigravity-cli. Google’s notice.

Scale at $43/mo unlocks the Ad Network, but the subscriber math says it earns its keep around 5,000-plus engaged readers, not before. We did the breakeven so you do not have to. Read it · Try beehiiv.

Kit’s free Newsletter tier carries you to 10,000 subscribers. Creator at $33/mo is the right move only when the automation wall, not the subscriber cap, is what slows you down. Read it · Try Kit.

Fifteen skills ship; ops teams should install competitive-intelligence, PDF-extraction, and API-change-monitor, and leave browser-automation for later. Read it · Try Firecrawl.

CustomGPT now trains on closed Zendesk tickets. Pick it when your knowledge is buried there; choose Zendesk native AI to add zero new vendors, or Chatbase for the budget play. Read it · Try CustomGPT.

Kit MCP went GA with 76 read-write endpoints on the Creator plan ($33/mo). Our setup guide runs the connection start to finish for both Claude and ChatGPT. Read it · Try Kit.