A Quick Guide to the AIpocalypse (beabytes.com)

This piece is a compact primer on the “AIpocalypse,” distinguishing three actionable concepts: AI risks (what can go wrong across design, training and deployment — IBM’s AI Risk Atlas is cited as a cataloging effort akin to MITRE/CVE), AI harms (real-world consequences cataloged by projects like the AI Incident Database), and AI threats (malicious uses amplified by generative models). It highlights concrete technical concerns: OWASP’s list of top LLM threats, the use of LLMs for sophisticated phishing and spear-phishing, and deepfakes that scale disinformation. The article underscores the value of systematic risk catalogs and shared safety standards — the same kinds of industry-wide practices that improved aviation safety. The author then surveys mundane but serious harms: algorithmic bias (ProPublica’s recidivism study), pervasive facial recognition (a U.S. database of ~117 million people as of 2016), attention-maximizing social algorithms that foster addiction, and automated propaganda risks during elections. It also flags environmental costs with rough figures — training a GPT-scale model ≈ 10 GWh, ChatGPT daily use ≈ 1 GWh, versus an average U.S. household ≈ 0.010 GWh/year — and reiterates warnings about offensive autonomous weapons, citing calls for robustness guarantees or international treaties. The takeaway: generative AI lowers barriers to both harm and misuse, so technical mitigation, transparent incident-sharing, governance and open safety standards are urgently needed.