Episode #195: 21 October 2025

The episode tracks two arcs: DeepSeek-OCR’s optical context compression, and a sweeping AWS outage that exposed platform dependence. DeepSeek’s 3B MoE vision model compresses pages into ~100 vision tokens with 97 percent accuracy, runs at ~2500 tokens per second on A100s via vLLM, and hints at gains for STEM tasks and long-context workflows. Meanwhile, a DNS failure in AWS US-EAST-1 cascaded into DynamoDB, EC2, and Lambda issues, taking down Snapchat, Fortnite, Coinbase, Slack, and more, sparking memes, status-page humor, and fresh debate over crypto’s reliance on centralized clouds. X reported smooth sailing, Hetzner bragged about uptime, and engineers joked through the chaos while calls for multi-region design resurfaced. Also noted: an arXiv study on LLMs trained on viral junk content showing reasoning decay, Modi’s Diwali stop on INS Vikrant, Ricky Gervais’ 2020 monologue trending again, Bitcoin whipsaws, and a nod to Karpathy’s hands-on rigor.