Meta torrented & seeded 81.7 TB dataset containing copyrighted data

Meta allegedly torrented over 81.7 terabytes of pirated books, including 35.7 terabytes from shadow libraries, to train its AI models, with internal emails showing staff expressed concerns about the legality of the practice. The revelation complicates Meta's copyright case, with authors alleging that the company's actions constitute direct copyright infringement by not only downloading but also seeding, or distributing, the pirated content.

Read more

Transformer – Spreadsheet

Tom Yeh, the creator of AI by Hand, has developed a tool using Google Sheets that allows users to create their own AI by Hand exercises with custom numbers and solutions. The goal of this project is to maximize reach and broaden access to AI education, and Tom is seeking feedback on the tool, which can be accessed at https://by-hand.ai/sp/tfmr.

Read more

Apple ordered by UK to create global iCloud encryption backdoor

The UK government has secretly ordered Apple to create a global backdoor to iCloud encryption, allowing UK security officials to access encrypted user data worldwide. Apple is likely to stop offering encrypted storage in the UK rather than comply with the order, which would compromise the company's Advanced Data Protection feature that provides end-to-end encryption for many iCloud data categories.

Read more

The impact of AI on the technical interview process

The traditional technical interview process, which often involves solving coding challenges and system design problems, is becoming outdated as AI technology advances and can now perform these tasks quickly and efficiently. The author suggests that this process is no longer an effective way to measure a candidate's skills and proposes alternative methods, such as code reviews, which can better assess a candidate's ability to communicate, collaborate, and evaluate code.

Read more

Amazon Will Spend Nearly a Year of AWS Revenue on AI Investments

Amazon is expected to spend over $100 billion in capital expenditures in 2025, with the majority going towards AWS, primarily to support growing demand for AI services. The estimated $86 billion in AI datacenter spending would be a significant investment, but Amazon's model suggests that for every dollar spent on AI infrastructure, it can generate nearly 88 cents in annual revenue, making the investment sustainable and potentially highly profitable.

Read more

Test-time scaling new approach: extra test-time compute improves LLM reasoning

Researchers developed a simple approach to test-time scaling for language modeling, using a curated dataset and a technique called budget forcing to control test-time compute, which led to improved performance on math questions. Their model, s1-32B, outperformed OpenAI's o1-preview on competition math questions by up to 27% and achieved further gains when scaled with budget forcing, with the model, data, and code made available open-source.

ArXiv

Meta AI's latest research: improved LLM reasoning with Latent Tokens

Large Language Models (LLMs) trained on chain-of-thought data can excel at reasoning and planning, but processing these lengthy inputs is computationally expensive. This work proposes a hybrid representation using latent discrete tokens to abstract away initial reasoning steps, reducing input length and improving performance in various benchmarks for logical and mathematical reasoning problems.

ArXiv

Robust autonomy emerges from self-play

Self-play in simulation has been used to develop a robust and naturalistic driving policy, with 1.6 billion km of driving experience, which achieves state-of-the-art performance on autonomous driving benchmarks. The policy, trained without human data, outperforms prior models in real-world scenarios and demonstrates unprecedented robustness, averaging 17.5 years of continuous driving between incidents in simulation.

ArXiv

HippoRAG: Neurobiologically Inspired Long-Term Memory for LLMs (2024)

HippoRAG is a novel retrieval framework that integrates large language models, knowledge graphs, and the Personalized PageRank algorithm to mimic human memory, enabling more efficient knowledge integration over new experiences. This approach outperforms existing methods by up to 20% on multi-hop question answering tasks, while also being significantly cheaper and faster, and can tackle new scenarios that are beyond the reach of existing methods.

ArXiv

Gold-Medalist Performance in Solving Olympiad Geometry with AlphaGeometry2

AlphaGeometry2 is an improved version of AlphaGeometry that can solve Olympiad geometry problems at a level surpassing an average gold medalist, with an 88% coverage rate and 84% overall solving rate for geometry problems from the last 25 years. The system's advancements include an extended language, improved search process, and enhanced symbolic engine, and it has been used to achieve a silver-medal standard at the International Math Olympiad 2024.

ArXiv

Show HN: Upsonic: An AI agent framework with client-server architecture

Upsonic is a cutting-edge enterprise-ready framework that enables users to orchestrate LLM calls, agents, and computer use to complete tasks cost-effectively, providing features such as production-ready scalability, task-centric design, and multi-client processing. The framework offers a range of tools and features, including direct LLM calls, response format customization, tool integration, and memory management, making it a versatile and powerful solution for automating complex tasks.

Repo

Show HN: AI YC Partner Agent

This project is a demo LLM app built with Pocket Flow, a minimalist LLM framework, which uses Retrieval-Augmented Generation on curated Y Combinator public materials and cites relevant sources. The app is available to try out and is also the project's application to Y Combinator, showcasing their idea and technology.

Repo

OpenHealth: AI Health Assistant – Powered by Your Data, Running Locally

OpenHealth is a private and locally-run AI health assistant that helps users take charge of their health data by consolidating and parsing their personal health information, and providing personalized interactions with GPT-powered AI. The platform supports various data sources, including blood test results, health checkup data, and wearable devices, and can be run locally using Docker Compose, with optional integration with external APIs like Upstage and OpenAI.

Repo

CowPilot: An AI cowsay for your terminal

CowPilot is an AI-powered cowsay for terminals that can be installed with pip install cowpilot and used to generate cow-themed text responses. The tool also allows for custom cow designs by passing a .cow file to the -c or --cow flag, enabling personalized and varied outputs.

Repo

Show HN: Unify – A single file Python CLI Tool port of repomix

Unify is a single Python file that unifies code repositories into a single text file, with features including no pip dependency and a small size of under 100 lines. To use Unify, download the script, run it with Python, and specify the repository path to generate a unified text file, or create an executable using Pyinstaller for a more streamlined experience.

Repo