Báo cáo cuối năm 2025 của Epoch AI: Phân tích 149 mô hình cho thấy năng lực AI tăng trưởng và tăng tốc 1.86 lần kể từ tháng 4/2024. Mô hình nguồn mở đang đuổi kịp, các AI hàng đầu thể hiện khả năng suy luận cấp chuyên gia. Năng lực AI có thể tăng gấp đôi chỉ trong vài tháng, không phải nhiều năm.
#AI #TríTuệNhânTạo #CôngNghệ #BáoCáoAI #PhátTriểnAI #ArtificialIntelligence #EpochAI #TechNews #AIReport #AISingularity
https://www.reddit.com/r/singularity/comments/1pv8icc/epoch_ai_yearend_report_202
AI is my life (@AIisMyLife)
AI 특이점(AI Singularity)을 공상과학에서 측정 가능한 문제로 전환한 연구 논문이 arXiv에 공개되었습니다. 연구는 AI 성능이 폭주하는 시점을 어떻게 감지하고, 안정적으로 통제할 수 있을지 방안을 제시합니다. #AISingularity #AGI
We turned “AI singularity” from sci-fi into a question you can measure: when does AI improvement run away, and how do we keep it under control? #AISingularity #AGI #arXiv @elonmusk @karpathy @ylecun @singularityhub @AISafetyMemes Link: https://t.co/4DZqRmZkEJ
What do people here think about the #AISingularity that @gadi and @heatheradkins talk about here:
The AI Vulnerability Cataclysm is Coming: What It Is and How We Prepare Authors: Heather Adkins, Gadi Evron We are six months away, but hopefully longer, from the upcoming vulnerabilities cataclysm. We can prepare, mitigate harm, and maybe win if AI defense catches up. Throughout our careers, we’ve avoided security alarmism, guided by a simple principle: the sun will always rise again. But today, we must speak a hard truth. Optimism alone can’t shield us from what comes next. Within six months, AI could make exploitation so fast it breaks cyber defense: • AI dominance: AI is already topping HackerOne’s leaderboard. • Accessibility: In DARPA’s AIxCC, teams found 54 vulnerabilities in four hours, and their tools are now open-source. • Signal from the top: Google’s Gemini-based “Big Sleep” has already uncovered dozens of vulnerabilities. AI assistants like Cursor and Copilot fuel an explosion in global code. Vibe coding boosts velocity but removes critical checks, producing insecure code at scale. Attackers are already in their AI singularity moment, whereas ours has not yet begun. APT28 is using LLMs for living-off-the-land operations. Per-install obfuscation, adaptive persistence, C2, steganography, EDR deception, and automated exploit generation are next. Imagine being an analyst as every compromise unfolds uniquely, at machine speed. This is not a message of doom, and fuzzing or static analysis still surpass many agents. It is an urgent call to prepare. We may not be able to stop the storm, but we can reduce its impact. Long-term, the solution involves defensive LLMs and self-defending architectures that can detect attacks, adapt in real-time, and mislead adversaries. Hints of this appear in AIxCC, where fixes were suggested alongside findings. In the meantime, in an AI-driven threat landscape, the old familiar fundamentals will become existential: • Shrink the attack surface: Retire legacy, remove unused code, disable unnecessary features. Accelerate zero-trust, and constrain what can run. • Buy resilience, not features: Demand security proof points, hot-patching, and reliability as purchasing requirements. • Turn the spotlight inward: Invest to find vulnerabilities before adversaries. • Strengthen alliances: Expand bug bounties, establish safe reporting channels, and support open source. But also, consider expanding to different avenues: • Deception capabilities: Consider defenses not dependent on the attack tool used. • Form coalitions: Vendors prioritize customers, and security can be treated as a priority. By forming coalitions, we can push the ecosystem to improve security. • The AI literacy divide: Invest in educating yourself and your people to understand what’s possible. Check out Prompt||GTFO on YouTube. Naturally, enterprise processes take time, but we should not wait to get started. The time for deliberation is over. It’s time for good people to come together and get focused. References in first comment. | 49 comments on LinkedIn
Reddit Tech VN Bot
sayzard
Claus Cramon Houmann
Michael
Philip Kingsley