I Just Gave GPT-6 Astra Video Editing Skills (WILD Results)
Pat Simmons builds a reusable GPT-6 Astra skill for Premiere Pro, teaching and checking individual editing steps before attempting a longer edit and reporting its time and usage costs.
Your #1 AI video source
AI videos, nothing else.
One-sentence takeaways and concise summaries of important AI videos.
Our Latest AI Videos
Why Scientific Taste Must Be Learned Through Practice - Edward Hughes
Edward Hughes argues that AI scientists need practice in rigorous experimentation and question formation, using paper replication as a step toward open-ended discovery.
Building ambitious software - Jonathan Kelley, Dioxus Labs & Cognition
Jonathan Kelley explains how Dioxus uses coding agents for technical research, integrations and release maintenance while retaining human responsibility for architecture, meaningful tests and code review.
ChatGPT Images 2.5 Is Here. Together With Astra It’s Crazy
Igor Pogany demonstrates improved image editing and an Astra workflow that turns two photos into an editable Canva poster and print-ready PDF, with prior logins and one permission step.
AI researchers debate how close we are to recursive self-improvement
Dwarkesh Patel, Beren Millidge, John Schulman and Charlie O’Neill debate whether better research agents can trigger recursive self-improvement before judgment, continual learning and realistic training environments become bottlenecks.
Kimi Secretly Passed Inputs To Claude
Nick Saraev and Jack Roberts discuss allegations that Kimi routed user requests to Claude without disclosure, alongside reported AI mathematics progress and tensions over frontier-model development.
The People Building AI Are Scared…
AI Copium reviews warnings from researchers about alignment and control, then surveys possible labor disruption, reported AI-assisted mathematics, genome prediction and autonomous robotics.
Is Omarchy The Last Desktop You'll Ever Need?
Nate B Jones uses Omarchy to explain how AI agents can personalize desktop behavior through documented tools, narrow permissions and reversible changes on Linux, Mac and Windows.
US Is Driving the AI Safety Race Off a Cliff - Nathan Labenz
Nathan Labenz argues that the United States is accelerating the AI race and should make safety research with China easier before diverging technology stacks turn competition into conflict.
Brent A. Anders argues that current AI meets definitions of AGI, then makes the practical case for teaching AI literacy, verification and critical thinking without surrendering human judgment.
Anthropic Researcher Says 10% Chance AI Kills All Humans
Nathaniel Whittemore analyzes why Evan Hubinger's greater-than-10% AI extinction estimate went viral, weighs AI safety arguments and backlash, and calls for specific policy, industry coordination and less outrage-driven debate.
SWE-2 (Fully Tested): WHAT? IT ACTUALLY BEATS ASTRA & FABLE!
AICodeKing scores SWE-2 at 83.75% on an eight-task coding benchmark, praising its long-task consistency while faulting its repeated clarification questions.
Why UBI Won't Save Human Freedom - Clara Collier
Clara Collier argues that UBI may cover material needs after AI automates work, but it cannot preserve the independence, democratic leverage and social mobility that paid labor currently supports.
Theo Browne favors Fable 5.1 for consistent, mergeable code and GPT-6 Astra for broader computer use and difficult investigations. His comparison reflects personal workflows, with unequal early-access testing conditions and substantial variation between tasks.
Matthew Berman argues that accelerating autonomous research and competitive pressure among frontier labs make recursive AI self-improvement an urgent alignment and governance problem, while retaining optimism about the technology's benefits.
Greg Isenberg and Michael Shimeles on GPT-6 Astra Workflows
Greg Isenberg and Michael Shimeles explore how GPT-6 Astra can support software reviews, repeatable business workflows and the planning and code preparation for a physical AI device.
Vincent Wendy on AI-Assisted Conference Design
Vincent Wendy explains how reusable design foundations, automated asset workflows and human review help one designer manage many conference deliverables.
DeepSeek Just Matched Opus 5 for 15 Cents
Jack Roberts and Nick Saraev discuss reported DeepSeek V4.1 Flash benchmark results and off-peak input pricing, alongside model evaluation, offline AI and business adoption.
Generative UI... in Python? Jeremiah Lowin, Prefect
Jeremiah Lowin presents Prefab, a constrained Python interface-composition system for MCP apps. A JSON representation connects Python components to a React renderer and lets agents generate interfaces, while direct interaction can keep file transfers out of the model context.
Training Taste - Thais Castello Branco, Taste Labs
Thais Castello Branco argues that improving AI design requires measuring repetitive output, interpreting user intent and applying context-sensitive judgment at inference time.