LLMs understand nullability

Large language models have made significant progress in writing code, but there are still many unanswered questions about their capabilities, such as how often they can write correct code on their own and whether they truly "understand" the code they are writing. Researchers are now studying the internal processes of these models to better understand their strengths and weaknesses, including their ability to reason about nullability, a fundamental concept in programming that determines when a variable can take on a null value.

Read more

Agenda Behind the Facial Recognition Tech Used by ICE and the FBI Revealed

Clearview AI, a facial recognition company founded by Hoan Ton-That, has compiled a massive biometric database by scraping billions of images from the internet and social media without users' knowledge or consent. The company's technology has been used by law enforcement agencies and other clients to surveil and profile individuals, including immigrants and activists, and has been the subject of numerous lawsuits and fines for its alleged violations of privacy laws.

Read more

Shopify CEO: "AI usage is now a baseline expectation"

Shopify is embracing reflexive AI usage as a baseline expectation, recognizing its potential to revolutionize the way work is done and enable merchants to build successful businesses. The company is making a concerted effort to integrate AI into all aspects of its operations, with employees expected to learn and adapt to using AI effectively, and AI usage being incorporated into performance reviews and peer feedback.

Read more

John Carmack on AI in game programming

The development of AI tooling in programming, art, and design will likely lead to increased productivity and accessibility, allowing smaller teams and new creators to produce high-quality content, while also enabling top talent to reach new heights. The impact on job numbers is uncertain, but the overall effect will be a vast increase in available content, with the potential for creative entrepreneurship to flourish at various scales.

Read more

Deep Learning, Deep Scandal

Major tech companies, including Meta, OpenAI, and Google, have failed to create a "GPT-5" level AI, despite massive investments, with their latest models, such as Llama 4, not meeting expectations. A rumor suggests that Meta may have attempted to cheat to improve Llama 4's performance, which, if true, would be a serious violation of research integrity and follows a pattern of questionable practices in the AI community, including data leakage and manipulation of benchmark tests.

Read more

Beyond Quacking: Deep Integration of Language Models and RAG into DuckDB

FlockMTL is an extension for database management systems that integrates large language models and retrieval-augmented generation to streamline the development of knowledge-intensive analytical applications. It provides features such as model-driven functions, cost-based optimizations, and novel SQL abstractions to simplify the implementation of data pipelines and reduce the effort required to manage heterogeneous data systems and low-level implementation details.

ArXiv

SeedLM: Compressing LLM Weights into Seeds of Pseudo-Random Generators

SeedLM is a novel post-training compression method that uses pseudo-random generators to encode and compress model weights, reducing memory access and speeding up memory-bound tasks. The approach has been shown to achieve better zero-shot accuracy retention and significant speed-ups, particularly in large models like Llama 3 70B, while maintaining performance comparable to FP16 baselines.

ArXiv

Model Context Protocol (MCP): Landscape, Security Threats

The Model Context Protocol (MCP) is a standardized interface that enables AI models to interact with external tools and resources, facilitating interoperability and breaking down data silos. This paper provides an overview of MCP, analyzing its components, security risks, and current landscape, while also exploring future directions and offering recommendations for its secure and sustainable development.

ArXiv

Transformers Are Efficient Compilers, Provably

This paper explores the use of transformer-based large language models as compilers, introducing a programming language called Mini-Husky to investigate their expressive power. The researchers found that transformers can efficiently handle compilation tasks with a logarithmic number of parameters, outperforming recurrent neural networks (RNNs), and validated their theoretical results with empirical comparisons between the two models.

ArXiv

Proof or Bluff? Evaluating LLMs on 2025 USA Math Olympiad

State-of-the-art large language models, such as Gemini-2.5-Pro, have achieved impressive performance on mathematical competitions, but a new evaluation reveals they struggle with rigorous reasoning and proof generation, with most models scoring less than 5% on a set of challenging math problems. The results highlight the need for significant improvements in reasoning and proof generation capabilities, as current models are inadequate for real-world mathematical tasks despite their ability to produce correct numerical answers.

ArXiv

Auto-Sync Your Docs, SDKs and Examples for LLMs and AI Agents

The VideoDB Agent Toolkit is an AI agent toolkit that exposes VideoDB context to LLMs and agents, enabling integration with AI-driven IDEs and chat agents. The toolkit provides context files, including llms-full.txt and llms.txt, which offer comprehensive and lightweight metadata for deep integration and quick discovery, and automates context generation, maintenance, and discoverability through GitHub Actions.

Repo

Show HN: An open-source code scanner to find issues in prompts and LLM calls

Kereva LLM Code Scanner is a static analysis tool designed to identify potential security risks, performance issues, and vulnerabilities in Python codebases that use Large Language Models (LLMs). The tool analyzes code without execution to detect problems such as hallucination triggers, bias potential, and prompt injection vulnerabilities, and provides features like static code analysis, specialized LLM scanners, and flexible reporting.

Repo

AI Policy Update Proposal · servo/servo · Discussion #36379

Servo is a prototype web browser engine written in Rust, currently developed on various 64-bit platforms including macOS, Linux, Windows, OpenHarmony, and Android, and welcomes contributions from everyone. To get started, developers can check out the Servo Book and follow platform-specific instructions to set up their environment, install dependencies, and build the Servo browser engine.

Repo

Node-Llama-cpp – Run AI models locally on your machine with Node.js

Node-llama-cpp is a package that allows users to run AI models locally on their machine, with features such as pre-built binaries, automatic hardware adaptation, and a complete suite of tools for using LLMs in projects. The package provides a simple way to interact with AI models, including a CLI for chatting with models without writing code, and supports various features like JSON output, function calling, and embedding and reranking support.

Repo

Show HN: MCP server wrapper for Google Lighthouse

The Lighthouse MCP Server is a tool that wraps around Google's Lighthouse to measure performance metrics for web pages, allowing users to run comprehensive audits, configure device emulation, and control network throttling. It can be installed and run in various ways, including using npx, global installation, or local development, and provides tools such as running audits and getting performance scores for URLs.

Repo