# Jun 13, 2024

## Daily

### Uncensor any LLM with abliteration
- **Abliteration** is a technique that uncensors any Large Language Model (LLM) by removing its built-in refusal mechanism, enabling responses to all types of prompts without retraining.

### AMD's MI300X Outperforms Nvidia's H100 for LLM Inference
- The **AMD MI300X accelerator** outperforms NVIDIA's H100 in **LLM inference**, achieving **33% higher throughput** in real-world chat use cases, as demonstrated by TensorWave and MK1's collaboration.

### Announcing the Open Release of Stable Diffusion 3 Medium, Our Most Sophisticated Image Generation Model to Date
- **Stable Diffusion 3 Medium** is **Stability AI's most sophisticated text-to-image model** to date, featuring **two billion parameters** for enhanced image generation, including **photorealism and prompt adherence**.

### How Meta trains large language models at scale
- **Meta has transitioned from training numerous smaller AI models to fewer, but significantly larger, generative AI models**, necessitating a comprehensive overhaul of their software, hardware, and network infrastructure to support the increased computational demands.

### MLow: Meta's low bitrate audio codec
- **Meta's MLow codec** significantly enhances audio quality for real-time communications on low-speed connections, achieving **twice the quality of Opus at 6kbps**.

### What If We Recaption Billions of Web Images with LLaMA-3?
- Leveraging **LLaMA-3**, researchers **fine-tuned** and employed it to **recaption 1.3 billion images**, aiming to **enhance model training** across various vision-language tasks by semantically aligning textual descriptions.

### Launch HN: Overwatch (YC S22): OSINT platform for cyber and fraud risk
- Overwatch automates **OSINT and threat intelligence**, transforming vast data into **actionable insights** for cyber and fraud risk management, leveraging AI and NLP to filter noise and identify relevant threats.

### Accessing Math Solutions via Monte Carlo Self-Refine with LLaMa-3 8B
- The **MCT Self-Refine (MCTSr) algorithm** innovatively combines **Large Language Models (LLMs) with Monte Carlo Tree Search (MCTS)** to tackle complex mathematical reasoning, enhancing LLMs' accuracy and reliability in strategic tasks.

### Show HN: Pathway – Build Mission Critical ETL and RAG in Python (NATO, F1 Used)
- **Pathway**, a Python ETL framework, excels in stream processing, real-time analytics, and LLM pipelines, powered by a **scalable Rust engine** for enhanced performance and efficiency in both development and production environments.

### Can LLMs invent better ways to train LLMs?
- Sakana AI has developed **Discovered Preference Optimization (DiscoPOP)**, a novel algorithm for training Large Language Models (LLMs) to align with human preferences, outperforming existing methods like Direct Preference Optimization (DPO) across multiple evaluation tasks. [Read the paper](https://arxiv.org/abs/2406.08414)

### Luma AI Dream Machine
- **Luma Dream Machine** is an AI model designed to **generate high-quality, realistic videos** quickly from text and images, marking a significant step towards creating a universal imagination engine.

### François Chollet Announces New ARC Prize Challenge – Is It the Ultimate Test for AI Generalization?
- **François Chollet introduces the ARC Prize** to tackle the ARC-AGI benchmark, focusing on **AI's ability to generalize** from minimal examples, akin to human learning processes. [Tweet](https://x.com/fchollet/status/1800577019979411560)

### An Empirical Study of Mamba-Based Language Models
- **Selective state-space models (SSMs) like Mamba** address **Transformers' computational and memory inefficiencies**, showing potential in **language modeling** with capabilities that can match or exceed those of Transformers in controlled comparisons.

### Is grokking "solved"?
- The **[Grokfast paper](https://arxiv.org/abs/2405.20233)** introduces a method that **accelerates grokking by a factor of 50** in algorithmic datasets, suggesting a significant leap forward in understanding and applying this phenomenon.

### Can LLMs invent better ways to train LLMs?
- **Large Language Models (LLMs) can autonomously invent more effective training methods**, specifically by generating new preference optimization algorithms that outperform traditional, manually-crafted ones.