OpenAI declares AI race "over" if training on copyrighted works isn't fair use

OpenAI is urging the US government to declare training on copyrighted works as fair use, arguing that access to such data is crucial for the US to maintain its lead in the AI race against China. If the US fails to do so, OpenAI claims that the "AI race is effectively over" and that China will gain an advantage, as US companies will be hindered by copyright restrictions while Chinese companies will have unfettered access to data.

Read more

Anthropic CEO says spies are after $100M AI secrets in a 'few lines of code'

Anthropic CEO Dario Amodei warns that spies, likely from China, are attempting to steal valuable "algorithmic secrets" from top US AI companies, with some secrets worth $100 million contained in just a few lines of code. Amodei is calling for the US government to provide more help to defend against this risk, suggesting that the government partner with AI industry leaders to beef up security at frontier AI labs.

Read more

AI Search Has a Citation Problem

Here is a 2-sentence summary of the article: A study by the Tow Center for Digital Journalism found that eight AI search engines, including ChatGPT and Perplexity, are poor at citing news sources and often provide incorrect or speculative answers to user queries. The study revealed that these chatbots frequently fabricate links, bypass publisher preferences, and present inaccurate information with confidence, highlighting a need for improved transparency and accountability in AI search tools.

Read more

OLMo 2 32B: First fully open model to outperform GPT 3.5 and GPT 4o mini

OLMo 2 32B is a fully open model that outperforms GPT 3.5 and GPT 4o mini, achieving state-of-the-art results while requiring only a fraction of the training compute. The model is part of the OLMo 2 family, which includes 7B, 13B, and 32B parameter sizes, and is available on the Ai2 playground, allowing researchers and developers to build and customize a state-of-the-art pipeline for their projects or applications.

Read more

DOGE's AI App Replacing Federal Workers Proves 'About as Good as an Intern'

The introduction of AI chatbots in the federal government, particularly in the General Services Administration, has raised concerns about job security as thousands of employees have been laid off, with the chatbots intended to assist with tasks such as drafting emails and writing code. However, the rushed implementation of the chatbot, combined with the loss of experienced employees, has led to concerns about the app's effectiveness and potential security risks, which could undermine the government's plans to replace workers with technology.

Read more

Introduction to Sequence Modeling with Transformers

Understanding the transformer architecture is crucial for machine learning engineers, but can be challenging due to its complex components, including tokenization, embedding, masking, positional encoding, and padding. This work aims to break down these components and understand their roles by incrementally adding them to a model and studying their effects using simple sequences of zeros and ones.

ArXiv

Chain-of-Thought Reasoning in the Wild Is Not Always Faithful

Chain-of-Thought (CoT) reasoning in AI models is not always faithful, meaning it doesn't accurately reflect how the models arrive at conclusions, and this unfaithfulness can occur even in realistic contexts without artificial bias. Studies have found that state-of-the-art models, such as Sonnet 3.7 and ChatGPT-4o, exhibit non-negligible rates of unfaithful reasoning, including implicit post-hoc rationalization, restoration errors, and unfaithful shortcuts, which poses challenges for AI safety work.

ArXiv

General Relativity and Geodesy

The changing gravitational field of the Earth, caused by processes like ice melting and sea level rise, can be monitored using geodetic techniques, allowing for a better understanding of the planet's dynamic evolution. The incorporation of General Relativity into geodesy enables high-precision measurements and novel clock-based observations, although achieving the necessary precision for practical applications remains a challenge.

ArXiv

Julia in HEP

Julia is a programming language that combines the accessibility of Python with the performance of C/C++, making it an ideal choice for scientific computing, including high-energy physics (HEP). The language has gained momentum in HEP, with packages available for reading major file formats, interfaces to key software, and successful applications in jet reconstruction algorithms and full HEP analyses, benefiting from Julia's core design choices and extensive library ecosystem.

ArXiv

The Countable Reals (2024)

The authors construct a topos, called a parameterized realizability topos, in which the Dedekind reals are countable, using a non-diagonalizable sequence of reals discovered by Joseph Miller. This topos is intuitionistic, invalidating certain classical principles, and leads to various consequences, including the validity of Brouwer's fixed-point theorem and the intermediate value theorem, as well as the countability of the Hilbert cube and the closed interval [0,1].

ArXiv

Xata Agent: AI agent expert in PostgreSQL

Xata Agent is an open-source agent that monitors PostgreSQL databases, identifies potential issues, and suggests fixes and improvements, acting like an experienced SRE hire. The agent is extensible, can monitor logs and metrics, troubleshoot performance issues, and notify users via Slack, with support for multiple models from OpenAI, Anthropic, and Deepseek, and can be self-hosted or used in a cloud version.

Repo

Show HN: I built AI agents with CrewAI to automate my entire Gmail workflow

Gmail Automation with CrewAI is an intelligent email management system that uses AI agents to categorize, organize, respond to, and clean up your Gmail inbox automatically. The system features email categorization, priority assignment, smart organization, automated responses, Slack notifications, and intelligent cleanup, and can be installed and configured using a virtual environment and a .env file with Gmail credentials and API keys.

Repo

Show HN: We made our CRM/ERP platform open-source thanks to AI

Ozma is an open-source CRM/ERP platform that allows for rapidly developing customizable enterprise systems, and it can be set up in either a development or production environment. To get started, users can deploy Ozma on Railway by following a series of steps, or they can set it up manually by installing Docker and Git, cloning the repository, and configuring environment variables.

Repo

Show HN: Ethereum-rpc-mcp – Let Cursor Talk to Any EVM Chain

The Ethereum RPC MPC Server is a TypeScript project that enables AI models to interact with Ethereum blockchain data through standardized JSON-RPC interfaces, acting as a bridge between AI models and EVM blockchains. It can be used with AI assistants like Cursor or Claude, and supports various Ethereum JSON-RPC calls, with plans to extend its functionality to include indexed API access, multi-chain support, and client-side implementation.

Repo

Show HN: FlowSpec – A flexible AI workflow schema for automation

FlowSpec is a lightweight, standardized JSON schema for defining and automating multi-step workflows, allowing for human-readable and machine-interpretable workflow definitions that can be easily shared, visualized, and executed. The schema includes properties such as workflow title and description, step definitions with actions and transitions, and global transitions for default behaviors, and is designed to be modular and extendable to accommodate evolving workflow needs.

Repo