Project Synapse Weekend: When Cheap Local Models Hallucinate + Anthropic's Fable-5 Controversy

Jim Love and Marcel host the Project Synapse weekend edition without Jon Pinard, discussing Jim's three-month experiment building his Nexus AI agent. Trying to avoid expensive API calls, Jim switched to a small open-source Gemma model and found that limited context windows and growing memory files led to severe hallucinations and corrupted "memories," suggesting he may need periodic context compression or paid access to larger models. They explain quantization and compare AI memory to human memory reconstruction, then pivot to Anthropic's newly released Fable-5 (a guardrailed Mythos-5 variant) via Project Glasswing: powerful but slow, expensive, temporary for subscribers, and controversial for silently downgrading or misdirecting users. They also discuss AI-written apps replacing utilities, debates over AI creativity, Canadian proposals on youth social media and humanlike AI, information control by big tech, job disruption risks, and a case where lawyers cited AI-fabricated precedents in court.

00:00 Weekend Edition Kickoff 00:27 Cheap Model Reality Check 03:37 Quantization and Brain Metaphors 08:16 Nexus Agent Memory Spiral 13:26 Compressing Context Strategy 17:17 Conversational Programming Demo 23:52 What Is Fable 5 26:52 Fable Controversy and Guardrails 30:30 AI Feels Human and New Rules 34:21 LLMs Mirror You 35:29 Kids and Age Gates 36:24 Dumbed Down Society 39:07 Extraneural Fan Site Backlash 43:33 AI Art and Creativity Lines 48:17 Waterloo and Mennonite Tech 53:05 Weaponized Narratives 58:30 Billionaires and Ubiquity 01:01:19 Jobs and AI Economics 01:04:15 Lawyers Citing Fake Cases