# Jan 28, 2025

## Nvidia’s $589B DeepSeek rout

- **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.

- **DeepSeek releases Janus Pro**, a text-to-image generator [pdf]

- 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.

- **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.

- **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.

- **ggml achieves a remarkable x2 speed increase for WASM** by optimizing SIMD instructions in the `qX_K_q8_K` and `qX_0_q8_0` dot product functions, significantly enhancing performance for web applications.

- **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.

- **SLAP and FLOP** are speculative execution attacks targeting Apple CPUs, exploiting the **Load Address Predictor (LAP)** and **Load Value Predictor (LVP)**, respectively, to access sensitive data through incorrect memory operations.

- **DeepSeek v3** achieves state-of-the-art performance with only **2.8 million H800 hours** of training, significantly less than Llama 3.1, by innovating with **multi-head latent attention (MLA)** and **mixture-of-experts (MoE)** techniques.

- **Cleveland police's reliance on AI facial recognition** to justify a search warrant led to the dismissal of key evidence in a murder case, as the technology's results are deemed inadmissible in court, highlighting the **critical need for regulatory oversight** in law enforcement practices.

- **Berkeley researchers** have successfully replicated **DeepSeek R1's core technology** for under **$30**, demonstrating that small models can achieve complex reasoning capabilities, thus democratizing AI research.

- Researchers at the University of Toronto have developed **nano-architected materials** that combine the **strength of carbon steel** with the **lightness of Styrofoam**, utilizing machine learning to optimize their design for enhanced performance.

- **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.

- **Sam Altman's assertion** that startups with only **$10 million** are "totally hopeless" against giants like OpenAI has been challenged by **DeepSeek**, which claims to have trained its model for just **$5.6 million**, showcasing a potential shift in the competitive landscape of AI startups.

- **Qwen2.5-Max** is a large-scale **Mixture-of-Expert (MoE) model** pretrained on over **20 trillion tokens**, utilizing advanced techniques like **Supervised Fine-Tuning (SFT)** and **Reinforcement Learning from Human Feedback (RLHF)** to enhance its intelligence.

- **Diffusion models**, particularly the **Diffusion Transformer (DiT)**, have transformed content synthesis, yet they struggle with **generation diversity**, which this work addresses through targeted attention feature injection for consistent image edits.