AlphaQubit: AI to identify errors in Quantum Computers

Google has developed an AI system called AlphaQubit that can accurately identify errors inside quantum computers, a major step towards making this technology more reliable. AlphaQubit is a collaborative effort between Google DeepMind and Google Quantum AI, and its state-of-the-art accuracy could help make quantum computers capable of performing long computations at scale.

Read more

Between the Booms: AI in Winter – Communications of the ACM

Science fiction writer Ted Chiang argues that the term "artificial intelligence" is misleading, as it implies sentience, and instead suggests "applied statistics" as a more accurate description. However, this view is disputed, as AI research has evolved over the years, shifting from rule-based systems in the 1960s-1980s to probabilistic methods and neural networks, with various approaches emerging during the "AI winter" of the 1990s and early 2000s.

Read more

Show HN: Rebuild of Blossom, an open-source social robot

The Blossom robot, an open-source platform for human-robot interaction research, has been rebuilt with a new design inspired by robot model kits, featuring a modular and customizable inner frame and improved hardware. The robot's software has also been refactored into a Python library called r0b0, a message-oriented middleware for connecting hardware peripherals and software applications.

Read more

Strava closes the gates to sharing fitness data with other apps

Strava has updated its API terms, restricting third-party apps from displaying user data to anyone other than the user themselves, prohibiting the use of Strava data for AI model training, and preventing apps from replicating Strava's look and feel. The changes, effective November 11, are expected to affect less than 0.1% of existing applications, but some popular services like VeloViewer and Final Surge may need to find alternative ways to provide features to Strava users.

Read more

U of T computational imaging researchers harness AI to fly with light in motion

Researchers at the University of Toronto have developed an advanced camera setup that can visualize light in motion from any perspective, allowing for the capture of ultrafast moments of a scene, such as a pulse of light speeding through a pop bottle or bouncing off a mirror. This technology has the potential to unlock new capabilities in several research areas, including non-line-of-sight imaging, imaging through scattering media, and 3D reconstruction, and could also inspire creative applications in the arts and improve LIDAR sensor technology used in autonomous vehicles.

Read more

Wave Network: An Ultra-Small Language Model

The Wave network, a new ultra-small language model, uses complex vectors to represent tokens and achieves high accuracy in text classification tasks, outperforming a single Transformer layer with BERT pre-trained embeddings. The Wave Network also reduces video memory usage and training time by 77.34% and 85.62% compared to the BERT base model while achieving comparable accuracy.

ArXiv

Adding Error Bars to Evals: A Statistical Approach to Language Model Evaluations

Evaluations of large language models can be improved by applying principles from experiment analysis and planning in other sciences. This article provides statistical formulas and recommendations for analyzing and reporting language model evaluation data to minimize noise and maximize informativeness.

ArXiv

Can the noble metals (Au, Ag and Cu) be superconductors?

Researchers have developed a theory suggesting that ultra-thin films of non-superconducting metals like gold, silver, and copper could exhibit superconductivity at low temperatures if their thickness is precisely controlled. According to the theory, a specific film thickness, around half a nanometer, would be required to induce superconductivity in these metals.

ArXiv

Birdie: Advancing State Space Models with Reward-Driven Objectives and Curricula

Efficient state space models (SSMs) struggle with tasks requiring long-range in-context retrieval, but a novel training procedure called Birdie significantly enhances their capabilities without altering their architecture. Birdie combines bidirectional input processing with dynamic mixtures of specialized pre-training objectives, resulting in improved performance on retrieval-intensive tasks and narrowing the performance gap with Transformers while retaining computational efficiency.

ArXiv

MolGrapher: Graph-Based Visual Recognition of Chemical Structures (2023)

MolGrapher is a system that automatically recognizes chemical structures from images, using a deep keypoint detector, graph representation, and Graph Neural Network to classify atoms and bonds. The system was trained on a synthetic data generation pipeline and evaluated on a large-scale benchmark, USPTO-30K, showing significant improvement over existing methods.

ArXiv

Show HN: Open-Source Apple Intelligence Writing Tools for Windows and Linux

Writing Tools is an AI-powered writing assistant for Windows and Linux that offers features like grammar correction, text summarization, and language translation, all accessible through a customizable hotkey. It supports various AI models, including local and cloud-based options, and prioritizes user privacy by storing data locally and not collecting logs or tracking user activity.

Repo

Show HN: Langrocks – tools like computer access, browser etc., for LLM agents

Langrocks is a collection of tools for Large Language Models (LLMs) that can be installed using pip and provides features such as web browsing, remote computer control, and file operations. It comes with example scripts and can be used to extract information, take screenshots, and interact with web pages and remote computers.

Repo

Argilla: Build high quality datasets for your AI models

Argilla is a collaboration tool for AI engineers and domain experts to build high-quality datasets for their projects, allowing for continuous evaluation and model improvement. It provides a programmatic approach to data management, enabling users to achieve and maintain high-quality standards for their data and improve the quality of their AI output.

Repo

Jax AI Stack

The JAX AI stack is a collection of packages for specialized numerical computing, built around the JAX Python package for array-oriented computation and program transformation. It can be installed with a single command, pip install jax-ai-stack, and includes core packages like JAX, flax, and optax, as well as optional packages like grain and tensorflow-datasets.

Repo

Show HN: Neuronic – Define AI functions in your apps with predictable responses

Neuronic is a Python library that leverages OpenAI's GPT models to transform, analyze, and generate data in various formats, serving as a versatile tool for data manipulation. It supports features such as data transformation, smart analysis, data generation, and context-aware transformations, making it suitable for various use cases including data processing, content generation, analysis, and development support.

Repo