# Aug 17, 2024

## The future of Deep Learning frameworks

- **PyTorch's design philosophy** of flexibility and ease of use for quick prototyping has led to significant technical debt, making it ill-suited for scalable, high-performance applications, unlike **JAX**, which was developed by DeepMind with scalability and performance in mind.

## Does Reasoning Emerge? Probabilities of Causation in Large Language Models

- The paper introduces a **framework** to assess **large language models' (LLMs')** ability to **replicate real-world reasoning** using the **probability of necessity (PN)** and **probability of sufficiency (PS)**.

## DifuzCam: Replacing Camera Lens with a Mask and a Diffusion Model

- **DifuzCam innovates by replacing traditional camera lenses with an optical element and a diffusion model**, enhancing image quality significantly from raw sensor data.

## Open-source tool helps you convert PDF documents, web pages, etc., into Markdown

- **MinerU, an open-source tool for converting PDFs to machine-readable formats like markdown and JSON, was developed during the pre-training of [InternLM](https://github.com/InternLM/InternLM)**, aiming to address symbol conversion challenges in scientific literature.

## [R] Beating gpt-4o structured output with gpt-3.5 and haiku on cost, latency and accuracy

- **BAML significantly outperforms** native function calling APIs in **accuracy, cost, and speed**, achieving near-solution status on the **[Berkeley function-calling benchmark (BFCL)](https://gorilla.cs.berkeley.edu/leaderboard.html)** with models like **gpt-3.5-turbo** and **claude-haiku**, challenging the performance of **gpt-4o**.

## Colorize Lidar point clouds with camera images

- **Colorizing Lidar point clouds** by mapping them onto synchronized camera images enhances object identification, leveraging both the depth accuracy of Lidar and the color detail of photographs.

## BMX: A Freshly Baked Take on BM25

- **BMX, a new lexical search algorithm developed by Mixedbread and the Hong Kong Polytechnic University, outperforms BM25** by incorporating **entropy-weighted similarity** and **weighted query augmentation (WQA)**, enhancing search accuracy through better handling of query-document similarity and semantic understanding. [BMX Paper](https://arxiv.org/abs/2408.06643), [Baguetter](https://github.com/mixedbread-ai/baguetter)

## Synthesizing Abstract Transformers for Reduced-Product Domains

- The **new algorithm** proposed tackles the challenge of synthesizing **abstract transformers** for **reduced-product domains**, overcoming the limitations of previous methods by handling an expanded feature set and requiring component cooperation.

## [R] MAMBA 2 Head Dimension

- **MAMBA 2** introduces a significant change in tensor dimensionality by altering the head dimension from **1 to 64**, indicating a shift in how data is processed.

## [P] New LLM Pre-training and Post-training Paradigms: Comparing Qwen 2, Llama 3.1, Gemma 2, and Apple's FMs

- **New LLM training paradigms** now include both pre-training and post-training phases, with post-training encompassing supervised instruction fine-tuning and alignment, a methodology popularized by ChatGPT.

## [D] Question about shapes of Mamba algorithm

- The **S6 algorithm** from the Mamba paper introduces **input-dependent shapes** for transformations, raising questions about the unconventional shape of **C** as **(B,L,N)** instead of **(B,L,D,N)**. [Read the Mamba paper](https://arxiv.org/pdf/2312.00752)

## TwoMinutePapers - NVIDIA’s New AI Will Supercharge Gaming!

## Abstracts: August 15, 2024

- **Microsoft Research's "Abstracts" podcast episode on August 15, 2024, discusses the paper "[Personhood credentials: Artificial intelligence and the value of privacy-preserving tools to distinguish who is real online](https://www.microsoft.com/en-us/research/publication/personhood-credentials-artificial-intelligence-and-the-value-of-privacy-preserving-tools-to-distinguish-who-is-real-online/),** highlighting the increasing challenge of online deception by AI and proposing a privacy-preserving credential system to verify human users.
