US Copyright Office found AI companies breach copyright. Its boss was fired

The US Copyright Office has found that AI companies often breach copyright by using vast amounts of copyrighted material to train their models without permission or compensation. The day after the office released its report, its head, Shira Perlmutter, was fired, with some speculating that the move was related to the report's findings, which could impact major AI companies and donors to the Trump administration.

Read more

Continuous Thought Machines

Neurons in biological brains use timing and synchronization to compute, a property that is essential for flexibility and adaptability, but is often discarded in modern AI systems for efficiency and simplicity. Researchers have introduced the Continuous Thought Machine (CTM), a novel neural network architecture that incorporates neural timing as a foundational element, and have found surprising and encouraging results in bridging the gap between biological plausibility and modern AI scalability.

Read more

Avoiding AI is hard – but our freedom to opt out must be protected

As AI becomes increasingly integrated into daily life, from job applications to healthcare and finance, it is becoming nearly impossible to opt out of its influence without facing significant disadvantages. The growing reliance on AI is creating a divide between those who can navigate and benefit from these systems and those who are left behind, highlighting the need for policies that protect individual freedoms and provide transparency and accountability in AI decision-making.

Read more

Intellect-2 Release: The First 32B Model Trained Through Globally Distributed RL

INTELLECT-2 is a 32B parameter model trained using globally distributed reinforcement learning, marking a paradigm shift in decentralized training. The model was trained using a custom framework called PRIME-RL, which enables fully asynchronous reinforcement learning across a dynamic swarm of permissionless compute contributors, and has been open-sourced along with its code and data.

Read more

Paul McCartney, Elton John and other creatives demand AI comes clean on scraping

Over 400 UK creatives, including Paul McCartney and Elton John, have signed a letter to the prime minister demanding that AI companies disclose which copyrighted works they have used to train their models. The letter supports an amendment to the Data (Use and Access) Bill that would require AI firms to be transparent about the copyrighted material they ingest, in order to protect the rights of creators and prevent copyright theft.

Read more

Byte latent transformer: Patches scale better than tokens (2024)

The Byte Latent Transformer (BLT) is a new architecture that matches the performance of tokenization-based large language models while improving inference efficiency and robustness by encoding bytes into dynamically sized patches. BLT demonstrates the feasibility of scaling models trained on raw bytes, showing improved training and inference efficiency, as well as better scaling than tokenization-based models for fixed inference costs.

ArXiv

AI Powered Energy Management Systems – Prospects and Challenges

Microgrids, a key solution for a sustainable energy future, face challenges such as forecasting renewable energy demand and protecting against cyberattacks, which can be addressed through the use of artificial intelligence (AI) in energy management systems (EMS). The integration of AI in microgrid EMS has shown immense potential for optimizing energy management, and future research directions include the development of self-healing microgrids, blockchain integration, and the use of Internet of Things (IoT) technology.

ArXiv

TransMLA: Multi-head latent attention is all you need

Multi-head Latent Attention (MLA) is a method that reduces communication bottlenecks in large language models by using low-rank matrices and compressed latent key-value states, leading to faster inference. A new post-training method called TransMLA can convert models that use Group Query Attention (GQA) into MLA-based models, allowing for more efficient and expressive models without increasing the key-value cache size.

ArXiv

Toward a Sparse and Interpretable Audio Codec

Most modern audio codecs, including Ogg Vorbis and MP3, use block-coding to compress audio into fixed-size frames, but this method doesn't produce an intuitive representation. A new proof-of-concept audio encoder represents audio as a sparse set of events and their times, using physics-based assumptions to model the sound and create a more interpretable representation.

ArXiv

Zero-shot forecasting of chaotic systems

Foundation models, pre-trained on vast amounts of time-series data, have shown promise in general-purpose time-series forecasting, particularly in challenging tasks such as forecasting chaotic systems with limited training data. These models can produce competitive forecasts and preserve the properties of chaotic attractors, even when point forecasts fail, by using mechanisms such as in-context learning and context parroting to capture long-term behavior.

ArXiv

Show HN: Airweave – Let agents search any app

Airweave is a tool that enables semantic search across various apps, databases, and APIs, allowing agents to retrieve information efficiently. It supports over 25 integrations, offers features like data synchronization, entity extraction, and semantic search, and is built using a technology stack that includes React, FastAPI, PostgreSQL, and Qdrant.

Repo

Show HN: MCP Browser Agent – Autonomous Browser Automation for Claude Desktop

The MCP Browser Agent is a powerful integration that provides Claude Desktop with autonomous browser automation capabilities, allowing for advanced browser automation, API client functionality, and AI agent capabilities. It supports multiple browser types, including Chrome, Firefox, Microsoft Edge, and WebKit, and can be installed and run manually or configured to auto-start with Claude Desktop.

Repo

Show HN: Ragmate – Local RAG server for JetBrains with project-aware context

Ragmate is a local RAG server that scans your codebase, builds a context index, and connects to external LLMs for context-aware code generation, with features including integration with JetBrains IDEs and real-time file change tracking. It can be installed and set up using Docker Compose, and supports various OpenAI-compatible LLMs and embedding models, with plans for future integration with VS Code and additional language models.

Repo

Infio-copilot- A Cursor-inspired AI assistant for Obsidian

Infio-Copilot is a Cursor-inspired AI assistant for Obsidian that offers smart autocomplete and interactive chat with selected notes, allowing users to receive context-aware writing suggestions and edit notes directly within the current file. The plugin requires a recent version of the Obsidian installer and an API key from a supported platform, such as SiliconFlow, OpenAI, or Google, and can be installed and configured through the Obsidian settings.

Repo

OpenAI/GPT-2

The GPT-2 code and models are available in an archived state, with no updates expected, and are intended as a starting point for researchers and engineers to experiment with the technology. The models have limitations, including potential biases and inaccuracies, and users are advised to carefully evaluate GPT-2 for their specific use case, especially in safety-critical applications.

Repo