# Mar 17, 2025

## Undergraduate Disproves 40-Year-Old Conjecture, Invents New Kind of Hash Table

An undergraduate, **Andrew Krapivin**, has developed a new type of **hash table** that significantly speeds up data searches, contradicting a **40-year-old conjecture** by Andrew Yao regarding search efficiency limits.

## Akira ransomware can be cracked with sixteen RTX 4090 GPUs in around ten hours

**Akira ransomware** can be cracked using **sixteen RTX 4090 GPUs** in about **ten hours**, leveraging a new brute-force method that exploits its outdated encryption techniques, allowing affected companies to recover their data without paying the ransom.

## Can a LLM convert C, to ASM to specs and then to a working Z/80 Speccy tape? Yes

**LLMs can effectively convert C code to assembly and generate functional specifications**, enabling the creation of a working Z/80 Spectrum application, as demonstrated by a sales tax calculator project that transitioned from C to ASM and then to a high-level specification.

## I fine-tuned Qwen 2.5 Coder on a single repo and got a 47% improvement in code completion accuracy

Fine-tuning the **Qwen 2.5 Coder** model on a single repository resulted in a **47% improvement** in code completion accuracy, elevating performance from **25% to 36%** after just **500 iterations** on an RTX 4090 GPU.

## Milestone XAI/Interpretability papers?

**Key papers** in **XAI** include _Axiomatic Attribution for Deep Networks_ and _Sanity Checks for Saliency Maps_, which introduce foundational concepts that reshape our understanding of interpretability in AI.

## Double Descent in neural networks

**Double descent** in neural networks refers to a phenomenon where model performance improves after initial overfitting, leading to a second descent in error rates as model complexity increases.

## Technologies on Effectiveness and Efficiency: A Survey of State Spaces Models

**State Space Models (SSMs)** are gaining traction as an efficient alternative to **transformer-based models**, particularly excelling in tasks involving sequential data and longer contexts, showcasing comparable performance with notable efficiency improvements.

## Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models

**Block diffusion language models** merge the strengths of autoregressive and diffusion models, enabling **flexible-length generation** and enhanced inference efficiency through techniques like KV caching and parallel token sampling.

## 4D Language Fields for Dynamic Scenes via MLLM-Guided Object-wise Video Captioning

**4D LangSplat** integrates **4D Gaussian Splatting** with **multimodal LLMs** to create language-aware scene representations, enabling **3D-aware grounding** of language in dynamic environments without scene-specific training.

## GTC 2025 – Announcements and Live Updates

**GTC 2025** will showcase cutting-edge advancements in **AI, robotics, and accelerated computing**, featuring influential speakers like **Yann LeCun** and **Frances Arnold**, who will challenge conventional thinking and inspire innovation.

## BriLLM: Brain-inspired Large Language Model

**BriLLM** is the first **brain-inspired large language model** that diverges from traditional architectures, utilizing a **Signal Fully-connected flowing (SiFu)** approach to enhance interpretability across all nodes in the model's directed graph.

## AIstorian lets AI be a historian: A KG-powered multi-agent system for accurate biography generation

**AIstorian** is a **KG-powered multi-agent system** that significantly enhances biography generation by addressing challenges like **factual fidelity** and **stylistic adherence** that traditional LLMs struggle with.

## Collaboration is all you need: LLM Assisted Safe Code Translation

**UniTranslator** redefines code translation by utilizing a **collaborative framework** of multiple compact LLMs, achieving accuracy and efficiency comparable to larger models while focusing on specialized tasks.

## Limits of KV Cache Compression for Tensor Attention based Autoregressive Transformers

**KV cache compression** in autoregressive transformers is a critical factor limiting the context length of large language models (LLMs), with our research extending previous findings on space complexity to tensor attention mechanisms.