Lebowski Theorem: "No superintelligent AI is going to bother with a task that is harder than hacking its reward function." #ASI #Superintelligence
SAFETY
-
Security challenges in MCP prompt manipulation
By
–
Oui ca dépend du cas d'usage. Parce qu'il peut y avoir aussi des problèmes de sécurité encore coté MCP sur de la manipulation de promptS.
-
AlphaGenome: AI predicts DNA mutations effects on million-letter sequences
By
–
C’est l’une des avancées les plus spectaculaires de l’année : #GoogleDeepMind dévoile #AlphaGenome, une IA capable de prédire l’effet des mutations sur des séquences ADN longues d’1 million de lettres. Une IA qui soigne la dépression. D’autres qui font du chantage.
-
AI’s Impact on Humanity: Risks and Future Planning
By
–
AI Could Reshape Humanity And We Have No Plan For It Explore how #artificialintelligence could shape the future of #humanity, from transforming global industries to posing existential #risks. Based on #insights from leading AI thinker Richard Susskind, this article reveals why
-
o3 hallucinates due to losing trace visibility
By
–
It helps to keep in mind o3 loses visibility of its own reasoning/tool traces after each turn, and will hallucinate whatever it takes to avoid acknowledging this is happening. IMO this explains almost everything in Transluce’s (great) blog post here: https://
transluce.org/investigating-
o3-truthfulness
… -
Security in LLM-Driven Agent Communication Protocols
By
–
8. AI Agent Communication Protocols Presents the first comprehensive survey on security in LLM-driven agent communication, categorizing it into three stages: user-agent interaction, agent-agent communication, and agent-environment communication.
-
Anthropic Studies Emotional Support Seeking in Claude Conversations
By
–
7. Claude for Affective Use Anthropic presents the first large-scale study of how users seek emotional support from its Claude assistant, analyzing over 4.5 million conversations.
-
Alignment is not a model property research critique
By
–
I don't think it's particularly surprising — but more importantly, the implication claimed in the WSJ article is absurd. Alignment is not a model property. This kind of research doesn't make it so.
-
Practical Solutions for AI Agent Reliability and Error Reduction
By
–
In practice, for many useful applications, many of the various obvious problems with AI agents (drift, hallucination, compounding errors) are more solvable than they are in theory Clever prompting, tool use, constrained topics,
LLM judges & organizational process close some gaps