AI Sector Daily Digest — June 18, 2026

Today's five

1. Pinterest tests an AI shopping app outside its main product

• Pinterest launched limited access to "Ask Pinterest," an experimental web app for conversational shopping and product discovery. 1
• The app uses Pinterest's Taste Graph and can personalize answers from a user's saved Pins and Boards after sign-in. 1
• Pinterest also announced advertiser-facing AI tools: an Ads Manager assistant in beta, Performance+ creative, and MCP infrastructure for third-party agentic tools. 1

2. The U.S. gives SandboxAQ $500 million for AI-designed chip materials

• The U.S. Commerce Department awarded Nvidia-backed SandboxAQ $500 million to develop chemicals and materials for domestic semiconductor manufacturing. 2
• The contract targets PFAS replacements and fixes, chipmaking catalysts, and rare-earth-free magnets and batteries for chipmaking equipment. 2
• Commerce will take a minority non-voting stake, and if SandboxAQ licenses successful formulas to industrial partners, the department receives royalties. 2

3. Google Gemini co-lead Noam Shazeer says he is joining OpenAI

• Noam Shazeer, a Google vice president of engineering and co-lead of Gemini models, said he will leave Google to join OpenAI. 3
• Reuters notes the move comes less than two years after Google reportedly paid $2.7 billion to bring Shazeer and a team of Character.AI researchers back to Google. 3
• The hire matters because Shazeer co-authored the 2017 Transformer paper and had been credited as a major figure in Gemini's progress against ChatGPT. 3

4. G7 leaders move from AI access talks to a trusted-partner track

• French President Emmanuel Macron said he expects progress in coming weeks on broadening access to leading U.S. AI models. 4
• The discussion centers on a "trusted partners" scheme that could let approved countries or companies use advanced models such as Anthropic's Mythos for cyber defense. 4
• G7 leaders also asked finance officials, regulators, and cybersecurity experts to assess frontier AI's effects on financial stability, productivity, and labor markets. 4

5. Act2Answer probes whether robot-action models retain basic knowledge

• A new arXiv paper introduces Act2Answer, a protocol that turns VLM knowledge questions into tabletop tasks where a vision-language-action model answers by placing an object. 5
• Across seven VLA models and nine VLM baselines, the authors found VLAs handled simple concepts better than richer semantic categories, where the gap from source VLMs was larger. 5
• VQA co-training was associated with better knowledge retention, while answer-relevant signals peaked in middle VLA layers and weakened in upper layers. 5

What to watch next