Nvidia, ASML Plunge as DeepSeek Triggers Tech Stock Selloff

• Nvidia and ASML shares fell sharply as the Chinese AI startup DeepSeek demonstrated competitive performance against Western chatbots at a significantly lower cost, raising concerns about the sustainability of the current AI business model reliant on high-end chips.

• Qwen2.5-1M models introduce two new checkpoints capable of processing 1 million tokens, significantly enhancing long-context capabilities compared to previous versions.

• Nvidia's stock faces significant challenges as competition intensifies from innovative architectures like Cerebras and Groq, which threaten its data center dominance and high margins, while major customers develop custom silicon to reduce reliance on Nvidia's products.

• Nvidia's stock plummeted 17%, resulting in a staggering $600 billion loss in market cap, marking the largest single-day drop for any U.S. company, driven by fears of competition from China's AI lab DeepSeek.

• DeepSeek's training efficiency is 45x greater, suggesting they could have led the LLM market, yet they opted for open-sourcing, likely to foster collaboration and innovation within the community.

• The Janus Pro Technical Report details advancements in AI technology, specifically focusing on enhanced inference capabilities and model efficiency that significantly improve performance in various applications.

• Janus-Pro is an innovative autoregressive framework that enhances multimodal understanding and generation by decoupling visual encoding, allowing for improved flexibility and performance compared to previous models.

• Notion serves as a comprehensive workspace that integrates notes, tasks, wikis, and databases, enhancing productivity through its versatile features and user-friendly interface.

• AVX-512 instruction enables a phrase search algorithm that outperforms Meilisearch by up to 1600x, showcasing significant advancements in computational efficiency for phrase searches in large datasets. The source code is available here and the benchmarking methodology is detailed in the article.

• ErisForge is a Python library that enables the modification of Large Language Models (LLMs) by transforming their internal layers, allowing for the creation of both ablated and augmented model versions tailored to specific inputs.

• Fast Think-on-Graph (FastToG) enhances large language models (LLMs) by enabling them to reason "community by community" within knowledge graphs (KGs), addressing limitations of existing Graph Retrieval Augmented Generation (GRAG) systems.

• DeepSeek-R1 is a groundbreaking open weights model that excels in reasoning tasks by utilizing a unique training method that incorporates long chains of reasoning and reinforcement learning, resulting in a model capable of generating detailed thought processes.

• Flash Attention 2 significantly enhances transformer inference speed, but requires careful installation using a -devel Docker container and the ninja build tool for optimal performance.

• ZETA introduces a novel method for top-k attention by utilizing Z-order curves, enabling efficient parallel querying of past tokens, which significantly enhances training efficiency for long sequences.