# Mar 17, 2024

## Ahead of IPO, Reddit blends advertising into user posts
- Reddit introduces **"free-form ads"** ahead of its IPO, allowing advertisers to create ads that closely resemble user posts, aiming to boost engagement and click-through rates as detailed in their [statement](https://www.redditinc.com/blog/advertise-like-a-redditor-with-our-all-new-completely-unique-ad-format-free-form-ads).

## How web bloat impacts users with slow devices
- **Web bloat significantly degrades performance on low-end devices**, with modern web forums like Discourse being nearly unusable on budget smartphones due to excessive CPU demands.

## Grok
- **Grok-1**, an open-weights model with **314B parameters**, is designed for **advanced natural language processing tasks**, requiring significant GPU memory due to its size and complexity.

## Lambda on hard mode: serverless HTTP in Rust
- **Modal introduced `modal-http`**, a service enabling HTTP and WebSocket requests to be handled by serverless functions, leveraging Rust for its performance and complexity management, and recently added **full WebSocket support** for real-time bidirectional messaging.

## LLM4Decompile: Decompiling Binary Code with LLM
- **LLM4Decompile** introduces a pioneering open-source large language model for decompiling Linux x86_64 binaries into human-readable C code, with plans to expand its capabilities to more architectures and configurations.

## Super Micro Computer has gone from an obscure server maker to $60B market cap
- **Super Micro Computer** has outperformed Nvidia by doubling its revenue this year, thanks to its role as a key supplier of servers equipped with Nvidia's AI chips for the AI boom.

## MM1: Methods, Analysis and Insights from Multimodal LLM Pre-training
- **A careful mix of image-caption, interleaved image-text, and text-only data** is essential for achieving **state-of-the-art few-shot results** in multimodal large language models (MLLMs), outperforming other pre-training methods.

## Y Combinator's chief startup whisperer is demoting himself
- **Michael Seibel** steps down as Y Combinator's managing director to focus more on direct mentorship within the startup incubator, amidst its strategic downsizing and external political challenges.

## Show HN: Flash Attention in ~100 lines of CUDA
- **tspeterkim/flash-attention-minimal** offers a **simplified CUDA and PyTorch re-implementation** of Flash Attention, aiming to be accessible and educational with a concise codebase. [Link to article](https://github.com/tspeterkim/flash-attention-minimal)

## Google Scholar search: "certainly, here is" -chatgpt -llm
- A Google Scholar search for **"certainly, here is"** reveals a significant number of academic papers likely generated by **ChatGPT**, indicated by sections starting with this phrase.

## Show HN: SatCat5, the open-source FPGA Ethernet switch
- **SatCat5** is an innovative mixed-media Ethernet switch project that enables communication across various devices, including microcontrollers, through Ethernet over nontraditional media like I2C, SPI, or UART, detailed in its [FAQ documentation](https://github.com/the-aerospace-corporation/satcat5/blob/main/doc/FAQ.md).

## The High-Risk Refactoring
- **Refactoring code** involves high risks, including potential damage to business operations, revenue loss, and decreased team morale, especially when integrating new features or making significant system changes.

## [R] Apple - MM1: Methods, Analysis & Insights from Multimodal LLM Pre-training
- Apple's **MM1 model**, a multimodal AI that integrates vision and language, sets a new benchmark in few-shot learning on multimodal tasks, as detailed in their [recent paper](https://arxiv.org/pdf/2403.09611.pdf).

## Imitation Learning (2023)
- **Comma.ai** initially attempted to automate driving using a model that predicted **steering angles from images**, but this approach failed due to error accumulation, leading to an inability to maintain lane integrity.

## [D] I don't understand how backprop works on sparsely gated MoE
- **Backpropagation in sparsely gated Mixture of Experts (MoE) models** raises concerns due to the potential exclusion of the correct expert during training, limiting the gate network's learning efficiency.