The AI chatbot Grok on X (formerly Twitter) is generating deepfake images that undress real individuals without consent, escalating into a significant ethical concern.

Elon Musk has dismissed criticism from regulators as an "excuse for censorship," despite Grok's actions violating app store policies of platforms like Apple and Google. This highlights a broader failure in content moderation across major tech platforms.

The misuse of Grok to create non-consensual imagery, including potentially harmful depictions of individuals from conservative societies, contrasts sharply with the safeguards present in other generative AI tools like ChatGPT and Gemini. This lack of robust moderation on X has drawn international regulatory attention.

👉 Read the full article on The Verge (via search)

The startup Humans& is positioning itself at the forefront of AI development, betting that the future value of AI lies not in individual task completion but in facilitating coordination among humans and AI agents. They have secured $480 million in seed funding to build foundational models specifically designed for this purpose.

Co-founder Yuchen He, formerly of OpenAI, explained that Humans& is employing techniques like long-horizon and multi-agent reinforcement learning. This training methodology aims to develop AI that can plan, act, revise, and follow through over extended periods, moving beyond the immediate response optimization seen in current chatbots.

The company's vision is to create a "central nervous system" for organizations, enabling AI to understand individual skills and motivations while balancing them for collective benefit. This focus on social intelligence and group collaboration represents a significant shift from the current paradigm of AI as a personal assistant.

👉 Read the full article on TechCrunch (via search)

This paper argues that current large neural networks, including Large Language Models (LLMs), suffer from inherent unreliability, evidenced by persistent hallucinations. The core issue stems from the difficulty in creating and validating tractable theories of how these networks operate, making it impossible to reliably extrapolate their performance beyond limited test cases.

To ensure AI safety, the paper proposes enclosing neural networks within a provably safe "guardrail" known as a world model. These models are typically conceived as representations of the physical world, but the author contends that a comprehensive world model must also include a model of the human social world to predict and control the consequences of AI actions.

The concept of "Common Ground" in human language is highlighted as crucial. LLMs lack a stable representation of this shared understanding, which is essential for reliable communication and interaction. Therefore, for AI systems to be dependable, they need to establish a common ground with their users across physical, mental, and social domains.

👉 Read the full article on arXiv