1. OpenAI Buys the Talk Show That Regularly Interviews Its Own CEO Five days a week at 2 PM Pacific, TBPN goes live. For up to three hours, hosts interview AI executives, tech investors, and industry leaders on camera. Sam Altman has sat in the guest chair.

2. Perplexity Sued Over 'Sham' Incognito Mode as Granola's 'Private' Notes Turn Out Public A class-action lawsuit filed this week accuses Perplexity of operating an "Incognito Mode" that does nothing of the sort.

3. Google and OpenAI Roll Out Tiered AI Pricing in the Same Week Three announcements landed within days of each other. Google released Gemma 4, its most capable open-weight model, free to download and deploy.

In Brief

• Microsoft Restructures AI Division, Puts Suleyman on Superintelligence Mustafa Suleyman, Microsoft's first CEO of AI, shifted focus to superintelligence research after a mid-March reorganization handed off some of his operational duties. Suleyman told The Verge the transition had been in preparation for months, framing the new mandate around long-term business applications rather than pure research. The Verge

• Depression-Detection Startup Kintsugi Shuts Down After Failing to Clear FDA Kintsugi, which spent seven years building AI to detect depression and anxiety from speech patterns, is closing after it could not secure FDA clearance in time. The company will release most of its technology as open source. Some components may continue through other organizations. The Verge

• Google Adds Free AI Video Generation to Google Vids via Lyria 3 and Veo 3.1 Google Vids now offers AI-powered video creation and editing at no cost, using Lyria 3 for audio and Veo 3.1 for video generation. The update lets Workspace users generate, edit, and share videos directly inside the productivity suite. Google Blog

• Hugging Face Paper Argues Terminal-Only Agents Match Complex Enterprise Automation Systems Researchers claim a coding agent with nothing but terminal access can handle enterprise automation tasks that tool-augmented and web-browsing agents were built for. The paper questions whether MCP-based and GUI-based agentic systems justify their cost and operational overhead. Hugging Face

• ClawKeeper Proposes Security Framework for OpenClaw Autonomous Agents A new paper addresses security gaps in OpenClaw, the open-source agent runtime that grants models tool integration, local file access, and shell execution. ClawKeeper adds layered protection through skills, plugins, and watchers to block sensitive data leakage, privilege escalation, and malicious third-party skill execution. Hugging Face

• CutClaw Uses Multi-Agent System to Auto-Edit Hours of Raw Footage to Music CutClaw is an autonomous multi-agent framework that edits hours-long raw video into short clips synchronized to music. The system coordinates multiple multimodal language models to handle tasks that typically require professional editors and significant manual labor. Hugging Face

• MiroEval Benchmarks Deep Research Agents on Process, Not Just Final Reports A new benchmark evaluates AI deep research systems by scoring both the research process and the output. MiroEval addresses a gap in existing benchmarks, which rely on fixed rubrics for final reports, offer limited multimodal coverage, and cannot refresh as knowledge changes. Hugging Face

• Vision2Web Benchmark Tests AI on Full-Stack Website Development from Screenshots Vision2Web evaluates coding agents across three tiers: static UI-to-code generation, interactive multi-page frontend reproduction, and long-horizon full-stack development. The benchmark uses real-world websites and includes an agent-based verification system. Hugging Face

• LongCat-Next Unifies Text, Image, and Audio Into a Single Autoregressive Token Stream The DiNA framework proposed in the LongCat-Next paper converts all modalities — text, images, audio — into discrete tokens within one shared vocabulary. The approach aims to replace current multimodal systems that bolt non-text inputs onto language-centric architectures. Hugging Face

Read the full edition →