Karpathy's New Ed-tech Startup, FlashAttention-3, and more!

After Tesla and OpenAI, Andrej Karpathy’s startup aims to apply AI assistants to education

Andrej Karpathy, former head of AI at Tesla and researcher at OpenAI, is launching Eureka Labs, an AI-focused education platform. The initiative aims to integrate AI teaching assistants to support human educators, though practical implementation and testing are not yet evident. Eureka Labs plans an AI course, LLM101n, to train students on building their own AI. The startup's precise business model and funding details remain unclear. Read more here

FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision

FlashAttention-3 significantly improves the performance of attention mechanisms in large language models by using new hardware capabilities on Hopper GPUs. The update achieves 1.5-2x speed increases over FlashAttention-2 and makes better use of low-precision processing, leading to more efficient GPU utilization and the ability to handle longer context lengths without compromising accuracy. These advances are crucial for faster training and inference of AI models. FlashAttention-3 is available on GitHub. Read more here

Anthropic's Claude adds a prompt playground to quickly improve your AI appsRead more here

Anthropic has introduced new tools within their language model, Claude, to help developers improve prompt engineering for AI applications. The features, available on the Anthropic Console under the Evaluate tab, include a built-in prompt generator and tools to test and compare prompts, providing quick feedback to enhance outputs for specific tasks. This aims to assist both novice and experienced developers, potentially reducing the necessity for dedicated prompt engineers. Read more here

AMD to Acquire Silo AI to Expand Enterprise AI Solutions Globally

AMD is set to acquire Silo AI, Europe's largest private AI lab, for approximately $665 million. The acquisition aims to enhance AMD's AI capabilities and deliver end-to-end AI solutions. The Silo AI team, led by CEO Peter Sarlin, will join AMD's Artificial Intelligence Group. This move is part of AMD’s broader strategy, building on previous investments in AI companies. The acquisition underscores AMD’s commitment to integrating AI into various sectors and expanding its AI ecosystem. The deal is expected to close in the second half of 2024. Read more here

Run CUDA on AMD Hardware with SCALE

SCALE by Spectral Compute is a GPGPU programming toolkit allowing native compilation of CUDA applications for AMD GPUs without modifying the CUDA code or build system. It features an nvcc-compatible compiler, CUDA runtime, and driver APIs for AMD GPUs, and supports libraries like cuBLAS via ROCm. Key tested GPUs include AMD's gfx1030 and gfx1100. SCALE stands out by enabling direct compilation of CUDA code for AMD GPUs, aiming for full compatibility with NVIDIA CUDA, and provides some language extensions. Users can contact the team via Discord or email for support. Read more here