An AI company you never heard of just raised $1 billion

An AI company you never heard of just raised $1 billion

AI cloud computing startup CoreWeave has raised $1.1 billion, bringing its valuation to $19 billion. The New Jersey-based company, founded by three former commodities traders, pivoted from Ethereum mining to providing GPU-accelerated cloud infrastructure essential for AI model training and deployment. With significant backing and expansions, CoreWeave now supports advanced computational workloads and competes in the growing AI infrastructure market. (Read More)

Reproducing GPT-2 (124M) in llm.c in 90 minutes for $20

Andrej Karpathy discusses how to reproduce the GPT-2 model (124M) using the llm.c implementation in C/CUDA in about 90 minutes with a cost of around $20. The detailed guide includes installation steps, dataset preparation, compilation, and training procedures. The project outperforms the original OpenAI model on certain benchmarks and offers more efficient training options with planned future enhancements and larger model support. The process benefits from hardware like 8X A100 80GB GPUs, but can also run on less powerful single GPUs. (Read More)

LLMs Can't Do Probability

The article by James Ravey discusses the limitations of large language models (LLMs) in handling probabilistic requests. Specifically, when prompted to generate responses with specific probabilities (e.g., saying "left" 80% of the time and "right" 20% of the time), models like GPT-3 and GPT-4 fail to produce the desired distributions reliably. Instead, they output biased results, favoring one answer disproportionately. Ravey suggests that manipulation using external scripts might be necessary to achieve controlled randomness until LLMs improve their handling of probabilistic tasks. (Read More)

The AI Hardware Dilemma

AI hardware startups face an uphill battle against the dominant iPhone. While devices like the Humane Pin and Rabbit R1 struggle, significant investments continue to pour into new AI hardware. For success, these startups need to explore unique use cases, possibly focusing on niche markets or innovative interfaces. The challenge lies in delivering AI capabilities that can genuinely compete with top-tier smartphones. (Read More)

Apple Reportedly Building M2 Ultra and M4-Powered AI Servers

Apple is reportedly constructing its own AI servers using M2 Ultra chips, with plans to use M4 chips by late 2025. This move aligns with Apple's strategy to enhance the performance of its data centers and future AI tools. The company also aims to vertically integrate its supply chain. While focusing on on-device AI processing for upcoming tools, cloud operations will be necessary, supported by these AI servers. Apple is expected to announce related on-device AI features at WWDC in June, with plans to incorporate AI features heavily in the iPhone 16 and iPhone 17 models. (Read More)