Models & Labs
Google Developing New AI Chip for Gemini Models
TechCrunch AIhigh confidence
Why it matters
- →Custom AI chips can significantly enhance model efficiency and performance.
- →Reducing dependency on Nvidia could alter the competitive landscape in AI hardware.
- →Improved efficiency addresses concerns about AI computing capacity and costs.
Google is reportedly developing a new AI chip, 'Frozen v2', to enhance the efficiency of its Gemini models. Expected to launch in 2028, the chip could be six to ten times more efficient than existing AI chips, according to sources. This development is part of a broader trend where tech companies are creating custom chips to reduce dependency on Nvidia and improve AI model performance. The news has positively influenced Google's stock, indicating investor confidence in the company's AI strategy.
Read originalMore from TechCrunch AI
© TechCrunch AIMarket & Regulationbusiness
Anthropic's $1.5B Copyright Settlement Approved
Anthropic's $1.5 billion settlement over copyright infringement has been approved, marking a pivotal moment in the intersection of AI and copyright law. The settlement compensates authors and publishers $3,000 per work due to Anthropic's use of pirated books for AI training. While the court ruled that training AI on copyrighted text constitutes fair use, it highlighted the illegality of sourcing books from pirate sites. This decision remains a district court ruling and does not establish a binding precedent, leaving the broader legal landscape open. The case highlights ongoing tensions in the AI industry regarding the use of copyrighted materials, as similar lawsuits continue against other major tech companies.
TechCrunch AI
© TechCrunch AIModels & Labsagents
Model Context Protocol Update Eases AI Integration
The Model Context Protocol (MCP) is undergoing an update that aims to simplify its application in large-scale AI deployments. By adopting a stateless approach to session IDs, the update addresses the complexities faced by companies operating MCP servers across multiple machines. This change is expected to streamline operations and potentially reduce costs for businesses integrating AI agents. Although end users might not notice the difference, this development is a crucial step in refining AI infrastructure, which is necessary for deploying AI models effectively in real-world scenarios.
TechCrunch AI
© TechCrunch AIMarket & Regulationbusiness
Debate Over Open-Weight AI Models Intensifies
The rise of Moonshot's Kimi K3, a significant open-weight language model from China, has sparked a contentious discussion about AI's future in terms of innovation and regulation. OpenAI's Dean W. Ball initially suggested that the US government should impose regulatory barriers to protect American AI companies, but he later withdrew his statement. This debate highlights the tension between proprietary and open-source models, with data security and innovation concerns at the forefront. Advocates for open models argue they could democratize AI development, challenging the dominance of established US labs like OpenAI and Anthropic. Critics, however, worry about the potential impact on US AI leadership and data security. The outcome of this debate could significantly influence the balance of AI power between the US and China.
TechCrunch AI
More in Models & Labs
Models & Labsmodels
llama.cpp b10069 Release Enhances OpenCL Support
The b10069 release of llama.cpp brings notable improvements to OpenCL support, particularly targeting Adreno GPUs. By enabling broadcast for Adreno MUL_MAT and respecting view offsets, this update aims to boost performance for multi-stream operations on llama-server. The release also extends general GEMM/GEMV support for broadcast, which could optimize operations across different hardware setups. Although there are no revolutionary new features, these updates represent a consistent enhancement in compatibility and performance, especially for developers working with a range of hardware configurations.
llama.cpp Releases
Models & Labsmodels
llama.cpp b10075 release expands platform support
The b10075 release of llama.cpp marks a significant step in enhancing its compatibility across diverse hardware setups. With the addition of ROCm 7.2 support on Ubuntu, AMD GPU users can now enjoy improved performance. Windows users benefit from the inclusion of CUDA 13.3, ensuring better integration with NVIDIA GPUs. The update also brings Vulkan support, which optimizes GPU utilization for developers. Although no new model architectures are introduced, this release reinforces llama.cpp's role as a flexible and adaptable inference runtime for developers working in varied environments.
llama.cpp Releases
© FireshipModels & Labsmodels
Thinking Machines releases Inkling model
Thinking Machines has unveiled Inkling, a new open-weights model boasting 975 billion parameters. While the model is described as 'deliberately mid,' its release marks a significant step in the ongoing evolution of large language models. This development could provide new opportunities for developers seeking to leverage massive AI models with open access. The introduction of Inkling suggests a shift towards more accessible and customizable AI tools, potentially democratizing the use of advanced AI capabilities.
Fireship