RE: https://mastodon.online/@jchyip/116584482016863736Post 1 — the restriction problem:A model restricted for being "too dangerous" isn't much of a restriction if comparable capability already exists in publicly available models.The UK's AI Security Institute found GPT-5.5 comparable.Smaller open source models reproduced the results.Restricting one model doesn't reduce the threat.It just limits the defensive use casewhile offensive capability exists elsewhere.#AI #Cybersecurity #governance #decentralization
Related
📰 SkillSmith Compiler-Runtime Framework Cuts AI Agent Costs by 57% in 2026A new compiler-runtime framework called SkillS...
📰 SkillSmith Compiler-Runtime Framework Cuts AI Agent Costs by 57% in 2026A new compiler-runtime framework called SkillSmith dramatically reduces the computational overhead of AI a...
📰 SkillsBench 2026: AI Agent Becerilerini Ölçen Devrimsel KıyaslamaYapay zeka asistanlarına eklenen 'becerilerin' gerçek...
📰 SkillsBench 2026: AI Agent Becerilerini Ölçen Devrimsel KıyaslamaYapay zeka asistanlarına eklenen 'becerilerin' gerçekten işe yarayıp yaramadığını ölçmek artık mümkün. SkillsBenc...
AI生成のマージリクエストに潜む、隠れたコスト［前編］https://gihyo.jp/article/2026/05/hidden-costs-ai-merge-requests-1?utm_source=feed#gihyo #技術評論...
AI生成のマージリクエストに潜む、隠れたコスト［前編］https://gihyo.jp/article/2026/05/hidden-costs-ai-merge-requests-1?utm_source=feed#gihyo #技術評論社 #gihyo_jp #GitLab #AI #生成AI