# Jan 29, 2025

## Daily

### OpenAI Furious DeepSeek Might Have Stolen All the Data OpenAI Stole from Us

- **OpenAI and Microsoft are investigating** whether DeepSeek improperly trained its R1 model using data from OpenAI, raising concerns about **unauthorized data usage** and potential violations of terms of service.

### New speculative attacks on Apple CPUs

- **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.

### Promising results from DeepSeek R1 for code

- **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 proves the future of LLMs is open-source

- **DeepSeek's open-source model** is a strategic move to build trust in Western markets, countering skepticism towards a Chinese AI company by allowing users to self-host and maintain control over their data.

### Questions censored by DeepSeek

- **DeepSeek-R1**, a leading open-source AI model, censors approximately **85%** of sensitive prompts related to topics like Taiwanese independence and the Cultural Revolution due to compliance with **CCP policies**.

### An Analysis of DeepSeek's R1-Zero and R1

- **R1-Zero's significance** lies in its ability to operate without human supervision, relying solely on reinforcement learning, which marks a potential shift in AI training paradigms towards systems that can adapt without human bottlenecks.

### DeepSeek's multi-head latent attention and other KV cache tricks

- **Key-Value (KV) caches** significantly enhance the efficiency of language models like ChatGPT and DeepSeek by reducing computational costs from **O(n³)** to **O(n²)**, allowing for faster text generation while managing memory trade-offs effectively.

### Machine learning and nano-3D printing produce nano-architected materials

- 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.

### Adding concurrent read/write to DuckDB with Arrow Flight

- **DuckDB** faces **concurrency limitations** that hinder its use in real-time analytics, specifically lacking support for concurrent writers and simultaneous read/write operations, which are essential for streaming data workflows.

### TokenVerse: Multi-Concept Personalization in Token Modulation Space by Google

- **TokenVerse** introduces a novel method for **multi-concept personalization** using a pre-trained text-to-image diffusion model, allowing for the **disentanglement of complex visual elements** from a single image while enabling the generation of diverse combinations from multiple images.

### DeepSeek's Hidden Bias: How We Cut It by 76% Without Performance Loss

- **Hirundo's bias unlearning technology** applied to DeepSeek-R1-Distill-Llama-8B achieved a **76% reduction in bias** without sacrificing performance, demonstrating a significant advancement in AI fairness.

### The scale vs. intelligence trade-off in retrieval augmented generation \[Discussion\]

- **Retrieval Augmented Generation (RAG)** faces a critical **trade-off** between **long-context models** that excel in reasoning but are limited by training data, and **embedding-based approaches** that scale well but lack deep reasoning capabilities.

### Mamba-Shedder: Post-Transformer Compression for Efficient Selective Structured State Space Models

- **Mamba-Shedder** enhances **Selective Structured State Space Models (SSMs)** by compressing them, achieving a **speedup of up to 1.4x** during inference while preserving accuracy.

### Integration of 1,024 silicon quantum dots with on-chip electronics

- Researchers at Quantum Motion successfully integrated **1,024 silicon quantum dots** with on-chip electronics, enabling a **quantum computing system** to function at temperatures below **1 K**, which enhances the potential for **silicon qubit-based technologies**.

### FactCG: Enhancing Fact Checkers with Graph-Based Multi-Hop Data

- **FactCG** enhances fact-checking by utilizing **multi-hop reasoning** on context graphs, improving detection of hallucinations in large language models (LLMs) beyond traditional methods.