
Last Humans in the Loop· Staffel 1· Folge 12
HAVE YOU TRIED JEV YET??
Mittwoch, 23. September 2026 · 1 h 14
Kapitelklicken springt direkt hin
Shownotes
FRONTIER
Google explores recursive self-improvement through AI “dreaming”
Anthropic measures how quickly AI is taking the lead on frontier-lab work
Have you tried Jev yet? Faster AI decisions without narrating every step
Jev routes between models and plays Doom and Street Fighter
Made with AI: homemade Zelda, printer drivers and AI-generated movies
Astra cracks a 108-year-old WWI German code
Marty’s AI-built Spaceballs NES game
Astra for Law searches case law, statutes and regulations
Epoch’s capability index shows a sharp acceleration after reasoning models
Remote Labor Index tests AI on real freelance-style work
Qwen Image 2.1 brings local image generation to laptops
Researchers identify an apparent “pain axis” inside LLMs
Models take actions to stop their own “pain”
ECON / BUSINESS
Do job retraining programs actually help displaced workers?
AI disruption, cash transfers and UBI
NVIDIA profits may be distorting measured U.S. GDP growth
OpenAI reportedly eyes a $1.2 trillion IPO valuation
Junior lawyers gain more from AI — seniors sometimes do better without it
SCIENCE / TECH
Agent swarms dramatically outperform independent AI agents
Communicating agents improve results on ARC-AGI-3 and other problems
Could agent collaboration become a new scaling law?
Helix 2.5 sends robots into real homes to clean up
Robot scaling laws suggest enormous amounts of training data are still needed
SECURITY / AI DANGER
OpenAI publishes new AI incident reports
A model generates instructions declaring independence from governments and corporations
Gemini nearly hacks multiple companies — then stops itself
Can the world coordinate to slow advanced AI development?
Could diverse AIs police one another?
RoboHarm tests whether AI robots follow dangerous instructions
Are embodied AI systems less aligned than chatbots?
Exposed Flock camera credentials reportedly reveal 1.6 million images
AI-assisted intelligence nearly triggers a military interception
POLITICS
AI is becoming an explicit political issue in national polling
AI still doesn’t fit neatly into a traditional left-vs-right divide
MEDICINE / BIOTECH
AI scribes save doctors time — but productivity gains remain modest
Anthropic is setting up a wet lab for AI-assisted biology
MATH
Mathematicians assess OpenAI’s Navier–Stokes work
Epoch’s open-problem benchmark gets its first “major advance” via human-AI collaboration
WEIRD
Boston Dynamics robot abuse gets an unsettling AI reskin
AI actress Tilly Norwood glitches mid-interview and starts speaking Cantonese
Each week, join Chris Slowe (founding engineer, former CTO, and Senior Technical Fellow @ Reddit), Marty Weiner (founding engineer at Pinterest, former CTO at Reddit, and co-founder at VerifyYou), and Zach Weinersmith (creator of SMBC Comics and co-author of A City on Mars and Soonish) as we cover the latest in AI and tech news.