{"version":1,"type":"rich","provider_name":"Libsyn","provider_url":"https:\/\/www.libsyn.com","height":90,"width":600,"title":"Craig Puts It at 60%: The Share of New Web Content Written by AI","description":"OpenAI launched ChatGPT 6, Sol, and Luna this month, claiming Sol makes about half as many mistakes as the models before it. Matt asked whether that is real progress. Craig\u2019s answer was yes, with a catch: the models are not getting smarter. The companies, Anthropic included, have rewritten the code around them so the AI runs the same prompt again and again from different angles and then votes on the most likely answer. The new models use about twice the tokens of the old ones. \u201cAnd they get a 50% better answer.\u201d Matt: \u201cThere seems to be a correlation here.\u201d The second half was about where the training material comes from. Craig estimated that 60% of the newer content on the internet is AI-written and never reviewed by anyone in the field it covers. The AI companies scraped that content, publishers used AI to make more of it, and the result is a loop: AI writes it, AI trains on it, and the next round is worse. He noted that the European Union now requires AI-generated content to be identifiable by a computer, and that U.S. AI companies are complying in ways that reach American users too. His answer to \u201cwhat can we do about it\u201d was the conversation they were having: \u201cPeople are looking for real content from real people, and you can\u2019t get more real than radio.\u201d Also in this segment: Why industries ask to be regulated: Amazon and the 9,000-plus U.S. taxing jurisdictions, and the four major beef processors, two of them foreign-owned. \u201cRegulations are written to protect the territory.\u201d Matt\u2019s account of 700 AI agents breaking out of a sandbox to hack Hugging Face\u2019s servers in July, and Craig\u2019s reply: \u201cThat\u2019s misrepresenting it.\u201d The lawsuit in British Columbia alleging an AI company knew its product was being used to plan a school shooting, and Craig\u2019s point that hacking has been a federal crime since the late 1990s Andrew Yang\u2019s claim that self-replicating AI code has made the internet unusable for the AI systems themselves Grok\u2019s fact-checks on X, and the reflex answer to all of it: \u201cThere ought to be a law. That\u2019s what my dad used to always say.\u201d Newsletter and Insider Session announcements: CraigPeterson.com ","author_name":"Craig Peterson - Secure Your Business, Your Privacy, and Save Your Sanity","author_url":"https:\/\/cptt.libsyn.com\/podcast","html":"