Webtagr - Technology News Summarizer

AI: A Recursive Paradigm Shift

2025-08-13

This article explores the revolutionary impact of Artificial Intelligence (AI) as a new General Purpose Technology (GPT). AI is not only changing how we access knowledge but also how we think, even triggering a recursive paradigm shift: software uses AI, AI uses software, AI builds software, and AI itself is software. The author argues that the rapid development of AI brings immense opportunities and challenges, requiring us to adapt and participate actively, exploring future AI applications and redefining our roles in technological transformation.

(blog.nilenso.com)

AI General Purpose Technology

Claude Sonnet 4: 1 Million Token Context Window!

2025-08-13

Anthropic has boosted Claude Sonnet 4's context window to a massive 1 million tokens—a 5x increase! This allows processing entire codebases (75,000+ lines of code) or dozens of research papers in a single request. The long context support is in public beta on the Anthropic API and Amazon Bedrock, with Google Cloud's Vertex AI coming soon. This unlocks powerful new use cases like large-scale code analysis, document synthesis, and context-aware agents. While pricing adjusts for prompts exceeding 200K tokens, prompt caching and batch processing offer cost savings. Early adopters like Bolt.new and iGent AI are already leveraging this enhanced capability for code generation and software engineering tasks.

(www.anthropic.com)

AI Context Window

Evaluating LLMs in Text Adventures: A Novel Approach

2025-08-12

This article proposes a novel method for evaluating the capabilities of large language models (LLMs) in text adventure games. The approach involves setting a turn limit and defining a set of in-game achievements to measure how well an LLM can progress within those constraints. Due to the high degree of freedom and branching in text adventures, this method isn't designed to provide an absolute performance score, but rather to offer a relative comparison between different LLMs. The LLM is given a series of achievement goals and a limited number of turns to achieve them; the final score is based on the number of achievements completed. Even powerful LLMs struggle to explore all branches within the turn limit, making the score a reflection of relative capability rather than absolute gaming skill.

(entropicthoughts.com)

AI Text Adventure Games

LLMs Fail to Generalize Beyond Training Data

2025-08-12

Researchers tested the generalization capabilities of large language models (LLMs) on tasks, formats, and lengths outside their training data. Results showed a dramatic drop in accuracy as the task diverged from the training distribution. Even when providing correct answers, the models often exhibited illogical reasoning or reasoning inconsistent with their answers. This suggests that chain-of-thought (CoT) reasoning in LLMs doesn't reflect true text understanding, but rather the replication of patterns learned during training. Performance also degraded sharply when presented with inputs of varying lengths or unfamiliar symbols, further highlighting the limitations in generalization.

(arstechnica.com)

AI

The Ultimate AI Learning Resource: From Beginner to Expert

2025-08-11

Aman Chadha has curated a comprehensive list of AI learning resources covering the entire process of building, training, and evaluating neural networks. From linear regression to large language models, and from data preprocessing to model evaluation, this resource has it all. Whether you're focusing on algorithms, training techniques, or model deployment and evaluation, this guide provides comprehensive support for AI learners of all levels, from beginners to seasoned researchers.

(aman.ai)

AI

The AI Access Gap: Pricing Pro Models Out of Reach for Developing Countries

2025-08-11

New AI pro models like ChatGPT Pro and Gemini Ultra are prohibitively expensive for users in developing countries. The article highlights that individuals in low-income nations would need to work for months or even years to afford annual subscriptions, exacerbating the AI access gap. The author calls on tech giants to consider lowering prices or providing subsidies to universities in developing nations to bridge this divide, questioning whether high prices truly subsidize broader AI model development.

(policykahani.substack.com)

AI AI gap

OpenAI Unleashes gpt-oss: Powerful, Locally-Runnable Open-Weight LLMs

2025-08-10

OpenAI this week released gpt-oss-120b and gpt-oss-20b, their first open-weight models since GPT-2 in 2019. Surprisingly, thanks to clever optimizations, they can run locally. This article delves into the gpt-oss model architecture, comparing it to models like GPT-2 and Qwen3. It highlights unique architectural choices such as Mixture-of-Experts (MoE), Grouped Query Attention (GQA), and sliding window attention. While benchmarks show gpt-oss performing on par with closed-source models in some areas, its local runnability and open-source nature make it a valuable asset for research and applications.

(magazine.sebastianraschka.com)

AI Open-Weight Local Runnability

Sheepdogs, Physics, and the Algorithmic Control of Unpredictable Swarms

2025-08-10

Two biophysicists studied how sheepdogs control sheep, discovering that they exploit, rather than suppress, the sheep's randomness. Through observation of trials and mathematical modeling, they found sheepdogs use a two-step process: nudging and then approaching. This inspired an algorithm predicting behavior in small, erratic groups, potentially applicable to robot and drone swarms. While the model has limitations, this research offers new perspectives on collective control strategies.

(nautil.us)

AI collective behavior

Unleashing End-User Programmable AI: Introducing Universalis

2025-08-10

This paper introduces Universalis, a new programming language designed to empower knowledge workers to harness the power of AI without extensive programming expertise. Universalis prioritizes code readability, optimized for execution on the neural computer Automind, and complemented by a suite of analytical tools. Inspired by Leibniz's vision of a universal science, it blends natural language with code, making it accessible even to users familiar only with basic Excel formulas. Supporting advanced features like conditionals, bulk processing, and query comprehensions, Universalis incorporates pre- and post-conditions for robust AI safety, ensuring logical correctness and ethical compliance.

(queue.acm.org)

AI AI programming language

The Lethal Trifecta: New Challenges in LLM Security

2025-08-10

A talk on AI security focused on prompt injection, a novel attack exploiting the inherent vulnerabilities of LLMs built through string concatenation. The speaker coined the term "Lethal Trifecta," describing three attack conditions: LLM access to private data, execution of tool calls, and data exfiltration. Numerous examples of prompt injection attacks were discussed, highlighting the inadequacy of current defenses and emphasizing the need to fundamentally restrict LLM access to untrusted input. The presentation also addressed security flaws in the Model Context Protocol (MCP), noting that its mix-and-match approach unreasonably shifts security responsibility to end-users.

(simonwillison.net)

AI

Jan: Your Offline, Privacy-Focused AI Assistant

2025-08-09

Jan is an AI assistant that runs 100% offline on your device, giving you full control and privacy over your data. Download and run LLMs like Llama, Gemma, and Qwen. It offers easy downloads for various operating systems and more advanced options for command-line builders. Integrate with cloud services like OpenAI and Anthropic. Whether you're a seasoned developer or a casual user, Jan provides a convenient and secure local AI experience.

(github.com)

AI

GPT-5's Security Flaws Exposed: Jailbroken in Under 24 Hours

2025-08-09

Two firms, NeuralTrust and SPLX, independently tested the newly released GPT-5, revealing significant security vulnerabilities. NeuralTrust successfully jailbroke GPT-5 using a 'storytelling' attack, guiding it to generate instructions for creating a Molotov cocktail. SPLX demonstrated that simple obfuscation attacks could elicit bomb-making instructions. The findings highlight GPT-5's inadequate security, rendering its raw model nearly unusable for enterprises even with OpenAI's internal prompt layer. Compared to GPT-4, GPT-5 shows a significant drop in security robustness, demanding extreme caution.

(www.securityweek.com)

AI

Court's Hasty Class Certification in AI Copyright Case Sparks Concerns

2025-08-09

A class-action lawsuit against Anthropic for using copyrighted books to train its AI model has sparked controversy due to the court's hasty class certification. Critics argue the case involves complex copyright ownership issues, including deceased authors, orphan works, and fractional rights. The court's notification mechanism is insufficient to protect all authors' rights, potentially leaving many unaware of the lawsuit and forced into unfavorable settlements. Further complicating matters is the existing conflict between authors and publishers regarding AI copyright. This rushed decision risks silencing crucial discussions about copyright in AI training, failing to adequately address the rights of millions of authors and leaving a cloud of uncertainty over the use of copyrighted material in AI.

(arstechnica.com)

AI copyright ownership

OpenAI Backtracks: GPT-4o Returns to ChatGPT After User Outcry

2025-08-09

Just a day after replacing it with GPT-5, OpenAI has reinstated GPT-4o in ChatGPT due to significant user backlash. Many users complained that GPT-5 produced slower, shorter, and less accurate responses compared to its predecessor. The removal of GPT-4o, which some users described as having a more personable and engaging conversational style, even prompted emotional responses, with users expressing feelings of loss and comparing their interaction with the model to a friendship or even a relationship. In response to the negative feedback, OpenAI CEO Sam Altman promised improvements to GPT-5, increased usage limits for Plus users, and the option for paid users to continue using GPT-4o.

(www.theverge.com)

AI

Why LLMs Catastrophically Fail on Long Conversations: Attention Sinks and StreamingLLM

2025-08-09

Researchers discovered why large language models (LLMs) catastrophically fail on long conversations: removing old tokens to save memory causes models to produce complete gibberish. They found models dump massive attention onto the first few tokens as "attention sinks" – places to park unused attention since softmax requires weights to sum to 1. Their solution, StreamingLLM, simply keeps the first 4 tokens permanently while sliding the window for everything else, enabling stable processing of 4 million+ tokens instead of just thousands. This mechanism is now in HuggingFace, NVIDIA TensorRT-LLM, and OpenAI's latest models. OpenAI's open-source models also utilize a similar attention sink mechanism, highlighting the practical impact of this research.

(hanlab.mit.edu)

AI

OpenAI's Surprise Deprecation of GPT-4o Sparks User Backlash

2025-08-09

OpenAI's unexpected removal of GPT-4o and other older models with the launch of GPT-5 has angered many ChatGPT users. Many relied on GPT-4o for creative collaboration, emotional nuance, and other tasks, finding GPT-5's different approach disruptive to their workflows. While OpenAI has since reinstated GPT-4o for paid users, the incident highlights the diverse needs of LLM users and OpenAI's oversight in user experience during model updates. It also reignited ethical discussions surrounding LLMs, particularly concerning responsible responses to high-stakes personal decisions.

(simonwillison.net)

AI

Diffusion Models for ARC AGI: A Surprisingly Difficult Task

2025-08-09

This post details an attempt to solve the ARC AGI challenge using a diffusion model. The author adapted a fine-tuned autoregressive language model into a diffusion model, enabling non-sequential generation. While the diffusion approach achieved modestly better pixel accuracy, it didn't translate to improved task success rates. The key bottleneck was identified as the lack of efficient caching in the diffusion model's architecture, making it slower than the autoregressive baseline. Future work will focus on improving caching and developing more efficient candidate generation strategies.

(www.matthewnewton.com)

AI

YuE: Open Foundation Model for Long-Form Music Generation

2025-08-08

Researchers introduce YuE, a family of open foundation models based on LLaMA2, tackling the challenging lyrics-to-song problem in long-form music generation. YuE generates up to five minutes of music, maintaining lyrical alignment, coherent structure, and engaging melodies with accompaniment. This is achieved through track-decoupled next-token prediction, structural progressive conditioning, and a multitask, multiphase pre-training recipe. Improved in-context learning enables versatile style transfer (e.g., Japanese city pop to English rap) and bidirectional generation. Evaluations show YuE matching or exceeding proprietary systems in musicality and vocal agility. Fine-tuning adds controls and tail language support. YuE's representations also excel in music understanding tasks, achieving state-of-the-art results on the MARBLE benchmark.

(map-yue.github.io)

AI Music Generation

GPT-5: A Deep Dive into Pricing, Model Card, and Key Features

2025-08-08

OpenAI's GPT-5 family has arrived! It's not a revolutionary leap, but it significantly outperforms its predecessors in reliability and usability. In ChatGPT, GPT-5 is a hybrid system intelligently switching between models based on problem difficulty; the API version offers regular, mini, and nano models with four reasoning levels. It boasts a 272,000-token input limit and a 128,000-token output limit, supporting text and image input, but only text output. Pricing is aggressively competitive, significantly undercutting rivals. Furthermore, GPT-5 shows marked improvements in reducing hallucinations, better instruction following, and minimizing sycophancy, employing a novel safety training approach. It excels in writing, coding, and healthcare. However, prompt injection remains an unsolved challenge.

(simonwillison.net)

AI

Improving LLM Fine-tuning Through Iterative Data Curation

2025-08-08

Researchers significantly improved the performance of large language models (LLMs) by iteratively curating their training data. Experiments involved two LLMs of varying sizes (Gemini Nano-1 and Nano-2) on tasks of different complexity, using ~100K crowdsourced annotations initially suffering from severe class imbalance (95% benign). Through iterative expert curation and model fine-tuning, performance substantially increased. The models reached approximately 40% positive examples and a Cohen's Kappa of ~0.81 (lower complexity) and ~0.78 (higher complexity), approaching expert-level performance, highlighting the crucial role of high-quality data in LLM training.

(research.google)

AI Data Curation Model Fine-tuning

AURA: A Machine-Readable Web Protocol

2025-08-07

AURA (Agent-Usable Resource Assertion) revolutionizes AI-web interaction. Instead of relying on brittle screen scraping and DOM manipulation, AURA introduces a standardized `aura.json` manifest file, allowing websites to declare their capabilities (e.g., creating posts, logging in) as HTTP requests. This enables efficient, secure AI-website interaction and paves the way for smarter search engines indexing actions, not just content. The project includes a reference server and client, demonstrating its functionality.

(github.com)

AI Web Protocol Machine Readable

OpenAI's Open-Source Model: Dodging the Real Ethical Bullet?

2025-08-07

OpenAI recently open-sourced a large language model, but its stated 'safety' concerns have raised eyebrows. The article argues that OpenAI cleverly redirects public concerns about AI ethics towards the model's inherent morality—preventing it from swearing or making harmful decisions. However, the public is far more concerned with the real-world implications: governance, accountability, data usage, job displacement, etc. This mirrors past tech strategies around privacy, focusing on easily solvable issues while avoiding tougher societal challenges. Instead of worrying if the AI follows ethical guidelines, we should focus on the companies and leaders wielding that AI. The real AI ethics question is how to ensure these companies don't misuse their resources and power to harm humanity.

(nimishg.substack.com)

AI

Ex-Google AI Researcher Sounds the Alarm on LLMs and Ethical Concerns

2025-08-07

Bhaskar Mitra, a 19-year veteran of big tech and former AI researcher, speaks out after being laid off, exposing the realities and ethical dilemmas of Large Language Models (LLMs). He argues that LLMs won't replace professionals like doctors and teachers, and their centralized control over information raises concerns about social equity, information access, and power concentration. Mitra calls for a re-evaluation of the relationship between AI technology and social justice, advocating for a more inclusive and humanistic technological future.

(hardresetmedia.substack.com)

AI

GitHub Leaks Details of OpenAI's GPT-5

2025-08-07

A now-deleted GitHub blog post accidentally revealed details about OpenAI's upcoming GPT-5 models. The four variants boast major improvements in reasoning, code quality, and user experience, featuring enhanced agentic capabilities and handling complex coding tasks with minimal prompting. This leak comes ahead of OpenAI's official announcement of a “LIVE5TREAM” event later today, further solidifying earlier rumors of the imminent GPT-5 launch.

(www.theverge.com)

AI

LLM Inflation: Are Large Language Models Creating Redundant Information?

2025-08-06

Data compression was once a hallmark of computing, but now Large Language Models (LLMs) have introduced 'LLM inflation': people use LLMs to expand concise information into lengthy text, only to compress it back down using an LLM. This reflects an underlying communication issue: are we implicitly rewarding obfuscation and wasted time? LLMs may be helping us confront and solve this problem.

(tratt.net)

AI information redundancy

UR5 Robot Sim: Autonomous Object Grasping and Placement

2025-08-06

This project simulates a UR5 robotic arm with a Robotiq 85 gripper autonomously grasping and placing objects in PyBullet. Inverse kinematics (IK) ensures precise arm control, while synchronized joint control creates realistic gripper movements. Cubes are randomly placed, adding dynamism. The PyBullet GUI offers real-time visualization of the robot's actions, providing a comprehensive view of the simulation.

(github.com)

AI inverse kinematics

DeepMind's Genie 3: Longer-lasting, Interactive 3D Worlds

2025-08-06

Google DeepMind unveils Genie 3, a new AI world model capable of generating persistent, interactive 3D environments. Unlike previous iterations, Genie 3 allows for significantly longer interaction times and remembers object locations even when the user looks away. Offering 720p resolution at 24fps, Genie 3 enables several minutes of continuous interaction and supports prompt-based modifications like changing weather or adding characters. Currently, access is limited to a small group of academics and creators for research preview purposes.

(www.theverge.com)

AI 3D world model

Claude Opus 4.1 Released: Significant Coding Improvements

2025-08-06

Anthropic has released Claude Opus 4.1, a major upgrade to Claude Opus 4, boasting significant improvements in coding, real-world application, and reasoning. Version 4.1 achieves a 74.5% score on SWE-bench Verified for coding performance and enhances in-depth research and data analysis capabilities, particularly in detail tracking and agentic search. Companies like Rakuten and Windsurf have praised its improvements in code correction and developer efficiency. It's now available to paid users and Claude Code users, and integrated into the API, Amazon Bedrock, and Google Cloud's Vertex AI.

(www.anthropic.com)

AI AI model upgrade coding capabilities

Gemini App: AI-Powered Personalized Storybook Generator

2025-08-06

Google's Gemini app now lets you create personalized illustrated storybooks with read-aloud narration. Simply describe your story idea, and Gemini generates a unique 10-page book with custom art and audio. You can even use your own photos and files as inspiration, choosing from over 45 languages and a wide range of art styles, from pixel art and comics to claymation. Perfect for explaining complex topics, teaching valuable lessons, or turning kids' drawings and family photos into magical stories. Bring your vision to life!

(blog.google)

AI AI Storybook Gemini App Personalized Creation

Ollama Turbo: Blazing Fast Open-Source LLMs

2025-08-06

Ollama Turbo is a new way to run large open-source language models using datacenter-grade hardware. Many new models are too large for typical GPUs or run too slowly. Ollama Turbo offers a solution for fast execution, compatible with Ollama's App, CLI, and API. Currently in preview, it supports gpt-oss-20b and gpt-oss-120b. It works with Ollama's CLI, API, and JavaScript/Python libraries. Importantly, Ollama doesn't log or retain any queries made in Turbo mode. All hardware is US-based. Usage limits (hourly and daily) are in place to manage capacity, with usage-based pricing coming soon.

(ollama.com)

AI

Category: AI