May 8, 2024

• Apple introduces M4 chip

  • The M4 chip, utilizing second-generation 3-nanometer technology, significantly enhances the iPad Pro's performance and efficiency, featuring a new display engine for the Ultra Retina XDR display and a 10-core CPU and GPU for advanced computing and graphics capabilities.

• AlphaFold 3 predicts the structure and interactions of life's molecules

  • AlphaFold 3, developed by Google DeepMind and Isomorphic Labs, predicts the structure and interactions of proteins, DNA, RNA, and more, aiming to revolutionize our understanding of biology and drug discovery.

• DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model

  • DeepSeek-V2 introduces a Mixture-of-Experts (MoE) language model with 236B parameters, achieving top-tier performance with only 21B activated parameters per token and supporting a context length of 128K tokens.

• IBM Granite: A Family of Open Foundation Models for Code Intelligence

  • IBM introduces the Granite series of decoder-only code models, trained on code in 116 languages, achieving state-of-the-art performance in various code generative tasks such as bug fixing and code documentation.

• XLSTM: Extended Long Short-Term Memory

  • xLSTM introduces exponential gating and modified memory structures, enhancing traditional LSTM capabilities to rival modern Transformer and State Space Models in language modeling tasks.

• Stack Overflow upset over users deleting answers after OpenAI partnership

  • Stack Overflow users are editing or attempting to delete their contributions in protest against the site's partnership with OpenAI, which aims to integrate AI-generated content with user-generated answers.

• Gradient descent visualization

  • Gradient Descent Viz is a desktop application designed to visualize various gradient descent methods in machine learning, aiming to foster intuitive understanding among users of all expertise levels.

• TimesFM: Time Series Foundation Model for time-series forecasting

  • TimesFM, developed by Google Research, is a pretrained model for time-series forecasting, focusing on univariate forecasts with support for various horizon lengths and an optional frequency indicator. Paper \| Google Research blog \| Hugging Face checkpoint

• ScrapeGraphAI: Web scraping using LLM and direct graph logic

  • Scrapegraph-ai is an open-source library designed for AI-powered web scraping, enabling users to scrape thousands of web pages efficiently with minimal setup.

• Show HN: I made a better Perplexity for developers

  • Devv AI introduces lightning-fast answers, documentation, and code snippets for developers, enhancing productivity and efficiency in coding tasks.

• Show HN: AI climbing coach – visualize how to climb any route based on your body

• Consistency LLM: converting LLMs to parallel decoders accelerates inference 3.5x

  • Consistency Large Language Models (CLLMs) introduce a new method for parallel decoding, significantly reducing inference latency by efficiently decoding an $n$-token sequence per inference step, leveraging Jacobi trajectories for training. Read the paper for detailed insights.

• \[Research] xLSTM: Extended Long Short-Term Memory

  • xLSTM introduces exponential gating and modified memory structures, enhancing LSTM's performance to rival that of state-of-the-art Transformers and State Space Models.

• Model Spec

  • The Model Spec is a guideline for OpenAI's models, including ChatGPT, focusing on desired behavior, conflict resolution, and safe usage, aiming to enhance transparency and public engagement in model development.

• Deterministic Quoting: Making LLMs safer for healthcare

  • Deterministic Quoting is a technique developed by Invetech to ensure LLMs quote verbatim from source material, addressing the challenge of hallucinations in healthcare applications.