Apr 27, 2024

• Tiny rubber spheres used to make a programmable fluid
  Researchers at Harvard University have developed a metafluid that can change its properties under pressure, using tiny rubber spheres that buckle, offering a new approach to soft robotics. Adel Djellouli/Harvard SEAS

• Searchformer: Beyond A – Better planning with transformers via search dynamics*
  Facebook Research's Searchformer introduces a novel approach to planning with transformers, enhancing algorithmic efficiency by leveraging search dynamics bootstrapping, as detailed in their paper Beyond A*: Better Planning with Transformers via Search Dynamics Bootstrapping.

• LLMs: Why does in-context learning work? What exactly is happening from a technical perspective?
  In-context learning in Large Language Models (LLMs) operates not through cognitive processes but by statistically weighting tokens similar to provided examples, enhancing output relevance and accuracy.

• What can LLMs never do?
  LLMs struggle with tasks requiring iterative reasoning or memory of previous states, such as playing Wordle or predicting cellular automata, despite excelling in other complex areas like chess or coding.

• Removing Reflections from RAW Photos
  A new system removes real-world reflections from RAW photos, utilizing an optional contextual photo from the opposite direction to enhance clarity.

• The Landscape of Machine Learning on ArXiv
  The ArXiv Machine Learning Landscape presents a visual data map showcasing the distribution and relationships of machine learning papers.

• Show HN: Cognita – open-source RAG framework for modular applications
  Cognita is an open-source framework designed to transition RAG (Retrieval-Augmented Generation) systems from prototyping in Jupyter notebooks to scalable, modular production environments, utilizing Langchain/LlamaIndex for experimentation and offering a structured codebase with API-driven, extendable components.

• Llama-3 based OpenBioLLM-70B & 8B: Outperforms GPT-4, Gemini, Meditron-70B, Med-PaLM-1 & Med-PaLM-2 in Medical-domain
  OpenBioLLM-Llama3-70B & 8B models set a new benchmark in the biomedical domain, outperforming major models like GPT-4, Gemini, and Med-PaLM series, marking them as the most advanced open-source medical-domain LLMs currently available.

• Direct-Style Effects Explained
  Direct-style effects, also known as algebraic effects and effect handlers, aim to simplify coding by allowing natural style code without monads while retaining benefits of reasoning and composition, as seen in languages like Unison and OCaml, and discussions for Scala implementation.

• Relational Graph Convolutional Networks for Sentiment Analysis
  Relational Graph Convolutional Networks (RGCNs) enhance sentiment analysis by capturing complex relationships between entities, offering a significant improvement over traditional and deep learning models.

• Einsum for Tensor Manipulation
  Einsum offers a versatile approach for tensor manipulation in ML, enabling operations like matrix multiplication and dot products with explicit documentation of tensor dimensions and implicit reordering.

• WebSim, WorldSim and the Summer of Simulative AI
  Simulative AI is emerging as a pivotal frontier in generative AI, with Joscha Bach of Liquid AI, Karan Malhotra of Nous Research, and Rob Haisfield of WebSim.ai offering insights into its potential to explore adjacent multiverses and create immersive, interactive simulations.

• Let's Think Dot by Dot: Hidden Computation in Transformer Language Models
  Transformers can solve complex algorithmic tasks by using meaningless filler tokens instead of detailed chain-of-thought responses, challenging the necessity of human-like task decomposition.

• Large language models as research assistants
  Large language models like GPT-4 are increasingly being used as research assistants, capable of outperforming humans in complex tasks such as the bar exam and matching the abilities of math Olympians, indicating a shift towards AI-driven academic research.

• How do I convince my superior to do data preprocessing?
  Data preprocessing is advocated by an AI Engineer as crucial for enhancing chatbot performance, contrasting with their superior's preference for prompt engineering without preprocessing.