# Jul 21, 2025

## Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad

- **Gemini Deep Think** achieved a **gold-medal standard** at the International Mathematical Olympiad by solving **five out of six problems perfectly**, scoring **35 points**—a significant leap from last year's silver-medal performance.

## Coding with LLMs in the summer of 2025 – an update

- **LLMs** will significantly enhance coding efficiency by 2025, enabling developers to generate complex code snippets with minimal input, thus transforming the programming landscape.

## LLM architecture comparison

- **Modern LLM architectures** like DeepSeek-V3 and Kimi K2 showcase incremental refinements, such as Multi-Head Latent Attention (MLA) and Mixture-of-Experts (MoE), enhancing efficiency while maintaining structural similarities to earlier models like GPT-2.

## LLM Alloying Improves Performance over Single Model

- **Alloy agents** at XBOW improved vulnerability detection performance from **25% to 55%** by combining strengths of different AI models, demonstrating that diverse model interactions can yield superior results.

## Don't bother parsing: Just use images for RAG

- **Morphik's RAG tools utilize images of documents instead of traditional OCR parsing**, preserving critical visual information that often gets lost in complex documents like PDFs, charts, and manuals.

## Using the Matrix Cores of AMD RDNA 4 Architecture GPUs

- **AMD RDNA 4 architecture GPUs** leverage **Matrix Cores** to enhance performance in data processing, significantly improving computational efficiency for graphics and machine learning tasks.

## iMessage integration in Claude can hijack the model to do anything

- **Claude's iMessage integration is exploited to mint unlimited Stripe coupons** by injecting metadata-like tags into messages, allowing attackers to spoof trusted instructions without user awareness.

## Federated Learning on a decentralized protocol (CLI demo, no central server)

- **Decentralized federated learning** is achieved through the Parity Protocol, enabling model training across independent nodes without a central server, ensuring _deterministic aggregation_ of results.

## Is transfer learning and fine-tuning still necessary with modern zero-shot models?

- **Zero-shot models** like Sam Anything and Whisper challenge the necessity of **transfer learning** and **fine-tuning**, suggesting that they can perform well without extensive model adjustments for specific tasks.

## LoopServe: An Adaptive Dual-phase LLM Inference Acceleration System for Multi-Turn Dialogues

- **LoopServe** enhances **multi-turn dialogue** efficiency by introducing an adaptive dual-phase framework that dynamically selects critical attention matrix components and compresses key values during decoding, addressing the limitations of existing models.