The US government lifts Claude Mythos 5 export restrictions; OpenAI previews GPT-5.6 Sol; 6,000 prompt injection attempts all fail against a Claude email assistant; and AI inference margins hold near 70–80%. Ben Thompson on memory chips, China, and Microsoft’s AI model play; Gergely Orosz on six months of AI agents remaking engineering; and a placebo that works even when the patient knows it’s fake.

Ben Thompson · Jun 25

Thompson and Figma CEO Dylan Field discuss AI as a design tailwind, new product launches, and competitive positioning after the failed Adobe acquisition.

Ben Thompson · Jun 23

Thompson examines memory chip makers’ regrets over Chinese market access and Microsoft’s strategic case for incorporating Chinese AI models.

Sean Goedecke · Jun 26

Goedecke calculates AI inference margins at 70–80%, disputing narratives that inference services are economically unsustainable long-term.

Gergely Orosz · Jun 23

Orosz surveys six months of AI-driven workflow changes at major tech companies, including reduced code review and shifting team dynamics.

Simon Willison · Jun 26

Willison covers a public security challenge where over 6,000 prompt injection attempts failed against a Claude Opus email assistant.

CNBC · Jun 26

US government cleared Anthropic to distribute Claude Mythos 5 to over 100 institutions after export controls had restricted broader access.

OpenAI · Jun 26

OpenAI previews the GPT-5.6 series—Sol, Terra, and Luna tiers—in a limited rollout for select institutional partners.

Andrew Nesbitt · Jun 26

Nesbitt’s satirical incident report dramatizes supply-chain attack risks from AI-assisted code review, including instruction injection and correlated blind spots.

Sean Goedecke · Jun 27

Goedecke argues that stating obvious truths grounds technical communication, validates tacit beliefs, and creates foundations for more nuanced workplace ideas.