DOJ proposal would require Google to divest from AI partnerships with Anthropic

The US Justice Department is seeking to unwind Google's partnership with artificial intelligence startup Anthropic as part of a proposal to resolve a landmark antitrust case over online search. The proposal would bar Google from acquiring, investing in, or collaborating with any company that controls where consumers search for information, including query-based AI products.

Read more

AI eats the world

Benedict Evans produces an annual presentation on macro and strategic trends in the tech industry, with recent titles including 'AI and Everything Else' for 2024 and 'The New Gatekeepers' for 2023. He also gives presentations for various companies and sends a weekly newsletter to over 150,000 people, analyzing and contextualizing key tech developments.

Read more

How did you do on the AI art Turing test?

A test of 50 images, half created by humans and half by AI, was given to 11,000 people to classify as either human or AI-generated, with the median score being 60%, only slightly above chance. The results showed that people have difficulty distinguishing between human and AI art, with biases towards certain styles and a tendency to misjudge human-generated digital images as AI-generated.

Read more

MIT researchers develop an efficient way to train more reliable AI agents

MIT researchers have developed an efficient algorithm for training more reliable AI agents, particularly for complex tasks that involve variability. The algorithm strategically selects the best tasks for training an AI agent, allowing it to effectively perform all tasks in a collection of related tasks, and has been shown to be between five and 50 times more efficient than standard approaches.

Read more

FLUX.1 Tools

Black Forest Labs has released FLUX.1 Tools, a suite of models that add control and steerability to their base text-to-image model FLUX.1, enabling modification and re-creation of real and generated images. The suite consists of four features: FLUX.1 Fill for inpainting and outpainting, FLUX.1 Depth and FLUX.1 Canny for structural guidance, and FLUX.1 Redux for image variation and restyling.

Read more

DroidSpeak: Enhancing Cross-LLM Communication

DroidSpeak is a framework that accelerates communication between agents in multi-agent systems using Large Language Models by reusing intermediate data, such as input embeddings and key-value caches. This approach achieves up to a 2.78x speedup in prefill latency with minimal loss in accuracy, enabling more efficient and scalable multi-agent systems.

ArXiv

A Critique of Unfounded Skepticism Around AI for Chip Design

A 2020 Nature paper introduced AlphaChip, a deep reinforcement learning method for generating superhuman chip layouts, which has been widely adopted and extended. A non-peer-reviewed paper questioned AlphaChip's performance claims, but its findings are disputed due to methodological differences and a meta-analysis by Igor Markov is also criticized for its lack of transparency.

ArXiv

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

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

ArXiv

Automating LLM Development with LLMs

Large Language Models (LLMs) are rapidly improving, but their advancements are limited by human-designed algorithms. The proposed Self-Developing framework allows LLMs to autonomously generate and learn model-improvement algorithms, resulting in models that surpass human-designed ones and demonstrate strong transferability.

ArXiv

Samurai: Adapting Segment Anything Model for Zero-Shot Visual Tracking

SAMURAI, an enhanced adaptation of the Segment Anything Model 2 (SAM 2), addresses the limitations of SAM 2 in visual object tracking by incorporating temporal motion cues and a motion-aware memory selection mechanism. This results in robust and accurate tracking performance, with significant improvements over existing trackers and competitive results compared to fully supervised methods in benchmark datasets.

ArXiv

Autoflow, a Graph RAG based and conversational knowledge base tool

Here is a 2-sentence summary of the text:

TiDB.AI is an open-source GraphRAG (Knowledge Graph) built on top of TiDB Vector, LlamaIndex, and DSPy, offering features such as conversational search and embeddable JavaScript snippets. The platform is designed to provide a comprehensive search experience and can be deployed using Docker Compose, with contributions from the community welcome under the Apache License, Version 2.0.

Repo

XGrammar: Efficient, Flexible and Portable Structured Generation for LLM

XGrammar is an open-source library for efficient, flexible, and portable structured generation, supporting general context-free grammar and optimized for fast execution. It features a minimal C++ backend for easy integration and is designed to enable zero-overhead structured generation in large language model (LLM) inference.

Repo

Show HN: Pull Request Reviewed by LLM

The eBank backend is a Go-based application that utilizes Docker for containerized services and relies on dependencies such as PostgresDB and Traefik proxy. To run the application, use commands like yarn db, yarn watch, and yarn proto to set up the database, start services, and generate protocol buffers.

Repo

Show HN: Best way to setup a fullstack webapp using AI to get >97% code accuracy

oneShotCodeGen is a command-line tool that generates complete full-stack web applications from a single prompt using AI, addressing issues with existing AI models that struggle to create accurate full-stack web apps. The tool divides the code generation process into distinct steps, forcing the AI to document and save its assumptions and code details, and provides various customization options and features.

Repo

Show HN: Open-Source Pull Request AI Reviewer

Presubmit's AI Code Reviewer is a tool that optimizes the code review process by catching bugs, suggesting improvements, and providing meaningful summaries before human reviewers take a look. It offers instant, in-depth PR analysis, interactive discussions, and seamless integration with GitHub Actions, allowing humans to focus on architecture and complex logic while the AI handles the basics.

Repo