# Sep 10, 2024

## Daily

### Benchmarking Chinese Knowledge Rectification in Large Language Models
- **Large Language Models (LLMs) struggle with Chinese ancient poetry, proverbs, and idioms**, leading to the generation of nonsensical information due to a lack of specific knowledge.

### Transfusion: Predict the next token and diffuse images with one multimodal model
- **Transfusion** introduces a **novel training recipe** that merges language modeling and diffusion techniques, enabling a single transformer to process both text and image data efficiently.

### Radiology specific foundation model released by Harrison.ai
- **Harrison.rad.1** significantly outperforms other AI models, including OpenAI's GPT-4o and Google's Gemini 1.5 Pro, in the **FRCR 2B Rapids exam**, scoring **51.4 out of 60 (85.67%)**, where most competitors scored below 30.

### Launch HN: Deepsilicon (YC S24) – Software and hardware for ternary transformers
- **Deepsilicon** is developing **software and hardware** to train and run **ternary transformer models**, which compress weight matrices by almost **8x** and reduce arithmetic intensity, offering a significant leap in efficiency for large transformer-based models.

### Deductive Verification for Chain-of-Thought Reasoning in LLMs
- **Chain-of-Thought (CoT) prompting** enhances Large Language Models' (LLMs) reasoning capabilities but risks introducing **hallucinations and errors**, necessitating a method for **rigorous deductive reasoning** and **self-verification**.

### Sail – Faster and cheaper PySpark-compatible computation framework
- **Sail aims to unify stream processing, batch processing, and compute-intensive AI workloads**, positioning itself as a versatile tool in data processing and AI fields.

### [R] Revisiting Sparse Convolutional Model for Visual Recognition
- **Sparse convolutional models**, bridging the gap between **interpretability** and **empirical performance**, employ **differentiable optimization layers** as replacements for standard convolutional layers in deep neural networks.

### We're in the brute force phase of AI – once it ends, demand for GPUs will too
- **Gartner's chief of research for AI, Erick Brethenoux, claims that the current "brute force" phase of AI, heavily reliant on GPUs, is temporary**, as history shows that specialized hardware becomes obsolete when general-purpose machines catch up.

### Satellites Spotting Aircraft
- **Umbra Space operates a fleet of Synthetic Aperture Radar (SAR) satellites** capable of capturing high-resolution images through obstacles like clouds and camouflage, with their first satellite launched by SpaceX in 2021.

### [R] Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
- **LLM-generated ideas** are judged **more novel** than those by human experts in a large-scale study involving **over 100 NLP researchers**, highlighting LLMs' potential in **creative ideation**.

### Talaria: Interactively Optimizing Machine Learning Models for Efficient Inference
- **Talaria** is a **model visualization and optimization system** designed to address the challenge of fitting machine learning models on devices with limited resources by enabling interactive optimization and visualization.

### [P] I built a tool to minimize hallucinations with 1 hyperparameter search - Nomadic
- The **Nomadic tool** significantly **reduces hallucinations** in Retrieval Augmented Generation pipelines by **4X** with a **single hyperparameter search**, enhancing the reliability of generated content.

### Is telling a model to "not hallucinate" absurd?
- **Instructing a Large Language Model (LLM) to "not hallucinate" is feasible**, especially if it has been preference-fine-tuned with such directives, challenging initial skepticism about the effectiveness of this approach.

### [R] LowFormer: Hardware efficient Transformer Backbone Design
- **LowFormer** introduces a **hardware-efficient design** for vision backbones by blending **convolutions and transformer blocks**, focusing on actual **throughput and latency** rather than just MACs for a more accurate efficiency metric.

### [P] `costly`: a package for estimating costs & running times of LLM projects in advance
- **`costly` is a Python package designed to estimate costs and running times for LLM projects**, addressing the need for pre-expenditure planning in complex LLM workflows.