Latent Space AI Anthropic's Claude 'Dreaming': What to Know
May 7, 2026
They dig into Claude's new 'dreaming' ability for offline self-improvement. Big cloud deals with Google and SpaceX and what that means for compute capacity. A major security exposé about apps leaking medical and corporate data is discussed. Google shelving Project Mariner and how platforms respond to safety gaps are also covered.
AI Snips
Chapters
Transcript
Episode notes
Browser Agents Folded Into Gemini Due To Cost
- Google shut down Project Mariner and folded its browser-agent tech into Gemini, likely due to high compute costs and slow performance.
- Jaeden notes computer-use agents that take screenshots are ~85x more compute intensive than text/code-based methods.
Claude Edits Files Directly Instead Of Clicking Through Apps
- Jaeden describes Claude Cowork editing files by modifying the underlying file rather than driving the UI, which is more efficient.
- He recounts better results when Claude edits CapCut files directly rather than controlling the CapCut app via screenshots.
Make AI Generated Projects Private Until Secured
- Avoid assuming AI-built apps are secure; many legacy projects defaulted to public and leaked sensitive data until owners fixed settings.
- If you maintain a Replit/Base44 project, set it private and verify authentication before exposing any customer or medical data.