🧠 Anthropic discovers Claude's self-built J-space reasoning layer 🧠
New interpretability research reveals Claude built an internal reasoning integration nobody programmed.

Anthropic published new interpretability findings showing that Claude developed an internal structure researchers call "J-space" — a global workspace for active concepts during reasoning that was never explicitly designed. The structure resembles neuroscience theories of conscious thought, acting as an internal notepad for complex problem-solving. Experiments showed that editing thoughts in J-space changes Claude's responses, while deleting it preserved basic chat abilities but collapsed multi-step reasoning — it offers a concrete tool for monitoring hidden AI reasoning chains. Researchers caution the discovery does not reveal whether Claude is conscious. The finding underscores Anthropic's continued leadership in mechanistic interpretability and opens new safety inspection avenues for frontier models.

🚀 xAI rebrands to SpaceXAI after $250B merger 🚀
The $250 billion integration reshapes AI's corporate landscape and resource access.

Elon Musk's AI company has officially rebranded from xAI to SpaceXAI, completing its integration into SpaceX following the February merger valued at approximately $250 billion. The move unifies Musk's AI, social media, and space businesses under a single corporate brand, accompanied by a new logo and updated social media handles. SpaceXAI is ramping up investments in AI infrastructure and large-scale compute capacity — this concentration of resources under one entity intensifies competitive pressure across the industry. The consolidation signals Musk's ambition to build a vertically integrated conglomerate spanning rockets, communications, and artificial intelligence.

🐲 Tencent ships Hy3, a 295B open-source MoE rival 🐲
A 295B-parameter open-weight model rolls out with free access on OpenRouter.

Tencent's Hunyuan team released Hy3, a 295-billion-parameter Mixture-of-Experts model with 21 billion active parameters, under a permissive Apache 2.0 license. Hy3 outperforms similarly sized open-source models and rivals flagship systems with two to five times more parameters on multiple benchmarks. While GLM-5.2 still leads on coding tests like SWE-bench, Hy3 tops open rivals on web research and tool-use tasks. The model is available for free on OpenRouter until July 21, lowering the barrier for developers to experiment — a clear adoption signal as Chinese tech giants deepen their push into the open-weight frontier race.

INTERESTING TO KNOW

⚡ Apple extends Broadcom AI chip deal through 2031 ⚡

Broadcom and Apple have extended their chip partnership through 2031, collaborating on multiple generations of custom ASIC silicon for AI workloads — a rollout that secures long-term chip supply for Apple's AI infrastructure push. Apple plans to deploy its own advanced AI servers as early as 2027, signaling a deeper move into in-house compute. The deal reflects the broader industry trend of major tech companies investing in custom silicon to optimize AI performance and reduce reliance on third-party processors.

📊 Study: AI-adopting firms boosted hiring by 10% 📊

A new study of more than 21,000 U.S. companies found that firms investing most heavily in AI increased their overall workforce by 10% over the past two years, with entry-level hiring rising 12% — an adoption signal that challenges the prevailing narrative of AI-driven job displacement. The data suggests AI helps companies grow faster and create roles that complement automated processes, carrying significant implications for labor policy and workforce planning.