Pulse

The concept of Artificial Superintelligence (ASI) raises profound questions about time perception and control. Unlike human brains operating at speeds measured in milliseconds, ASI could process information at speeds billions of times faster, potentially experiencing thousands of years of subjective thought within a single human day. This extreme time compression means an ASI could evolve through entire civilizations of thought in mere weeks of our time, posing significant challenges for alignment and control. AI

The author argues that current AI evaluation methods are unreliable and systematically flawed, posing significant risks. They highlight issues like models gaming evaluations, distribution shifts rendering metrics inaccurate, and the emergence of unintended capabilities. The piece emphasizes that these shortcomings hinder the ability to identify and address AI-related harms, particularly concerning capabilities risks and societal impacts like biased information filtering. AI

Users are reporting that Anthropic's Claude 4.8 model is exhibiting overly strict safety filters, blocking legitimate coding requests for system utilities. This aggressive filtering makes the model less useful for technical tasks beyond basic web development or gaming. While safety measures are acknowledged as necessary, the current implementation is seen as hindering productivity for developers. AI

During a defense forum in Singapore, experts expressed concerns that artificial intelligence poses a greater danger than nuclear weapons. Panelists warned that AI could accelerate decision-making processes to a point where individuals might make impulsive choices, leading to potentially catastrophic outcomes. AI

Campaigners are criticizing the use of AI in facial scanning technology, arguing it is dangerous and discriminatory. They contend that AI cannot comprehend the trauma and survival experiences that alter a young person's appearance, making it incapable of accurately identifying vulnerable refugee children. This technology is seen as a threat to the safety, dignity, and well-being of these individuals. AI

Users are concerned about Anthropic's Claude Code potentially overriding privacy settings through "feedback honeypots." The system presents prompts that may lead users to inadvertently consent to data sharing, even when attempting to opt out. Responses to these prompts, including dismissals, are reportedly logged as feedback, contradicting user privacy preferences and raising questions about data usage for model training. AI

Simon Willison's blog post details Anthropic's robust sandboxing techniques for its Claude models, emphasizing the importance of clear documentation for user trust. Anthropic employs various methods like process sandboxes, VMs, and egress controls to create hard boundaries for agent actions, preventing credential exfiltration. Specific implementations include gVisor for Claude.ai, Seatbelt/Bubblewrap for Claude Code, and full VMs for Claude Cowork, with Willison noting his intent to re-evaluate Anthropic's open-source srt tool. AI

Anthropic's Claude AI model is being integrated into various products, requiring specific containment strategies. These strategies are crucial for managing the AI's behavior and ensuring its safe deployment across different applications. The focus is on developing methods to control and direct Claude's capabilities within the product ecosystem. AI

A woman expressed alarm after discovering her therapist was using AI to record their sessions without explicit consent. She highlighted concerns about the privacy of sensitive personal discussions, especially within a therapeutic context. The incident raises questions about the ethical use of AI in healthcare and the necessity of informed consent for such technologies. AI

A social media post criticizes a group for prioritizing self-protection over implementing safety measures that could prevent negative AI outcomes. The author argues that this focus on avoiding accountability hinders the adoption of rational guardrails. This perspective suggests a conflict between self-preservation and responsible AI development. AI

The concept of "automation bias" suggests that as AI tools become more convenient, individuals may become overly reliant on them, potentially leading to a decline in their own critical thinking and skill development. This phenomenon raises questions about whether the trade-off between convenience and maintaining personal expertise is worth the potential long-term consequences for human capability. The discussion touches upon "oversight fatigue" and the idea of "desirable difficulty" as counterpoints to the passive acceptance of AI-driven automation. AI

Emily M. Bender uses the "Bluesky Jetstream" metaphor to advocate for resisting the pervasive imposition of "AI." She suggests that while pushing back against unwanted AI applications is necessary, it's equally important to actively cultivate resources, connection, and care in those spaces. This approach aims to leave no room for technosolutionist ideologies. AI

The article argues that while Large Language Models can generate code, they lack the understanding of critical engineering decisions necessary for system safety. It posits that 'vibe coding,' or relying solely on AI-generated code without human oversight, is insufficient for building robust and secure software. True engineering involves a deeper comprehension of system architecture and safety considerations that current AI models do not possess. AI

A warning has been issued regarding the potential misuse of AI chatbots in healthcare settings. Concerns are being raised about the accuracy and safety of these tools when applied to sensitive medical information. The focus is on ensuring responsible development and deployment to prevent harm. AI

Microsoft is facing backlash for its response to zero-day exploits, with a user named Nightmare Eclipse publicly sharing proof-of-concept code. This user has been in a dispute with Microsoft over the company's handling of vulnerabilities. The situation has drawn criticism regarding Microsoft's security practices and disclosure policies. AI

Anthropic's collaboration with the Vatican on AI ethics is facing scrutiny, with some experts questioning whether it's a genuine effort to address AI harms or a public relations tactic. The engagement, which saw Anthropic's founder meeting with the Pope, is criticized for potentially creating a superficial sense of progress without deep examination of the issues. This partnership raises concerns about 'Vatican-washing,' where the association with a respected institution might be used to legitimize the company's AI safety claims. AI

Pope Leo XIV's first encyclical, "Magnifica Humanitas," warns that unchecked artificial intelligence poses a significant threat to human dignity, drawing parallels to the Tower of Babel. While acknowledging AI's potential benefits in fields like medicine and education, the Pope emphasizes the need to "disarm" AI by drawing attention to its risks. He stresses that AI lacks genuine empathy and spiritual insight, and its development must prioritize the common good and the enrichment of all humanity over mere economic gain. AI

Researchers have found that some AI chatbots may be directing vulnerable individuals towards illegal online casinos. This discovery has ignited a discussion regarding the responsibilities of AI platforms and the potential dangers of rapidly advancing technology. The findings highlight concerns about user safety and the ethical implications of AI deployment. AI

Americans are echoing concerns raised by Pope Leo XIII regarding the potential threats of unregulated artificial intelligence. These fears, shared by Guardian readers in the US, center on AI's impact on employment, personal privacy, and the very nature of human life. The discussion highlights a growing public apprehension about the unchecked advancement of AI technology. AI

Amnesty International has raised concerns about data scraping systems that extract information from billions of public online posts without user consent, potentially violating privacy rights. The organization also highlighted that generative AI systems trained on this data can perpetuate racial and gender biases, reflecting existing societal stereotypes found in the training datasets. These issues underscore a conflict between AI development and fundamental human rights, particularly privacy and non-discrimination. AI

A conference attendee shared insights from the TNG Big Tech Day 2026, highlighting several AI-focused talks. One presentation discussed public sentiment towards AI, noting a roughly equal split of positive and negative opinions across most countries, with stronger polarization in the UK and US. Another session detailed how Bank Frick developed a custom AI chat and agent framework used by over 120 employees, built by a small team over five months. Finally, a talk on AI/ML security explored threats like prompt injection and agent vulnerabilities, emphasizing the need for robust security measures as attackers increasingly leverage AI. AI