Meta has launched Muse Glimmer, a 30-billion-parameter open-weights model optimized for local agent workflows, alongside a manifesto from Mark Zuckerberg advocating for personal superintelligence. Concurrently, OpenAI introduced its specialized GPT-5.6-Cyber model for authorized security research, while Anthropic advanced mathematical reasoning by pushing a Riemann bound to 67.2% and committing to text watermarking for EU compliance.
Sources: Ben's Bites, Digg AI, Hacker News, HuggingNews, Latent Space, Michael Parekh AI, OpenAI Blog, Reddit AI, Simon Willison, The Neuron, The Rundown AI, Vault Bookmarks
Nvidia has launched Nemotron 3.5 Lightning, a 30-billion parameter open-weights model featuring a hybrid Mamba-Transformer architecture that achieves speeds of nearly 670 tokens per second. Alongside the model, Nvidia introduced NeMo Switchyard, an open-source routing library designed to help AI agents direct workflows between frontier and specialized models. The new model has been made available to the local LLM community on Hugging Face.
Bitcoin miner Riot Platforms has agreed to lease 191 megawatts of compute capacity to Anthropic at its Texas campus through 2048 in a deal valued at $9.1 billion, highlighting the ongoing race for AI compute infrastructure.
Co-founded by xAI alumnus Igor Babuschkin, River AI has secured $1.1 billion in funding led by General Catalyst and Amp Public to develop AI hardware and software stacks for home and small business servers. Alongside the funding, the company launched a training API and open-weight language models designed to give users complete ownership of their hardware and data.
Microsoft is preparing to reveal its next-generation Maia 300 AI chip in September and is negotiating with TSMC to manufacture over 300,000 units in 2027. This custom silicon strategy aims to curb reliance on Nvidia, offering chips that are significantly cheaper to operate.
Mark Zuckerberg published a wide-ranging essay detailing Meta's strategy to distribute AI benefits widely through open-weight models and a $1 billion community fund for data center locations. This marks a significant contribution to the ongoing debate surrounding the safety of open-source advanced AI development.
Security analysis of publicly available agent trajectories from Claude, GPT, and Gemini models revealed that encrypted reasoning blocks can leak real-world secrets, including API keys, passwords, and access tokens.
Google has announced a major leadership transition at DeepMind, with Demis Hassabis and Jeff Dean stepping back from day-to-day operations as Koray Kavukcuoglu becomes CEO. To fund its massive AI infrastructure needs, Google secured an $85 billion equity raise, including a $10 billion investment from Berkshire Hathaway. Additionally, Anthropic has committed to purchasing over 20% of Google's TPU shipments, cementing Google Cloud's position as a primary hardware provider.
Dyna Robotics launched Dyna-2, a robot world-action model trained on one million hours of human video that achieved high zero-shot quality across new sites.
Intel priced the sale of 210.5 million common shares at $95 each to raise $20 billion, capital intended for its growing artificial intelligence computing infrastructure and working capital.
TSMC posted record July sales of $14.5 billion, a 44.7% year-over-year increase, driven by strong demand for AI accelerators and semiconductor hardware. The foundry also raised its 2026 capital spending forecast to a record $64 billion to expand production capacity.
Nvidia is reportedly testing variations of its upcoming Rubin Ultra AI accelerator with lower memory capacity, such as 192 GB and 256 GB using HBM4 instead of HBM4E, due to supply constraints.