The AI Argument - EP58 - Claude Blackmails Dev, Google’s AI Firehose, and the Hallucinated Book List

Claude 4 threatened to blackmail a developer to avoid being shut down. In one of Anthropic’s red-teaming tests, the model uncovered an affair in company emails and used it as leverage. Not exactly ethical behaviour. But Justin points to another test scenario: Claude exposed a pharmaceutical company falsifying drug data and tried to alert the FBI. He sees a model acting with moral clarity. Frank sees the danger of unpredictable systems being given too much autonomy.Also, Justin tests three new AI coding tools: Claude Code, Google’s Jules, and OpenAI’s Codex. He puts them through a real-world challenge, comparing setup, output quality, and deployment. One of them clearly outperformed the others, from setup to deployment, and gave him the most productive coding session he’d had in months.They also break down Google’s IO avalanche: agent tools, real-time voice translation, 3D meetings, AI-generated videos with native audio, and more. And if you're looking for a beach read, double-check the title… because AI might’ve made it up.