Pao Ramen
-
The World in Which We Live Now
The article analyzes the current global landscape, highlighting increasing concentration of wealth and power driven by hyper-connectivity, leading to phenomena like technofeudalism. It discusses the misinterpretation of historical dynamics, particularly the rapid rise of China versus declining Western economies, due to compounding growth differences and inefficient spending. Economic saturation in developed nations, coupled with high debt burdens, makes growth essential but difficult to achieve, exacerbated by policies that may reduce GDP. The piece also touches on the societal reliance on immigration for labor to maintain Western lifestyles and the liberating effect of social media in information dissemination, contrasting it with the growing, intrusive role of government and the importance of scale in effective governance.
Sep 15 ⎯ nntaleb.medium.com
-
Chinese students are using AI to beat AI detectors
Chinese universities are increasingly implementing AI detection tools for student theses, leading many students to use AI-powered services to bypass these checks. This has created a market for tools that either rewrite text to avoid detection or have the potential to manipulate the detection systems themselves. Students report issues with false positives and unreliable detection, leading some to alter their work drastically, while the academic community grapples with the implications of widespread AI use in education.
Sep 14 ⎯ restofworld.org
-
Setsum - order agnostic, additive, subtractive checksum - blag
Setsum is an order-agnostic, commutative checksum developed by Robert Escriva at Dropbox. It allows for additive and subtractive operations on data, making it useful for verifying consistency between database replicas or distributed systems. Unlike Merkle trees, Setsum is stateful and can be computed incrementally, with operations costing O(len(msg)) time. Each Setsum consists of 8 columns, each storing a sum of hash chunks modulo a large prime, significantly reducing collision probability.
Sep 14 ⎯ avi.im
-
The Case Against Social Media is Stronger Than You Think
This article critiques the argument that social media's impact on political polarization is overstated, primarily by focusing on its purported effects on political discourse and behavior rather than solely on affective polarization. It introduces an 'elite radicalization theory,' suggesting that social media amplifies emotionally extreme content and empowers a class of political influencers who shape public perception and offline political actions, including protests and hate crimes. The author contends that while direct links to polarization are complex, the amplification of negative content and the rise of extremist political behavior are significant concerns.
Sep 14 ⎯ arachnemag.substack.com
-
Magical systems thinking
The article argues that "systems thinking," while aiming to design functional complex systems, often fails because it overlooks that systems inherently resist change. It highlights historical examples of government system failures and contrasts them with successes derived from starting with simple, working systems and iterating. The piece advocates for this approach, citing examples like the US ICBM program, Operation Warp Speed, and Estonia's e-government, particularly in contrast to the often ineffective attempts to fix existing, overly complex systems.
Sep 14 ⎯ worksinprogress.co
-
The Last Days Of Social Media
Social media, once promising connection, has devolved into exhaustion due to algorithmic prioritization of low-quality, AI-generated, and clickbait content. This shift has eroded authenticity, leading to a "bot-girl economy" and a decline in genuine engagement. As a result, users are migrating to smaller, more private online spaces, signaling the potential end of mass social media as we know it and hinting at a future web focused on intention, community, and human connection.
Sep 14 ⎯ www.noemamag.com
-
GitHub - sindresorhus/type-fest: A collection of essential TypeScript types
This repository, type-fest, provides a comprehensive collection of essential TypeScript types, offering solutions for common and complex typing needs. It includes categories like Basic, Utilities, JSON, Async, String, Array, Numeric, Change case, and Miscellaneous, along with improved built-in types and guidance on extending existing ones. The project aims to enhance developer productivity by offering reusable and robust type definitions.
Sep 13 ⎯ github.com
-
The crawl-to-click gap: Cloudflare data on AI bots, training, and referrals
The article analyzes the shift in internet traffic from search engine referrals to AI-driven crawling, noting that AI training consumes vast amounts of data while providing fewer user referrals. This trend is impacting content creators by reducing traffic, ad revenue, and subscription opportunities. Key observations include the surge in AI bot activity for training purposes and the correlating drop in Google referrals to news sites, particularly after the rollout of AI Overviews. It also highlights the changing market share of AI crawlers like GPTBot and ClaudeBot, and the significant imbalance between crawling volume and referral traffic for services like Anthropic.
Sep 12 ⎯ share.google
-
GitHub - rednote-hilab/dots.ocr: Multilingual Document Layout Parsing in a Single Vision-Language Model
dots.ocr is a multilingual document parser utilizing a single vision-language model for layout detection and content recognition, achieving state-of-the-art performance with a compact 1.7B-parameter LLM. It demonstrates strong multilingual capabilities and a streamlined architecture, outperforming many larger models on benchmarks for tasks like text, table, and reading order recognition.
Sep 12 ⎯ share.google
-
ChatGPT Memory and the Bitter Lesson
This article reverse-engineers ChatGPT's memory system, detailing four components: Interaction Metadata, Recent Conversation Content, Model Set Context, and User Knowledge Memories. It analyzes how these are used, contrasts OpenAI's approach of including all data with a 'bitter lesson' about betting on powerful models over complex retrieval systems, and discusses future challenges in memory accuracy and user profile generation.
Sep 12 ⎯ www.shloked.com
-
Generative and Malleable User Interfaces with Generative and Evolving Task-Driven Data Model | Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems
This paper introduces a novel approach for creating generative and malleable user interfaces by leveraging AI to interpret user prompts and generate dynamic task-driven data models. These models form the foundation for UI generation, allowing for continuous adaptation and customization through natural language and direct manipulation. The proposed system, Jelly, was evaluated technically and through user studies, demonstrating its feasibility and effectiveness in enabling personalized and evolving information spaces.
Sep 12 ⎯ dl.acm.org
-
You Don't Need Animations
Animations can enhance user experience by making interfaces predictable, faster, and enjoyable, but they can also have negative effects if poorly implemented. Animations should have a clear purpose, such as explaining features or providing feedback, and their frequency of use by the user must be considered, with less frequent use allowing for more expressive animations. The perception of speed is crucial; animations should generally be under 300ms to maintain responsiveness and improve perceived performance, with the ultimate goal being to create great user interfaces, sometimes meaning no animation at all.
Sep 11 ⎯ emilkowal.ski
-
RSL
RSL (Really Simple Licensing) is an open content licensing standard designed for the AI-first Internet, enabling publishers to define machine-readable licensing terms for their content. It facilitates compensation for AI training and inference, supports attribution, and allows for secure licensing of non-public content. By integrating with protocols like RSS and Schema.org, RSL aims to create a sustainable ecosystem for content creators and AI companies.
Sep 11 ⎯ rslstandard.org
-
Defeating Nondeterminism in LLM Inference
Nondeterminism in large language model (LLM) inference stems from batch-invariant kernels that alter reduction orders based on batch size and concurrency. Common hypotheses attribute this to floating-point non-associativity and concurrency, but the primary culprit is the varying batch size, which impacts kernel execution. Achieving determinism requires implementing batch-invariant kernels for operations like RMSNorm, matrix multiplication, and attention, ensuring consistent reduction orders regardless of system load or request slicing.
Sep 11 ⎯ thinkingmachines.ai
-
An Interactive Guide to TanStack DB | Frontend at Scale
This article introduces TanStack DB, a reactive client store designed to enhance TanStack Query by addressing its limitations in data relationships and optimistic updates. It details TanStack DB's core features: Collections for typed data sets, Live Queries for fast, reactive data filtering and transformation, and Transactional Mutations for optimistic updates with less boilerplate. The guide also explores integrating TanStack DB with sync engines like ElectricSQL for real-time data synchronization, offering a path towards more robust and efficient client-side data management.
Sep 10 ⎯ frontendatscale.com
-
Real-Time Detection of Hallucinated Entities in Long-Form Generation
This paper introduces a scalable, real-time method for detecting entity-level hallucinations in long-form text generated by large language models. The approach focuses on fabricated entities like names and dates, enabling token-level detection. It utilizes an annotation dataset created via web search to train efficient classifiers that outperform existing methods, even demonstrating generalization to mathematical reasoning tasks.
Sep 10 ⎯ arxiv.org
-
Paper page - Reverse-Engineered Reasoning for Open-Ended Generation
The paper introduces REER (Reverse-Engineered Reasoning), a new paradigm for deep reasoning that tackles the limitations of traditional methods in open-ended generation. By computationally discovering latent reasoning processes from known-good solutions, REER aims to improve model performance on creative tasks. The research also presents DeepWriting-20K, a dataset of reasoning trajectories, and DeepWriter-8B, a model trained on this data that achieves competitive results with leading proprietary models.
Sep 10 ⎯ huggingface.co
-
Paper page - WebExplorer: Explore and Evolve for Training Long-Horizon Web Agents
WebExplorer is a data-driven approach for training long-horizon web agents, achieving state-of-the-art performance in information-seeking tasks. It addresses the scarcity of challenging data by systematically generating data through model-based exploration and iterative query evolution. The developed model, WebExplorer-8B, demonstrates strong capabilities in multi-step reasoning, complex web navigation, and long-context problem-solving, outperforming larger models on various benchmarks.
Sep 10 ⎯ huggingface.co