Bill Atkinson has died

Bill Atkinson, a renowned computer programmer and key figure in Apple's history, passed away on June 5th, 2025, at the age of 74, due to pancreatic cancer. He is survived by his family and leaves behind a legacy of innovative contributions to the field, including the creation of MacPaint, HyperCard, and his influential dithering algorithm, which continue to inspire and influence technology today.

Read more

What was Radiant AI, anyway?

The Elder Scrolls IV: Oblivion's Radiant AI system was a highly ambitious technology that promised to deliver a dynamic and living world with over a thousand fully voice-acted non-player characters, but it was ultimately cut from the final game due to its complexity and limitations. The system's legacy has been shrouded in controversy and misinformation, with many questions remaining about its true capabilities and implementation, prompting a deep dive into its history and development to separate fact from fiction.

Read more

Reverse Engineering Cursor's LLM Client

The authors used TensorZero, an open-source framework, to create a gateway between Cursor and language model providers, allowing them to observe and optimize the API calls being made. By setting up a self-hosted proxy, they were able to experiment with and optimize Cursor for individual users, gaining insights into the prompts and models used, and paving the way for further optimization and evaluation techniques.

Read more

If it works, it's not AI: a commercial look at AI startups (1999)

The thesis "If it works, it's not AI: a commercial look at artificial intelligence startups" by Eve M. Phillips was submitted to the Massachusetts Institute of Technology's Department of Electrical Engineering and Computer Science in 1999. The thesis, advised by Patrick Winston, examines artificial intelligence startups from a commercial perspective and includes bibliographical references.

Read more

Tesla AI VP Milan Kovac Resigns After 9 Years Leading FSD and Optimus Projects

Tesla's VP of AI and Optimus Engineering, Milan Kovac, has announced his departure from the company after 9 years, citing family obligations as the primary reason for his exit. Kovac played a crucial role in developing Tesla's FSD technology and Optimus robot projects, and his resignation marks a significant shift in the company's leadership structure, with Ashok Elluswamy set to assume responsibility for both the FSD program and Optimus robot initiatives.

Read more

Oracular Programming: A Modular Foundation for Building LLM-Enabled Software

Large Language Models (LLMs) can effectively solve tasks with few examples, but their limitations in reliability and modularity hinder their ability to tackle complex problems. Oracular programming is proposed as a solution, allowing domain experts to express problem-solving strategies as programs with choice points that are resolved at runtime by LLMs, enabling more sophisticated and modular applications.

ArXiv

Why is AI hard and Physics simple?

The fields of artificial intelligence and physics are being compared, with the suggestion that physical intuition and theoretical physics approaches can be applied to machine learning. The principle of sparsity is seen as a key connection between the two fields, and a new book on deep learning theory is proposed as a step towards bridging the gap between physics and AI.

ArXiv

Why Understanding Software Cycle Time Is Messy, Not Magic

The analysis of a large dataset from 216 organizations found that factors such as coding time, task scoping, and collaboration patterns have a modest impact on cycle time, a metric measuring the time from ticket creation to completion. However, the effects of these factors are overshadowed by significant unexplained variation, suggesting that improving software delivery velocity requires a systems-level approach rather than focusing on individual performance.

ArXiv

Log-Linear Attention

The attention mechanism in Transformers has limitations due to its quadratic-compute and linear-memory complexity, and while linear attention models can address these issues, they are still restricted by their use of a fixed-size hidden state. Log-linear attention, a new attention mechanism, balances efficiency and expressiveness by replacing the fixed-size hidden state with a logarithmically growing set of hidden states, allowing for log-linear compute cost and matmul-rich parallelization.

ArXiv

Simulating Time with Square-Root Space

A new simulation shows that multitape Turing machines running in time $t$ can be simulated in space $O(\sqrt{t \log t})$, improving upon a 50-year-old result. This simulation has various implications, including more space-efficient evaluation of bounded fan-in circuits and progress on the $P$ versus $PSPACE$ problem, by leveraging a recent space-efficient algorithm for Tree Evaluation.

ArXiv

Show HN: Cpdown – Copy any webpage/YouTube subtitle as clean Markdown(LLM-ready)

Cpdown is a browser extension that allows users to copy the content of any webpage as clean, formatted markdown with one click or keyboard shortcut. The extension also supports copying YouTube subtitles as markdown, and offers various configuration options, including the use of Defuddle or Mozilla's Readability to extract the main content and remove unnecessary HTML elements.

Repo

Show HN: Resumable Web Streams

The resumable-stream library allows clients to resume streams of strings after a lost connection or for a second client to follow along, using a pubsub mechanism with Redis to minimize latency impact. The library provides an idempotent API for creating and resuming streams, with options for explicit resumption and support for different Redis clients, such as ioredis.

Repo

Show HN: Wicketkeeper - A self-hosted, privacy-friendly proof-of-work captcha

Wicketkeeper is a privacy-friendly, proof-of-work captcha system that protects web forms from bots without requiring users to solve frustrating puzzles, instead issuing a small computational challenge that is easy for modern devices to solve. The system consists of a Go backend, a JavaScript client widget, and a full-stack demo application, and works by having the client widget request a challenge from the Wicketkeeper server, solving the challenge in-browser, and then verifying the solution with the server upon form submission.

Repo

Show HN: qc-ai – Quick Config for Neovim with OpenAI

qc-ai is a Neovim plugin that allows users to configure their editor using natural language prompts, such as changing themes or formatting files. The plugin uses the OpenAI API to generate configurations based on user input, and features include caching responses, remembering previous conversations, and configurable model settings.

Repo

Dear Coders and AI Users

Lumen is a command-line tool that automatically scans, structures, and formats a codebase into a prompt suitable for large language models (LLMs), eliminating the need for manual context generation and allowing for more effective AI interactions. With features such as effortless context generation, optimized performance, and intelligent file handling, Lumen aims to unlock the full potential of AI in software development by providing a deep, structured understanding of the codebase.

Repo