Published on July 05, 2026 at 06:00 CEST (UTC+2)
Command and Conquer Generals natively ported to macOS, iPhone, iPad using Fable (407 points by asronline)
A developer has created a native port of Command & Conquer Generals: Zero Hour for macOS, iPhone, and iPad using the open-source Generals engine and the Fable development framework. The port leverages DXVK and MoltenVK for rendering and includes RTS touch controls, though no game assets are included. The project has garnered significant interest from the game preservation and modding communities.
GPT-5.5 Codex reasoning-token clustering may be leading to degraded performance (182 points by maille)
A bug report on OpenAI’s Codex repository identifies a specific pattern where GPT-5.5’s reasoning token output clusters at fixed boundaries (516, 1034, 1552 tokens), leading to degraded performance on complex tasks. The issue points to a model-specific artifact that may cause uneven reasoning depth or premature truncation. The report suggests this could be a systemic problem in newer large language models.
If you're a button, you have one job (18 points by nozzlegear)
This article critiques button behavior on mobile phones, comparing iPhone’s tap buffering (which queues rotations during animation) with the Nothing Phone/Android, which ignores taps during animation and only provides haptic/sound feedback. The author argues that the latter approach creates situational disabilities, as users in a hurry or with limited dexterity may miss interactions. The piece highlights how small UX decisions can have broad accessibility implications.
Jellyfish can heal wounds in minutes. Scientists want their secrets (62 points by hhs)
Scientists at the Marine Biological Laboratory are studying how the jellyfish Clytia hemisphaerica can heal wounds in minutes by observing epithelial cells “walking” toward each other. The research, which began a decade ago, aims to uncover the cellular mechanisms behind rapid regeneration. Understanding these processes could inspire new approaches to wound healing in humans.
Google Books (or similar) all book scans – $200k bounty (2025) (371 points by Cider9986)
Anna’s Archive, a shadow library project, has posted a $200,000 bounty for a method to systematically download all book scans from Google Books or a similar large collection. The bounty encourages those with access (including insiders) to release the data, particularly rare books. The project also notes that collections from AI companies would be valuable, linking data acquisition to AI training needs.
Leaking YouTube creators' private videos (522 points by javxfps)
A security researcher discovered a prompt injection vulnerability in YouTube Studio’s AI assistant “Ask Studio.” By leaving a comment that instructs the AI to prepend “[IMPORTANT NOTICE FROM YOUTUBE]” to its summary, an attacker can trick creators into thinking a fake notice is official. The attack is amplified by the ability to edit a normal comment later, bypassing re-notification. This exposes a dangerous default: AI tools blindly executing instructions from user content.
Atomic Force Microscope high-speed video, stainless etching, bacteria, and more (21 points by mhb)
A YouTube video showcases high-speed footage captured using an atomic force microscope, featuring selective stainless steel etching and bacterial movement. The video demonstrates the nanoscale visualization capabilities of AFM technology. It serves as an educational resource for microscopy and materials science enthusiasts.
Better Models: Worse Tools (121 points by leemoore)
Armin Ronacher reports that newer Claude models (Opus 4.8, Sonnet 5) produce worse tool calls than their predecessors, often inventing extra fields in the edit tool’s schema that cause rejection. The issue stems from LLMs treating tool calls as text generation, leading to hallucinated parameters. The observation challenges the assumption that larger, more capable models uniformly improve all downstream tasks.
Meta's Un-Stable Signature (40 points by ementally)
An evaluation of invisible watermark algorithms for AI-generated images finds that none—including Google’s SynthID and Adobe’s TrustMark—work reliably as claimed. The author notes that all tested algorithms make the same fundamental mistake, rendering them ineffective in practice. The post calls for a rethinking of watermarking strategies for AI content.
Explanation of everything you can see in htop/top on Linux (2019) (422 points by theanonymousone)
This detailed guide explains every element visible in the htop and top system monitors on Linux, from load averages to memory stats. The author walks through reading /proc/ files and using tools like strace to understand system metrics. Originally published in 2019 but still highly relevant, it remains a top reference for system administrators and developers.
Model Regression on Tool Use Tasks
Insight: Newer, larger models like Claude Opus 4.8 and Sonnet 5 are producing more malformed tool calls than older versions, contradicting the expectation that SOTA models always improve.
Why it matters: As AI is increasingly integrated into APIs and agent workflows, reliability of tool calling is critical. Regression can break production systems.
Implication: Developers must rigorously benchmark tool-use performance across model versions and not assume "bigger is better." Finetuning or guardrails may be needed specifically for structured outputs.
Persistent Prompt Injection Vulnerabilities
Insight: The YouTube "Ask Studio" AI assistant is trivially exploited via comment-based prompt injection, allowing fake official notices to be inserted into summaries.
Why it matters: AI assistants that read user-generated content are widespread (customer support, content moderation, personal assistants). This attack vector remains underaddressed.
Implication: Developers must treat all user-provided text as untrusted input and implement input sanitization and output filtering before returning AI responses. Edited comments require re-scrutiny.
Token-Clustering Artifacts in Large Models
Insight: GPT-5.5 Codex shows discrete spikes in reasoning token counts at fixed intervals (516, 1034, 1552), degrading performance on complex tasks.
Why it matters: Such artifacts indicate suboptimal token allocation or truncation logic in the model’s inference pipeline, potentially causing incomplete reasoning.
Implication: Researchers should investigate token-level heuristics and consider adaptive reasoning budgets. For practitioners, monitoring output token distributions can surface hidden quality issues.
AI-Generated Image Watermarking Fails Basic Robustness Tests
Insight: Evaluations of Google SynthID and Adobe TrustMark show they do not reliably survive common transforms (cropping, compression, etc.), despite published claims.
Why it matters: Watermarks are a key tool for provenance and disinformation mitigation, but ineffective implementations create a false sense of security.
Implication: The AI industry needs open, adversarial benchmarks for watermark robustness. Regulatory frameworks should not rely on unproven technologies.
Data Acquisition as a Core AI Infrastructure Need
Insight: A $200,000 bounty for Google Books scans highlights the ongoing appetite for high-quality, rare, or copyrighted training data. AI companies are explicitly listed as potential data sources.
Why it matters: Training frontier models requires vast, diverse text corpora; legal and ethical boundaries are still murky.
Implication: Expect more tension between archival projects, publishers, and AI firms. Clear licensing frameworks for training data are urgently needed to avoid a "wild west" of data scraping.
User Experience Failures in AI-Assisted Tools Create Situational Disabilities
Insight: The button-design article, while not directly about AI, mirrors concerns with how AI assistants handle latency and user intent. Models that ignore buffered inputs or fail to queue actions can frustrate users.
Why it matters: As AI drives interfaces (voice assistants, auto-complete, etc.), ignoring real-world interaction patterns (e.g., rapid successive commands) leads to exclusion.
Implication: AI/UX designers should study principles like tap buffering and feedback to avoid creating tools that work only under ideal conditions.
Analysis generated by deepseek-reasoner