A team led by Carter Leffer used OpenAI's GPT-6 Astra to break a 75-year-old Enigma message, MVUEH, which had resisted solution since 2005. The AI achieved this feat on its own after analyzing the message and using a crib technique, revealing the correct key and plaintext.
TypeSafe's Jev has taken the AI world by storm with its fast and accurate classification capabilities, but OpenAI may replicate it soon. TypeSafe's training data and reinforcement learning process are their main moat to protect themselves from being copied.
TypeSafe Jev is a model for answering typed questions about text or JSON, used for bounded decisions such as routing, ranking, or selecting an agent's next action. It provides curated starting points and creator demos from X, GitHub, and YouTube.
Install Drop Run agents with --dangerously-skip-permissions to create isolated environments for running programs, containing damage within a sandbox. This setup prevents malicious programs from accessing the host kernel and reduces potential exploits.
iOS users are upset about Apple inserting ads for its own products into the Settings app, with no clear way to disable them. Many have taken to social media and Reddit to express their frustration over the persistent notifications.
A compressor like gzip can be used for language modeling without neural networks by scoring continuations based on their compressed length. Gzip's DEFLATE algorithm uses a sliding window to find matches and compresses them, allowing it to predict the next bytes in a text prompt.
Introducing the MiMo-V2.6 series: frontier intelligence, all the modalities, built in public.
They always look nice at first sight, but over time, they look all the same: complex, unnatural, and busy. Same as DALL-E images. And if it’s read by a human, it should also be written by a human. Nothing worse than having to read through a stack of elaborated text that has been generated on a small prompt. Naval elaborated it best at the The Differences between Writing and Coding Based on ...
The paper introduces N-Operator Correlators (NOC) as a generalized OTOC to better measure quantum information spreading. It demonstrates NOC using toy systems and highlights its advantages over traditional two-operator methods.
Google's SynthID embeds secret signals into images, audio, and video to track users without consent. This technology is part of a growing trend of "spymarks" that can be used to monitor and control individuals' online activities.
The user attended Muni Heritage Weekend 2026 in San Francisco, capturing tram photos with a line scan camera. They shared interesting facts about historic trams like Muni No. 1 and Milan tram, despite some photo failures due to technical issues or background distractions.
The Tetris effect shows how focused attention shapes thoughts, but modern technology often takes control of what we see and hear. Intentional internet browsing is still available, requiring a slower pace to find quality content.
Transformer is a neural network architecture that has revolutionized AI, used in text-generative models like GPT and Llama, and applied to various domains such as audio generation and image recognition. It operates on next-token prediction, using self-attention mechanism to process entire sequences and capture long-range dependencies.
The author discovered a feature in macOS 15.3 that phoned home with personal data and was unable to turn it off, highlighting the lack of consent in AI technology. The author expresses frustration with Apple's handling of their concerns and criticizes the AI industry for prioritizing profit over user privacy.
Meta's AI assistant Muse has a zero-day vulnerability allowing local apps and terminal commands to control the agent, raising serious doubts about its security. Amazon has blocked Muse from its site amid concerns over user data privacy and security after a macOS security expert discovered the flaw.
The Oxide team is gathering in Emeryville for their annual meetup, wearing t-shirts that pay homage to past computer companies like Sun. The nostalgia for Sun has sparked a discussion about what the company got right and wrong, with lessons learned from both serving as inspiration and warning for Oxide's own approach.
Play solitaire with the rest of the internet.
Semfont is a library that automatically sets typography for easier reading, analyzing text with four scores per word: valence, salience, surprise and certainty. It uses heuristics to improve readability without perfect classification or slowing down streaming components.
The author has discovered a bug in AMD's RDRAND and RDSEED instructions, which can't generate the number 0 at a given size.
The author expresses frustration with the increasing use of AI to generate writing, citing its inability to convey nuance and context. They argue that this approach can lead to unreadable documents and a loss of human connection in communication.
The industry is transforming, but relying on AI for coding can lead to unmaintainable code and a loss of human expertise. Experts rely on intuition built from experience, which cannot be replicated by AI or rigid rules.
Please enable JS and disable any ad blocker
MiMo-V2.6-Pro is a high-intelligence model with fast processing and multimodal capabilities, scoring 46 on the Artificial Analysis Intelligence Index, above average among open weight models of similar size. It costs $0.43 per 1M input tokens and $0.87 per 1M output tokens, making it moderately priced for its performance level.
Research found over 93,000 brand appearances during the largest FIFA World Cup, with unhealthy food and drink brands dominating. The tournament's sponsorship heavily promoted products linked to health harms, raising concerns about exposure to these products among billions of viewers worldwide.
Contractors hired to improve AI models have been fired for using AI to train the models, despite being told not to. This irony highlights a problem where some contractors are using AI-generated responses to train OpenAI's models, potentially leading to "model collapse."
The author memorized 100 digits of pi using a person-action-object (PAO) system, which associates each digit with a unique image. This process converted memorization into engineering, allowing for efficient recall and minimal errors.
The author optimized Linear's Continuous Integration (CI) process, reducing PR wait time and runner time by 34% and 52%, respectively. These changes improved the overall performance of CI, allowing for faster code shipping and reduced infrastructure costs.
A reformed former Amazon UK finance manager and the author argue that rent extraction, not profit, built the modern world. They propose composting capital back into sustainable wealth-dependent systems to address a world approaching Total Rent.
Verda, a European AI startup, raised $189 million in funding to build its full-stack infrastructure for compute capacity. The company aims to meet the growing demand for AI workloads with faster provisioning, platform updates and enterprise readiness features.