Pao Ramen
-
Figma's $300k Daily AWS Bill Isn't the Scandal You Think It Is
The article argues that Figma's $300k daily AWS bill is not a scandal, but rather a normal expenditure for a large-scale, compute-intensive SaaS platform. It explains that the perceived high cost is a result of misunderstanding AWS contracts, which include discounts and annual increases, and that Figma's spending aligns with industry benchmarks for similar companies. The author emphasizes that Figma's transparency in disclosing these costs is more notable than the cost itself.
Jul 28 ⎯ www.duckbillgroup.com
-
The Big LLM Architecture Comparison
The article provides a comprehensive comparison of recent Large Language Model (LLM) architectures, highlighting features like Multi-Head Latent Attention (MLA) and Mixture-of-Experts (MoE) in DeepSeek-V3, normalization layer placements in OLMo 2 and Gemma 3, and sliding window attention in Gemma 3. It also discusses advancements in models like Mistral Small 3.1, Llama 4, Qwen3, SmolLM3 with NoPE, and Kimi 2, noting the trend towards architectural diversity and efficiency improvements.
Jul 28 ⎯ magazine.sebastianraschka.com
-
Texts as Toys
The article posits that Large Language Models (LLMs) enable a "texts as toys" paradigm, where writing becomes toy-making and reading becomes playing. It argues that engaging with AI playfully is crucial to fully realizing its potential, counteracting negative perceptions by focusing on its ludic qualities. The piece suggests a shift in who will find fluency with text, favoring those amenable to new technological affordances and supervisory engagement over traditional craftsmanship.
Jul 28 ⎯ contraptions.venkateshrao.com
-
How to Keep Your Writing Weird in the Age of AI
The article discusses how AI writing tools, acting as 'timid scribes,' tend to smooth out unique and specific language in favor of general, professional phrasing. It emphasizes that while AI can assist with structure and idea generation, writers must actively preserve their distinctive voice and 'weirdness' by understanding AI's limitations and consciously choosing where to retain or smooth out text for different contexts. This involves a deliberate process of guiding AI with style guides and custom instructions to maintain character and impact.
Jul 28 ⎯ every.to
-
AlphaGo Moment for Model Architecture Discovery
This paper introduces ASI-Arch, the first Artificial Superintelligence for AI research (ASI4AI) system focused on neural architecture discovery. It autonomously hypothesizes, implements, trains, and validates novel architectural concepts, moving beyond traditional Neural Architecture Search (NAS). ASI-Arch discovered 106 innovative, state-of-the-art linear attention architectures, demonstrating emergent design principles that surpass human-designed baselines and establishing a scaling law for computational scientific discovery.
Jul 27 ⎯ arxiv.org
-
Textual - Home
The Textual framework is a Python-based Rapid Application Development tool for building sophisticated user interfaces in terminals or web browsers. It emphasizes ease of use, cross-platform compatibility, and integration with the command line. The documentation provides resources for getting started, tutorials, guides on various features like CSS and widgets, API references, and examples of applications built with Textual.
Jul 26 ⎯ textual.textualize.io
-
The Right Thing and the Right Reason:
The article explores the concept of "doing the right thing for the right reason," differentiating between mechanical compliance and ethical responsibility. It argues that "the right thing" is often an "undecidable" – a situation that cannot be fully resolved by logic or rules, requiring personal judgment and ownership of reasoning rather than blind adherence to instructions. The text emphasizes that responsibility cannot be outsourced to systems or phrases, and true ethical action involves accepting the inherent uncertainty and complexity of life, much like using a ladder that must be discarded after serving its purpose.
Jul 26 ⎯ harishsnotebook.wordpress.com
-
Becoming a High Taste Tester
The article discusses the concept of a "high taste tester" (HTT) in the context of frontier models, defining them as influencers who creatively test models in unique ways. It provides a curated list of examples and resources, including links to various blog posts, tests, and discussions that illustrate different approaches to evaluating AI capabilities, from image generation to legislative document analysis and code generation.
Jul 26 ⎯ www.swyx.io
-
Copyright your faults
The article explores the concept of "copyrighting your faults," advocating for embracing personal quirks and constraints as unique strengths. It uses examples like Dan Carlin's loud voice and the "Disco" DVD burning app's smoke simulation to illustrate how limitations can be turned into distinctive features. The author connects this to the Japanese concept of ikigai, suggesting that identifying and leveraging what one "can't not do" is crucial for personal and professional development.
Jul 26 ⎯ interconnected.org
-
Tattoy
Tattoy is a text-based compositor for modern terminals that enhances the user experience with visual effects rendered using UTF8 half-blocks and GPU shaders. It supports Shader Toy and Ghostty shaders, animated cursors, a second terminal in the background for visualizations or monitors, a live-updating scrollback minimap, and an intelligent auto text contrast feature that adjusts foreground colors for readability. Tattoy also allows for plugins written in any language to interact with terminal content.
Jul 26 ⎯ tattoy.sh
-
Never write your own Date Parsing Library—zachleat.com
The article strongly advises against creating custom date parsing libraries, a mistake the author made with Eleventy by adopting Luxon. Luxon's large size and lack of tree-shaking prompted a search for alternatives, leading to the development of `@11ty/parse-date-strings`, a new RFC 9557 compatible library. This change reduces Eleventy's bundle size and prepares it for native Temporal API support.
Jul 26 ⎯ www.zachleat.com
-
An Holistic Framework for Shared Design Leadership
This article proposes a holistic framework for shared design leadership, likening a design team to a living organism. It divides leadership responsibilities between a Design Manager (tending to the 'mind' or psychological safety and growth) and a Lead Designer (tending to the 'body' or craft and execution). Both roles collaborate on the 'circulatory system' of strategy and flow, emphasizing communication, clear feedback loops, and mutual respect to ensure overall team health and effective output.
Jul 25 ⎯ alistapart.com
-
To See the World like Humans
This paper introduces the Turing Eye Test (TET), a challenging perception-oriented benchmark with four diagnostic tasks designed to evaluate how well Multimodal Large Language Models (MLLMs) perceive the world compared to humans. State-of-the-art MLLMs exhibit significant failures on these tasks, indicating fundamental limitations in their visual attention mechanisms and generalization capabilities, which cannot be resolved by language-focused training alone.
Jul 25 ⎯ turingeyetest.github.io
-
Generating Image Placeholders on Cloudflare Workers | Jeremy Morrell
This article explores generating low-quality image placeholders (LQIP) using Cloudflare Workers. It details methods like dominant color extraction, palette-based dominant color, Blurhash, and CSS Blobhash, leveraging Cloudflare's Image bindings. The author found dominan_color_from_palette to offer the best balance for social media feeds.
Jul 24 ⎯ jeremymorrell.dev
-
How to Build an Agent
This article explains how to build a functional code-editing agent using a large language model (LLM), an API key, and a simple loop. It details the process of setting up a Go project, integrating with the Anthropic API, and implementing tools like `read_file`, `list_files`, and `edit_file`. The guide emphasizes that creating such an agent is straightforward, requiring minimal code and practical engineering rather than moments of genius, and highlights the power of LLMs in interacting with external systems through tools.
Jul 24 ⎯ ampcode.com
-
Crawling a billion web pages in just over 24 hours
This article details the process and challenges of crawling one billion web pages within a 24-hour period and a budget of a few hundred dollars in 2025. It explores architectural designs, hardware choices (emphasizing optimized nodes with Redis, fetchers, and parsers), and identifies parsing as a significant bottleneck, partly due to increased page sizes and SSL handshake computations impacting fetching. The author discusses practical learnings, including the effectiveness of the `selectolax` library for parsing and the need for managing memory for large domain frontiers, while also considering future directions like dynamic page rendering for crawls.
Jul 24 ⎯ andrewkchan.dev
-
Hacking the JavaScript Lottery - Independent Security Evaluators
This article explains how to exploit the XorShift128+ pseudo-random number generator used in JavaScript by browsers like Chrome and Firefox. By using symbolic execution with Z3, the author demonstrates how to recover the generator's state from a few of its outputs, enabling prediction of future numbers. This technique was applied to the Los Angeles Times' Powerball simulator, allowing the author to predict the winning lottery numbers.
Jul 24 ⎯ blog.securityevaluators.com
-
Tavily - The Web Access Layer for AI Agents
Tavily provides a web access layer for AI agents, offering an API for real-time, accurate search results tailored for LLMs and RAG workflows. It aims to reduce hallucinations and enhance decision-making by providing concise, relevant data from multiple sources with citations. Tavily is trusted by over 600k developers and integrates with tools like LangChain and LlamaIndex.
Jul 24 ⎯ www.tavily.com
-
Context7 - Up-to-date documentation for LLMs and AI code editors
Generate context with up-to-date documentation for LLMs and AI code editors
Jul 24 ⎯ context7.com
-
GitHub - janwilmake/flaredream: Wrangler Compatible Cloudflare Deployment API
This repository introduces Flaredream Deploy, a free API version of `wrangler deploy` that runs directly on a worker. It allows for deployment of Cloudflare Workers with key features like support for Service Workers, ES Modules, Durable Objects, and parsing of `wrangler.toml`. However, it currently has limitations regarding building/bundling, environments, environment variables, and route patterns.
Jul 24 ⎯ github.com