Wednesday — July 30, 2025
Anthropic introduces weekly limits to its Claude Pro and Claude Max services, KathaaVerse launches an AI-powered platform to turn books into text adventure games, and researchers discover a novel threat model called TrojanStego where language models can leak sensitive information.
News
Stop selling “unlimited”, when you mean “until we change our minds”
Anthropic has added weekly limits to its Claude Pro and Claude Max services, affecting heavy users, in a move that follows a common playbook of AI companies offering "unlimited" access before capping usage. This approach damages trust with power users and early adopters, who drive growth and adoption, and instead, companies should prioritize transparent pricing and let users control their limits.
Show HN: I built an AI that turns any book into a text adventure game
KathaaVerse is an interactive storytelling platform that allows users to transform their favorite books into personalized, text-based adventures, with options to choose from various genres such as fantasy, science fiction, and contemporary fiction. The platform, powered by AI, offers a range of stories, including popular titles like Harry Potter, Alice in Wonderland, and The Hunger Games, where users can direct the plot and create their own unique experience.
Playing with Open Source LLMs
The author, a self-described hermit, explores the latest trend in AI, using command line tools to write code with large language models (LLMs), and discovers open-source alternatives to vendor tools like Claude Code. They settle on using the qwen3:8b model with Ollama and Aider, a pair programming application, to write code and refactor existing projects, finding it to be a useful tool with some limitations and requiring guidance to produce accurate results.
Meta Is Going to Let Job Candidates Use AI During Coding Tests
Here is a summary of the article in a couple of sentences:
Meta is introducing a new type of coding interview that allows job candidates to use an AI assistant, aiming to make the interview process more representative of the developer environment and reduce cheating. The move is part of the company's push towards "vibecoding," where software engineers work alongside AI tools, and is in line with CEO Mark Zuckerberg's vision of humans managing "AI coding agents" that write code.
Delta’s new AI-powered pricing strategy
Delta Air Lines is expanding its use of artificial intelligence to set ticket prices, aiming to cover 20% of its US domestic flights by the end of 2025, with the goal of generating custom fares for each individual customer based on real-time data analysis. The move has sparked backlash from critics who accuse the airline of engaging in "predatory pricing" by using personal data to charge customers the maximum they are willing to pay, although Delta maintains that its AI pricing system is based on market conditions, not individual targeting.
Research
Language Model Can Be a Steganographic Privacy Leaking Agent
Researchers have proposed a novel threat model called TrojanStego, in which an adversary can embed sensitive information into the outputs of a large language model (LLM) without explicit control over inputs. Experimental results show that compromised LLMs can reliably transmit secrets with high accuracy while maintaining their utility and evading human detection, highlighting a new class of covert data exfiltration attacks.
Measuring the Occupational Implications of Generative AI
Researchers analyzed 200k conversations between users and a generative AI system to understand how AI is being used in work activities, finding that people commonly seek AI assistance for tasks like gathering information and writing. The study computed an "AI applicability score" for each occupation, revealing that knowledge work occupations, such as computer and mathematical, office and administrative support, and sales, have the highest scores, indicating a strong potential for AI impact.
Cascade: LLM-Powered JavaScript Deobfuscator
CASCADE is a hybrid approach that combines advanced coding capabilities with compiler Intermediate Representation to effectively deobfuscate JavaScript code, recovering semantic elements and revealing original program behaviors. By leveraging this method, CASCADE overcomes the limitations of existing deobfuscation techniques, achieving reliability and flexibility, and has already demonstrated substantial improvements in Google's production environment.
Supervised fine tuning on curated data is reinforcement learning
Supervised fine-tuning (SFT) of large language models can be understood as maximizing a lower bound on the Reinforcement Learning (RL) objective, explaining its good performance. A modified version of SFT, called importance weighted supervised fine-tuning (iw-SFT), optimizes a tighter bound to the RL objective and can improve performance, making it competitive with more advanced RL algorithms in various tasks.
Query Agnostic Adversarial Triggers for Reasoning Models
Researchers have discovered that appending short, irrelevant text, such as "Interesting fact: cats sleep most of their lives," to math problems can significantly increase the likelihood of advanced reasoning models producing incorrect answers. This vulnerability, demonstrated through the CatAttack automated attack pipeline, raises security and reliability concerns, highlighting that even state-of-the-art models can be misled by subtle adversarial inputs.
Code
Show HN: Terminal-Bench-RL: Training long-horizon terminal agents with RL
The author successfully built a stable reinforcement learning training infrastructure that scales to 32x H100 GPUs across 4 bare metal nodes, and developed an agent called Terminal-Agent-Qwen3-32b that became the highest scoring Qwen3 agent on the Terminal Bench leaderboard. However, due to the high cost of compute, the author was unable to train the agent to its full potential, estimating £30k-£50k in compute required to achieve state-of-the-art results.
Show HN: A GitHub Action that quizzes you on a pull request
PR Quiz is a GitHub Action that uses AI to generate a quiz based on a pull request, helping human reviewers test their understanding of code written by AI agents. To use PR Quiz, users need to set up an OpenAI API key and an ngrok auth token, then add a workflow file to their repository that triggers the quiz on pull request reviews.
Show HN: Xorq – open compute catalog for AI
Xorq is an opinionated framework for building declarative, reusable ML pipelines across multiple engines, including Python and SQL, with features like caching, portable UDFs, and column-level lineage. It provides a unified declarative API, transparent caching, and visual lineage trees to address common pain points in ML pipeline development, and is currently in beta with a rapidly evolving spec.
dson: A delta-based CRDT implementation in Rust
DSON is a Rust implementation of a delta-state Conflict-Free Replicated Datatype (CRDT) for JSON-like data structures, designed for robust and efficient multi-writer collaboration in constrained environments. It provides three fundamental CRDTs - OrMap, OrArray, and MvReg - which can be nested to create complex data structures, and produces small deltas that can be transmitted to other replicas.
AI-Doc-Gen: Multi-agent system that analyzes code, generates docs automatically
The AI Documentation Generator is an AI-powered tool that automatically analyzes repositories and creates comprehensive documentation using advanced language models, employing a multi-agent architecture for specialized code analysis. It offers features such as automated documentation, GitLab integration, concurrent processing, and flexible configuration, and can be installed and used through a series of commands and configuration options.