Models & Labs
Inherent's AI Outperforms Larger Models in Research Task
TechCrunch AIhigh confidence
Why it matters
- →Inherent's approach challenges the notion that larger models are always superior by achieving better results with a smaller model.
- →The use of reinforcement learning to develop 'research taste' could lead to AI systems that contribute meaningfully to scientific discovery.
- →This development positions Inherent as a potential competitor to established AI giants like OpenAI and Anthropic.
Inherent, a London AI lab founded by former DeepMind alumni, has announced that its AI agent, Faraday, has outperformed larger models from Anthropic and OpenAI in replicating scientific research findings. Faraday, which operates on a smaller model with 27 billion parameters, uses reinforcement learning to develop an instinct for valuable experiments. This achievement marks a significant step for Inherent, which aims to build AI capable of discovering new scientific knowledge. The startup, which recently emerged from stealth with a $50 million seed round, plans to expand its team as it continues to develop its AI capabilities.
Read originalMore from TechCrunch AI
© TechCrunch AIModels & Labsmodels
Mystery Surrounds New AI Model Ox Alpha
The release of Ox Alpha, a new AI model on OpenRouter, has sparked widespread speculation about its origins. Described as a reasoning model for coding and sustained agentic work, Ox Alpha's developer remains anonymous, fueling intrigue. While some speculate it could be linked to Chinese company Z.ai's GLM models, others suggest it might be an unreleased version of Microsoft's MAI. The mystery has captivated the AI community, highlighting the growing interest in stealth AI projects and their potential impact.
TechCrunch AI
© TechCrunch AIMarket & Regulationother
Legal Complexity of AI Training on Copyrighted Books
The legality of training AI models on copyrighted books is a complex issue, as highlighted by recent legal cases. A notable ruling by Judge William Alsup found Anthropic's AI training lawful, penalizing the company for using pirated books rather than the act of training itself. This decision suggests that AI training might be seen as analogous to reading rather than copying, which could be advantageous for AI companies. However, the legal landscape remains uncertain, with fair use and copyright laws being interpreted in various ways as they struggle to keep pace with technological advancements.
TechCrunch AI
© TechCrunch AIGeneral AIagents
Harvard Bootcamp Uses AI Avatars for Feedback
Harvard Business School's Foundry bootcamp is integrating AI avatars to provide personalized feedback to participants. Created by the startup HeyGen, these avatars simulate instructors and offer guidance during practice pitches and board meetings. While some students initially expressed skepticism about AI, the avatars have been well-received for their interactive and engaging nature. This approach represents a novel use of AI in education, offering a more immersive and personalized learning experience. The program's success could signal a shift towards more AI-driven educational tools in the future.
TechCrunch AI
More in Models & Labs
Models & Labsmodels
Llama.cpp Adds PAD_REFLECT_1D Operation for Vulkan
Llama.cpp's latest update introduces the PAD_REFLECT_1D operation for its Vulkan backend, enhancing its capabilities in handling reflection logic. This addition is significant for developers working with Vulkan, as it provides a new compute shader implemented in GLSL, tested successfully on Intel Iris Xe. The update demonstrates improved performance metrics, with operations running efficiently at high data throughput. This release marks a step forward in optimizing Vulkan's functionality within the llama.cpp framework, offering developers more robust tools for their applications.
llama.cpp Releases
Models & Labsmodels
llama.cpp b10593 Release Fixes and Enhancements
The b10593 release of llama.cpp brings crucial improvements, particularly in model loading and rollback mechanisms. This update resolves issues with multi-sequence rollback and optimizes cache management for specific sequence IDs, enhancing the platform's robustness. Developers will notice a more stable environment, especially when working with complex model sequences. While there are no new models or architectures introduced, the release strengthens llama.cpp's position as a reliable inference runtime. It supports a diverse array of systems, from Apple Silicon to Windows with CUDA, ensuring developers can deploy across different hardware with confidence.
llama.cpp Releases
Models & Labsmodels
llama.cpp b10594 release optimizes GPU resource use
The latest update to llama.cpp, version b10594, introduces a significant optimization by skipping the device_info loop when log verbosity is not set to LOG_LEVEL_TRACE. This change prevents unnecessary GPU context creation and VRAM allocation, particularly with CUDA, where a 550 MB VRAM allocation was previously unavoidable. This update is particularly beneficial for users who do not wish to utilize GPU resources, as it reduces resource consumption without affecting functionality. By addressing this inefficiency, llama.cpp becomes more resource-efficient, especially in default configurations.
llama.cpp Releases