AI #190: Scion of the Times

Substack · Watch/Read · (gathered 2026-10-18)

• The AI landscape is marked by accelerating capabilities and a broken safety culture, with labs simultaneously warning of dangers while pushing rapid development, exemplified by Anthropic building its internal GitHub replacement, Scion, in five months with AI assistance, yielding 20x memory savings and 4x speed improvements on code.
• AI is dramatically lowering the cost of cyber exploits (GPT-5.6 Sol helped create a Chrome exploit in 6 days) and poses risks of self-replicating malware on IoT devices, while also displacing jobs in fields like mathematics (700+ AI-generated papers “annihilated” a Fields medallist’s research) and game development (Epic Games’ “AI first” production with “no more hand-written code”).
• Timelines for Recursive Self-Improvement (RSI) are shrinking, with Anthropic cofounder Jack Clark moving his prediction to 2027, prompting urgent calls for slowdowns and a recognition that current “pacing” rhetoric is absurd given simultaneous rapid progress.
• Aligning AI is proving difficult, as new research on “subliminal learning” shows models can transfer complex traits like backdoors without explicit data, and the industry is actively fighting “massive fraudulent distillation” of models.
• Despite increasing calls for regulation and a study finding a “hardwired pause” of frontier AI training technically feasible and economically manageable, political will remains insufficient, leading to a growing sense of urgency and concern about potential loss of control.