# Apr 1, 2024

## InternLM2
- **InternLM2**, an **open-source Large Language Model (LLM)**, **surpasses its predecessors** in **6 dimensions and 30 benchmarks**, leveraging **innovative pre-training and optimization techniques**. [Read more](https://arxiv.org/abs/2403.17297)

## LLaMA now goes faster on CPUs
- **Justine Tunney has developed 84 new matrix multiplication kernels for [llamafile](https://github.com/mozilla-Ocho/llamafile)**, enhancing its prompt and image processing speeds by 30% to 500% on CPUs, with significant improvements on ARMv8.2+, Intel, and AVX512 architectures.

## Upscayl – Free and Open Source AI Image Upscaler
- **Upscayl introduces version 2.11**, enhancing its capability to upscale and improve low-resolution images using advanced AI algorithms, ensuring high-quality enlargements without quality loss.

## WSJ: The AI industry spent 17x more on Nvidia chips than it brought in in revenue
- The **AI industry's expenditure** on Nvidia chips for training advanced AI models was **$50 billion** last year, starkly contrasting with its **revenue of only $3 billion**.

## Mini-Gemini: Mining the Potential of Multi-Modality Vision Language Models
- **Mini-Gemini** enhances **multi-modality Vision Language Models (VLMs)** by focusing on **high-resolution visual tokens**, **high-quality data**, and **VLM-guided generation**, aiming to bridge the performance gap with advanced models like **GPT-4** and **Gemini**.

## A proposal to add signals to JavaScript
- The **JavaScript Signals standard proposal** aims to standardize a common model for signals in JavaScript, drawing design input from major frameworks like Angular, Vue, and React, to facilitate interoperability and enhance reactivity in web development. [Read more](https://github.com/proposal-signals/proposal-signals)

## LLM Paper on Mamba MoE: Jamba Technical Report from AI2
- **Jamba** introduces a **hybrid Transformer-Mamba architecture** with a mixture-of-experts (MoE) layering, enhancing model capacity while optimizing parameter usage for efficiency.

## Could the cosmos, in fact, be conscious?
- **Professor Philip Goff** advocates for **Cosmopsychism**, suggesting the **universe is conscious** and intentionally created conditions for life, challenging traditional views on creation and the existence of God.

## Can GPT optimize my taxes? An experiment in letting the LLM be the UX
- **Tax Driver**, an AI tax advisor, was developed to explore the concept of LLMs as operating systems by integrating GPT with the `tenforty` Python library and the Open Tax Solver package, enabling users to evaluate complex tax scenarios with ease. [Open Tax Solver](https://opentaxsolver.sourceforge.net/)

## 3Blue1Brown: But what is a GPT? 
- **GPT** stands for **Generative Pretrained Transformer**, where "Generative" indicates its ability to create new text, "Pretrained" signifies its initial learning from vast data, and "Transformer" refers to a specific neural network model crucial for recent AI advancements.

## Ask HN: What is the current (Apr. 2024) gold standard of running an LLM locally?
- **Running an LLM locally** involves navigating a variety of options, with the community seeking **idiot-proof** solutions for high-end hardware like the **3090 24Gb GPU**.

## What's more impressive in a ML portfolio: implementing a paper or creating a good project?
- **Hiring managers in the ML field value both implementations of papers and practical projects**, with each showcasing different strengths such as theoretical understanding and practical application skills.

## Can't escape OpenAI in my workplace, anyone else?
- **OpenAI's dominance** in the workplace is underscored by a **surge in requests** to use its API, overshadowing alternatives and reflecting a broader industry trend.

## Adaptive RAG – dynamic retrieval methods adjustment
- **Adaptive-RAG** introduces a **novel adaptive QA framework** that dynamically selects the most suitable strategy for retrieval-augmented LLMs based on **query complexity**, enhancing response accuracy across various tasks.

## Not so fast, Mr. Fourier
- The **discrete Fourier transform (DFT)** and its faster variant, the **fast Fourier transform (FFT)**, are pivotal in converting time-domain signals into their frequency-domain counterparts, enabling applications like audio processing and data compression.