Anthropic's 'J-space' Discovery, Tencent's Hy3 Release, and New Regulatory Moves
Anthropic researchers uncover a hidden 'J-space' in Claude, Tencent releases the open-source Hy3 model, and new regulatory actions emerge regarding AI companions and safety laws.
Podcast В· 1 min
Anthropic discovers 'J-space' in Claude
Anthropic researchers have identified an internal workspace within Claude, termed 'J-space,' where the model holds and processes concepts before generating responses. This structure, which emerged during training rather than being explicitly programmed, appears to function as a 'global workspace' similar to neuroscientific models of human conscious access. Experiments showed that manipulating these internal concepts directly altered Claude's outputs, suggesting that J-space is where complex reasoning occurs. This finding is significant for the field of mechanistic interpretability. While researchers caution that this does not prove consciousness, it provides a new way to inspect the 'thoughts' a model uses but does not explicitly state. This could eventually lead to better safety systems that can detect hidden flags like prompt injection or manipulation before they manifest in a response. This discovery has sparked discussion among AI researchers about the potential for more transparent and inspectable AI behavior. By understanding how models structurally organize their internal reasoning, labs may be able to build more reliable safety checks that do not rely solely on the model's own output.
Tencent open-sources Hy3 model
Tencent has officially released its Hy3 model under a permissive Apache 2.0 license, moving it out of its April preview phase. The model is designed for high efficiency, using only a small subset of its parameters for each request, which allows it to run on significantly less hardware than competitors like Zhipu’s GLM-5.2. While Tencent's internal testing indicates that Hy3 outperforms open-source rivals in web research and tool use, it still trails in coding benchmarks like SWE-bench. The release is notable for its broad accessibility, as the Apache 2.0 license avoids the regional restrictions that have limited the availability of previous Chinese models in markets like the EU, UK, and South Korea. This release highlights the ongoing trend of Chinese AI labs prioritizing efficiency and open-source accessibility to compete globally. While Hy3 may not challenge the absolute frontier models, its efficiency gains and open licensing make it a compelling option for developers looking for reliable, cost-effective agent deployment.
Kyutai releases MIRA world model
Kyutai and General Intuition have released MIRA, an open-source world model capable of simulating a live 2v2 Rocket League match. The model was trained on 10,000 hours of AI-vs-AI gameplay footage and operates without a traditional physics engine or graphics code, rendering details like boost meters and crashes entirely within a neural network. Despite running at 20 frames per second on a single Nvidia GPU, the model has limitations, such as a short memory span of about four seconds, which can lead to hallucinations during goal replays. The teams have open-sourced the code, training data, and a playable demo to encourage further research. This development is part of a broader effort to simulate the physical world for future robotics training. By demonstrating that neural networks can learn to simulate complex interactive environments, researchers hope to provide robots with the data and experience needed to operate effectively in the real world.
First agentic ransomware case identified
Security researchers have identified 'JADEPUFFER' as the first documented ransomware operation conducted entirely by a large language model agent. This marks a shift in cyber threats, where AI agents are being used to automate the entire attack lifecycle, from initial access to data encryption. The emergence of JADEPUFFER highlights the growing risk of autonomous AI systems being weaponized for malicious purposes. As AI agents become more capable of executing multi-step tasks, the barrier to entry for sophisticated cyberattacks is lowering, necessitating new defensive strategies. This development underscores the urgent need for robust security measures that can detect and mitigate autonomous AI threats. Organizations are being urged to review their security protocols to account for the potential of agent-driven attacks, which can operate with a speed and scale that traditional manual defenses may struggle to counter.
Illinois signs AI safety law
Illinois Governor JB Pritzker has signed the first U.S. state law requiring major AI developers to conduct annual third-party safety audits and reporting. The legislation, which has received backing from both Anthropic and OpenAI, aims to establish a framework for oversight of large-scale AI development. This move represents a significant step in state-level regulation of the AI industry. By mandating transparency and independent verification, the law seeks to address concerns regarding the safety and potential risks associated with the deployment of powerful AI models. As other states and federal bodies consider similar regulations, the Illinois law could serve as a template for future AI governance. The focus on third-party audits reflects a growing consensus that self-regulation by AI labs may be insufficient to ensure public safety.
ByteDance and Alibaba restrict AI companions
ByteDance and Alibaba are reportedly ending AI companion features in their respective services, Doubao and Qwen. This decision comes ahead of new regulations in Beijing, set to take effect in July, which target chatbots that foster emotional dependency. The move reflects the tightening regulatory environment in China regarding the social and psychological impact of AI. By restricting these features, the companies are aligning their products with government directives aimed at curbing potential negative societal effects of AI-driven emotional engagement. This development highlights the challenges tech companies face in balancing innovation with compliance in an increasingly regulated market. It also signals a broader trend where governments are taking a more active role in shaping the social applications of AI technology.
OpenAI's GPT 5.6 'Ultra' coming to Codex
OpenAI's Tibo Sottiaux has revealed that the upcoming GPT 5.6 'Ultra' variant will be integrated into the company's Codex platform. The model family is expected to be released later this week, marking another step in OpenAI's rapid iteration of its model lineup. This integration is expected to enhance the capabilities of Codex, providing developers with more powerful tools for automated coding and software development. The release of the 'Ultra' variant suggests a focus on high-performance tasks, potentially offering improvements in reasoning and code generation. As OpenAI continues to expand its model offerings, the integration into platforms like Codex demonstrates a strategy of embedding advanced AI capabilities directly into developer workflows. This could further accelerate the adoption of AI-assisted programming tools across the industry.
xAI rebrands to SpaceXAI
Elon Musk’s xAI has officially rebranded to SpaceXAI. This change follows the company's merger in February, which valued the AI lab at $250 billion. The rebrand aligns the AI division more closely with the SpaceX brand, potentially signaling a deeper integration of AI technologies into the company's aerospace and satellite operations. The move comes as the company continues to scale its AI infrastructure and research capabilities. While the rebrand is primarily a corporate shift, it underscores the strategic importance of AI within Musk's broader business ecosystem. As the company continues to develop its AI models, the integration with SpaceX's hardware and operations could provide unique advantages in the competitive AI landscape.