I Read Goodfire's New Essay. AI May Be Far Easier To Understand Than We Thought.

AI safety may have a real reason for optimism. Goodfire's new research suggests the inside of an AI model is much more understandable than researchers assumed, and if that holds up, AI interpretability could change what we can trust AI to do. Full post: https://natesnewsletter.substack.com/p/mechanistic-interpretability-explained?utm_source=youtube&utm_medium=video&utm_campaign=mechanistic-interpretability-explained-launch&utm_content=description Nate's Library MCP (every prompt kit and guide, inside the AI you already use): https://unlock-ai.natebjones.com/guides/how-to-connect-nates-library?utm_source=youtube&utm_medium=video&utm_campaign=mechanistic-interpretability-explained-launch&utm_content=description My Links 🔗 👉🏻 Newsletter: https://natesnewsletter.substack.com/?utm_source=youtube&utm_medium=video&utm_campaign=mechanistic-interpretability-explained-launch&utm_content=description 👉🏻 X: https://x.com/natebjones 👉🏻 TikTok: https://www.tiktok.com/@nate.b.jones 👉🏻 Instagram: https://www.instagram.com/nate.b.jones What's really happening inside the AI models OpenAI and Anthropic are building? The common story is that AI gets more dangerous as it gets more powerful, but the real question is whether we can understand it well enough to correct it. In this video, I share the inside scoop on why Goodfire's essay is a green flag for AI safety: - What Goodfire found inside vision and language model neurons - How reading model internals could catch reward hacking during training - Why explainable AI tools may reach law, health, and finance first - Where you may soon correct what an algorithm learned about you The research still has to scale into real products, and if it does, we get much better tools for finding and fixing what goes wrong inside AI. Chapters: 00:00 Why Goodfire's research is good news for AI safety 01:26 The 99 to 100% finding in a vision model's final layer 02:01 What Goodfire is and four changes to watch 02:44 The black box problem I saw at Amazon 04:42 What interpretability is and what an AI neuron is 06:01 The cats and cars neuron, reread 07:26 A preliminary 90% of neuron poles in a language model 08:01 Forecast 1: catching reward hacking and misalignment 11:55 Forecast 2: building more capable models safely 13:28 Forecast 3: explainable AI for law, health, and finance 15:58 Forecast 4: correcting what an algorithm learned about you 17:52 Don't just buy the pessimism Listen to this video as a podcast. Spotify: https://open.spotify.com/show/0gkFdjd1wptEKJKLu9LbZ4 Apple Podcasts: https://podcasts.apple.com/us/podcast/ai-news-strategy-daily-with-nate-b-jones/id1877109372