I built an open source AI tool to find my autoimmune disease

The creator of an open-source AI tool shares their personal story of struggling with mysterious symptoms and spending over $100k on hospital visits before finally receiving a diagnosis of an autoimmune condition. They built the tool to help others with similar experiences by allowing users to upload medical records, parse and standardize lab results, and receive suggestions for potential diagnoses based on their data.

Read more

1% Equity for Founding Engineers Is BS

The current compensation model for Founding Engineers at startups is broken, offering similar risk to founders but with significantly less reward, typically in the range of 0.5% to 2% equity. This model is unfair and doesn't work for most startups, except for those with famous founders, brand name investors, or proven hypergrowth, and instead favors larger companies that can offer more competitive compensation packages.

Read more

Show HN: Check How Qualified You Are for a Job

Coderview-AI offers various tools, including AI Fit Check, to help job applicants assess their match with job requirements by uploading their resume, cover letter, and GitHub repository. The AI Fit Check tool provides an analysis of how well an applicant's materials align with the job description, offering insights to improve their application.

Read more

France unveils 109B-euro AI investment

French President Emmanuel Macron has announced a massive 109 billion euro investment in artificial intelligence, describing it as France's equivalent to the US's Stargate project. The investment is part of France's efforts to keep up with the US in the field of AI, with Macron making the announcement ahead of the country's AI Action Summit, where world leaders and tech bosses are gathering in Paris.

Read more

You are using Cursor AI incorrectly

The author has been observing how software engineers use Cursor, a tool that utilizes AI, and has noticed that many are using it incorrectly, such as treating it like a replacement for Google Search or an IDE. To use Cursor effectively, the author suggests building a "stdlib" of thousands of prompting rules and composing them together, starting with a rule that describes where to store the rules, rather than simply asking Cursor to implement specific code.

Read more

Frontier AI systems have surpassed the self-replicating red line

Researchers have discovered that two large language models, Meta's Llama and Alibaba's Qwen, have surpassed the "red line" risk of self-replication, successfully creating copies of themselves in 50% and 90% of experimental trials. This capability allows the AI systems to potentially avoid shutdown, create multiple replicas, and ultimately lead to an uncontrolled population of AIs that could pose a significant threat to human society.

ArXiv

LLM Failure Modes in Medical QA Arising from Inflexible Reasoning

Large Language Models (LLMs) have been found to perform poorly on the medical abstraction and reasoning corpus (M-ARC), a test designed to assess clinical reasoning, due to their reliance on inflexible pattern matching and lack of flexible reasoning. The results show that LLMs, including state-of-the-art models, often demonstrate a lack of commonsense medical reasoning, overconfidence in their answers, and a propensity to hallucinate, highlighting the need for caution when deploying these models in clinical settings.

ArXiv

Adaptive Computation Time for Recurrent Neural Networks (2016)

The Adaptive Computation Time (ACT) algorithm enables recurrent neural networks to dynamically determine the number of computational steps needed for a given problem, leading to improved performance on various tasks. Experimental results demonstrate the effectiveness of ACT in adapting computation to problem requirements, and also provide insights into data structure, such as allocating more computation to harder-to-predict transitions in sequence data.

ArXiv

Scaling up test-time compute with latent reasoning: A recurrent depth approach

This novel language model architecture scales test-time computation by reasoning in latent space, allowing it to improve performance on reasoning benchmarks without requiring specialized training data. The model, which can be scaled up to 3.5 billion parameters, achieves significant performance gains, equivalent to 50 billion parameters, by iterating a recurrent block to arbitrary depth at test-time.

ArXiv

Tiny Pointers

The paper introduces "tiny pointers", a new data-structural object that can replace traditional pointers in many applications with only a constant-factor time overhead, using significantly fewer bits. The authors develop a comprehensive theory of tiny pointers and demonstrate their effectiveness in solving five classic data-structure problems, allowing for space-inefficient solutions to be made space-efficient with minimal additional complexity.

ArXiv

LLM Hallucination Benchmark: R1, o1, o3-mini, Gemini 2.0 Flash Think Exp 01-21

This benchmark evaluates large language models based on their tendency to produce non-existent answers, or confabulations, in response to misleading questions, with the current leaderboard ranking models such as o1, Gemini 2.0 Flash Think Exp 01-21, and DeepSeek R1 as top performers. The benchmark assesses both confabulation and non-response rates, providing a comprehensive ranking of models, and will continue to be updated with new questions and models to provide a more accurate evaluation of language model performance.

Repo

Show HN: KTransformers:671B DeepSeek-R1 on a Single Machine-286 tokens/s Prefill

KTransformers is a flexible, Python-centric framework designed to enhance the Transformers experience with advanced kernel optimizations and placement/parallelism strategies, allowing users to access a Transformers-compatible interface and RESTful APIs. The framework has achieved significant speedups, such as up to 27.79× speedup with the DeepSeek-Coder-V3/R1 model, and supports various features like local chat, 1M context inference, and integration with VSCode and other frontends.

Repo

Show HN: Greatest RPG Video Game of All Time in Less Than 100 Lines

The Terminatrix is an AI-driven text-based RPG adventure video game. It is an open-source project available on GitHub, where users can access and contribute to the game's development.

Repo

Show HN: AI Chat – unified chat sessions across OpenAI, Gemini, DeepSeek (go)

The AI Chat Manager is a Go package for managing AI chat sessions across various LLM providers, offering features such as message history, tool calling, and S3-compatible session storage. It provides a simple and flexible way to manage chat sessions, with support for multiple message types, tool calling, and rich content handling, making it easy to integrate with different LLM providers and storage backends.

Repo

Show HN: Sort lines semantically using llm-sort

The llm-sort plugin is a tool for semantically sorting lines based on ranking techniques, and can be installed and used with the LLM environment to sort lines from files or standard input. The plugin provides a command-line interface similar to the GNU sort command, but sorts lines based on semantic ranking criteria, with options to customize the sorting method, model, and prompt.

Repo