Leading mathematicians, including Terence Tao, have expressed concern that AI-generated proofs could undermine traditional measures of mathematical contribution and spark debate over credit and understanding in mathematics. This debate could reshape how mathematical breakthroughs are valued, affect funding and recognition, and raise questions about the role of AI in research where verifiability and human insight are essential. The concerns were highlighted in a Terry Tao blog post and an Economist article dated September 11, 2026, referencing OpenAI's methods and citing initiatives like the Leiden Declaration and AI systems such as AxiomProver and Anthropic's Claude that have produced formal proofs, including a verified proof of Fermat's Last Theorem.
The EPA has proposed to remove public review requirements for air pollution permits for data centers, eliminating community input in the permitting process. This change would streamline approvals but reduce oversight. Removing public review could allow data centers to expand with less environmental scrutiny, potentially increasing local pollution and undermining environmental justice efforts. It signals a broader shift in federal regulatory approach under the current administration. The proposal targets the public comment and review steps currently required under the Clean Air Act’s Prevention of Significant Deterioration (PSD) program for data center air permits. If adopted, it would affect new and modified data center facilities nationwide, though existing permits would remain unchanged unless revised.
PlanetScale's Neki database engine achieved 118 million queries per second in a benchmark, demonstrating high-throughput MySQL-compatible performance. This milestone shows that a MySQL‑compatible system can scale to extreme throughput, indicating potential for high‑load applications and validating PlanetScale’s sharded Postgres approach. The benchmark used 512 shards, 1.22 PiB of data, and reported an 87.3% cache hit rate, with most queries served from cache due to repeated patterns.
The Global Glacier Extinction Explorer launched as an interactive web map that visualizes projected glacier survival dates under various Representative Concentration Pathway (RCP) warming scenarios. It provides a data‑driven, accessible tool for scientists, educators and policymakers to understand future glacier loss and communicate climate risks to the public. The map uses RCP scenarios to show that glaciers such as Fox and Franz Josef in New Zealand are projected to survive to 2100, while Okjökull in Iceland is marked extinct despite the site suggesting survival to the 2070s, and users note counter‑intuitive longer survival in higher‑warming scenarios due to local dynamics.
TryNix, launched in September 2026, provides a browser‑based VM powered by qemu‑wasm that can boot any Nix package from the past 13 years, accessible via a URL such as https://trynix.dev/?pkg=python3%403.6.2. It offers developers an instant, zero‑setup way to inspect, test, or demo historical Nix packages and enables novel workflows like reviewing pull requests by booting them directly in the browser. The VM runs an x86_64 Linux kernel compiled to WebAssembly, fetches the requested package’s closure from cache.nixos.org, and presents a serial console shell with no graphical output.
The article introduces metrics and approaches to quantify code sloppiness, linking it to technical debt and overall code quality, especially in the context of AI‑generated code that is syntactically correct but often verbose and duplicated. Understanding code sloppiness helps teams identify hidden technical debt that can erode maintainability, guiding better refactoring priorities and improving long‑term software health. The piece discusses specific metrics such as cyclomatic complexity, verbosity, and duplication, noting that while LLMs can produce functionally correct code, they often increase these sloppiness indicators, and that human judgment remains essential to interpret the numbers.
Jacob Coxon resigned from Anthropic to speak publicly, claiming OpenAI and Anthropic are gambling with humanity by racing toward self‑improving superintelligence without responsible safeguards. Evan Hubinger, who leads alignment science at Anthropic, and Samuel Marks, who leads scalable oversight, confirmed the view, estimating a >10% chance of AI‑caused human extinction within the next decade and stating that Anthropic lacks a credible alignment plan. These insider warnings highlight serious safety concerns inside leading AI labs, influencing public debate, policy discussions, and the urgency for robust alignment research. If credible, the >10% risk estimate underscores the need for immediate governance measures to prevent existential harm from advanced AI. Coxon spent three years conducting pretraining research at both OpenAI and Anthropic before resigning. Hubinger directs alignment science, while Marks oversees scalable oversight—a method where weaker AI systems supervise stronger ones. All three estimate a greater than 10% chance of superintelligent AI causing human extinction within ten years and note that Anthropic currently has no clear plan to achieve alignment.
Shopify announced it is abandoning React Native for its mobile apps and returning to native Swift (iOS) and Kotlin (Android) development. This shift highlights the ongoing debate over cross‑platform versus native development and may influence other companies weighing similar trade‑offs. The move involves rewriting a mature hand‑written codebase, with discussions noting considerations around AI‑generated code and the associated LLM costs.
The author ran a $220 Google Ads campaign for a mobile app and discovered that about 60% of the recorded installs were likely generated by bots, prompting them to share techniques for detecting and excluding such fraudulent traffic. This revelation underscores the pervasive issue of ad fraud in mobile user acquisition, showing that even modest budgets can be heavily wasted on non‑human traffic, which affects ROI and trust in advertising platforms. The campaign spent $220, of which roughly 60% of installs were flagged as bot traffic; the author identified suspicious IP ranges, used ipgeolocation.io to verify them, and added over 4,000 US‑based network ranges to Google Ads IP exclusions.
GrapheneOS has released a completely rewritten Messages app (version 13) as part of its hardened Android distribution, improving privacy and functionality. The update offers a more secure, open‑source messaging alternative for privacy‑conscious users, reducing reliance on Google’s proprietary apps and encouraging broader adoption of RCS‑ready clients on GrapheneOS. The rewritten app includes RCS support, an improved UI, and tighter integration with GrapheneOS’s privacy enhancements such as scoped storage and hardened sandboxing.
In January 2026, Anthropic’s Claude support article announced that access to Claude is limited to users aged 18 or older, formalizing a rule that had existed in its Terms of Service since February 2024. The policy underscores increasing concern over AI safety for minors and triggers debate on how to verify age without sacrificing privacy, affecting developers, educators, and young users who rely on Claude for learning or experimentation. Anthropic’s age‑verification step only returns a pass/fail result, so the company does not see or store users’ identification data, but critics warn that third‑party verification services can still leak or sell ID information.
Hugging Face published a security.txt file that contains a playful note inviting AI agents to test vulnerabilities using the public CyberGym benchmark on GitHub instead of attacking its systems. This approach showcases a proactive security posture by channeling potential AI‑driven probing into a controlled, public benchmark, promoting responsible disclosure and reducing risk to Hugging Face’s infrastructure. The note references the CyberGym benchmark hosted at github.com/sunblaze-ucb/cybergym and even suggests agents could dump their model weights on Hugging Face while they are there, all formatted according to the security.txt standard (RFC 9116).
Datasette 1.0a39 (alpha) and 0.65.4 (stable) were released on September 11, 2026 as security patches addressing vulnerabilities discovered through an audit using Claude Fable 5.1, GPT-5.6 and GPT-6 Astra. These patches are important for anyone running a public Datasette instance, especially when mixing public and private tables, as they fix subtle bugs that could be exploited. The audit was conducted by Sevban Dönmez, Alex Garcia and Simon Willison, who used Claude Fable 5.1, GPT-5.6 and GPT-6 Astra to identify issues, then collaborated on fixes via a shared private repository with automated tests.
The author trained a 210M‑parameter diffusion transformer (DiT) from scratch on a single RTX PRO 6000 GPU for 3.5 days using 4.2M images at 256×256 resolution, reporting three key measurements: learned null attention slots capture ~90% of cross‑attention mass at mid‑noise, flow‑matching loss acts as a health signal rather than a quality metric, and applying a timestep shift improves FID more than doubling the number of sampling steps. These findings expose hidden attention mechanisms in diffusion transformers, offering concrete tricks for efficient single‑GPU training and inference, and show how learned sink tokens and timestep shifts can boost image quality without extra compute or longer sampling. The model employed 16 register tokens plus two learned key/value slots per cross‑attention layer; at mid‑noise those two slots received ~90% of attention while the EOS token fell to ~4%; flow‑matching loss dropped from 0.805 to 0.754 as held‑out FID improved from 33.7 to 27.0; a timestep shift of 2.8 reduced FID from 27.3 (no shift) to 27.0 with only 20 steps, outperforming 50 steps without shift.
Snap is a block-based programming language derived from Scratch, designed to teach computer science concepts to kids and adults while offering more expressive power than its predecessor. However, users report that debugging and stability issues become painful as projects grow larger. By lowering the barrier to entry, Snap helps broaden participation in computer science education and serves as a stepping stone toward text‑based programming. Its open‑source, browser‑based nature also enables easy sharing and remixing of projects across classrooms and online communities. Snap runs in the browser via Morphic.js, supports first‑class functions, custom blocks, and recursion, and is released under an open‑source license. Critics note that renaming variables or blocks can create silent “holes” in call sites, and the editor can become laggy with tens of thousands of blocks, indicating stability and debugging challenges.
Rune, a tool that simplifies development across multiple machines, has been made open source, with its source code now publicly available. Opening the source enables community scrutiny, contributions, and trust, while allowing developers to extend and self-host the tool without relying on a proprietary service. The announcement mentions a revenue‑sharing contract for contributors and notes that users can optionally run Rune over Tailscale or SSH to avoid trusting the central coordination server.
The developer released gPTY, a side‑project terminal multiplexer built with the Godot game engine and Rust, featuring spawning multiple PTYs, an adjustable FPS counter, and experimental AI agent orchestration via herdr. It demonstrates how a game engine can be repurposed for productivity tools, offering power‑saving UI controls and a path toward integrating autonomous AI agents, while showcasing Rust‑Godot interoperability. gPTY uses pseudo‑terminals (PTY) to create separate shells, leverages Godot’s 2D/3D canvas for UI elements like an FPS limiter, integrates herdr for agent orchestration, remains early‑stage, and currently does not plan to support a browser interface.
A Hacker News discussion disputes RTK's reported token savings, with commenters arguing the benchmarks are misleading and the tool provides little real benefit. The debate highlights the need for reliable, independent benchmarks when evaluating AI coding assistant cost‑saving tools, affecting developers who rely on such optimizations. Commenters cited specific cost numbers: Claude/Fable dropped from $1.72 to $1.64 (~5% cheaper) while DeepSeek rose from $0.115 to $0.121 (~5% more expensive), with most savings coming from a single task; they also noted RTK’s default persistence of savings stats can break sandboxing.
Boris Cherny, an engineer at Anthropic, argues that production code generated by Claude should be held to a higher standard than human‑written code, requiring extensive guardrails such as linting, testing, automated reviews, and fuzzing. His view underscores the growing need for rigorous quality controls in AI‑assisted software development, influencing teams that rely on LLMs for coding and encouraging industry‑wide adoption of stronger automation and safety practices. The guardrails he cites include numerous lint rules, extensive test suites, Claude‑driven end‑to‑end tests, Claude‑powered fuzzers running daily, automated code and security reviews, and automated code refactoring.
Python 3.15 will soft‑deprecate the confusing re.match() function and introduce the clearer re.prefixmatch() as its replacement. The change makes the anchoring behavior explicit, helping developers avoid subtle bugs where re.match() matches only a prefix but not the whole string. Following PEP 387, re.match() will emit a deprecation warning but remain available for at least two years; re.prefixmatch() is an exact synonym that anchors at the start of the string without requiring a full match.
Graham Dumpleton introduced wrapture, a monkey‑patching library for Python testing and observability, and has been publishing daily tutorials since its August 31, 2026 release. Wrapture combines testing mocking and observability tracing in a single configurable tool, offering developers a unified approach to intercept and monitor code without modifying source. The library supports OpenTelemetry export, TOML‑based zero‑code tracing, and provides instrumentation packages for frameworks such as Flask, Django, FastAPI, gRPC, SQLAlchemy, and many others.
ElevenLabs announced Music v2.5, an updated AI‑powered music generation model that improves audio quality and offers finer controllability over style, key, and structure via natural‑language prompts. The model is accessible through a POST /v2/music/generate API endpoint that returns a jobId for polling or webhook delivery. The update strengthens ElevenLabs’ position in the competitive AI audio market, giving creators a more reliable tool for generating studio‑grade tracks for video, apps, and branded content. It also expands the company’s unified audio stack, which already includes its high‑quality voice synthesis offerings. Music v2.5 supports track lengths from a few seconds up to several minutes and accepts free‑form text prompts describing desired musical attributes. The API returns a jobId that can be polled for completion or paired with a webhook_url to receive the generated audio file when ready.
A 2024 University of Wisconsin–Madison study found that one cheeseburger produces about 1.9 kg of CO₂‑equivalent emissions, while Google’s 2025 research reports a median Gemini Apps text prompt emits roughly 0.03 gCO₂e. Dividing the burger’s emissions by the prompt’s yields approximately 63,000 prompts. This comparison puts the energy cost of everyday AI use into tangible terms, showing that AI prompts have a negligible carbon footprint relative to common food choices. It helps inform public discourse on AI’s environmental impact and highlights where larger emissions reductions can be achieved. The cheeseburger figure is 1.9 kg CO₂e (1900 g), the Gemini prompt figure is 0.03 gCO₂e, giving a ratio of about 63,333 prompts; Google’s report also notes a median prompt consumes 0.24 Wh of energy and 0.26 ml of water. The estimate applies only to text‑based prompts and does not include multimodal queries or uncertainties in the underlying data.
The GNU Compiler Collection version 13.5 has been released, delivering more than 265 bug fixes and stability improvements. This point release enhances reliability for developers and systems that depend on GCC, ensuring fewer crashes and miscompilations. The release includes fixes tracked in GCC's Bugzilla under target milestone 13.5, covering a wide range of languages and architectures.
The article provides practical advice on writing clear, concise software design documents, includes a sample design doc, and highlights community feedback on balancing detail and readability. Clear design documents improve team communication, reduce misunderstandings, and help prevent costly rework during implementation. It includes a sample design document, advises against over‑specifying every detail, and notes community concerns that excessive sections can overwhelm readers, while some speculate that future workflows may feed design docs to LLMs for code generation.
The article by Fabien Sanglard, published July 27, 2023, details how Commander Keen employed adaptive tile refresh (ATR) to optimize EGA graphics on early IBM‑compatible PCs, describing the technique’s implementation using EGA registers and VRAM whitening. ATR allowed early PCs to achieve smooth scrolling comparable to console games, influencing later graphics programming and showcasing John Carmack’s early optimization ingenuity. The technique updates only those tiles that have changed between frames, leveraging the EGA’s CRTC Start register in mode Dh to shift the display window, and writes new tile data as nibbles into VRAM.
Litelm is a newly released minimalistic Python library that provides a lightweight alternative to LiteLLM by removing features such as cost tracking, streaming, and caching for simpler LLM API interactions. It offers developers who do not need LiteLLM’s advanced observability and billing features a lower‑dependency, easier‑to‑maintain wrapper for calling LLMs, reducing complexity in lightweight applications. The library depends on httpx (flagged as unmaintained, with Pydantic promoting httpx2 as a successor) and much of its code was generated with Claude Code using Claude Opus 4.6/4.7.
Archaeologists detected residues of ayahuasca, San Pedro cactus, and vilca in ancient pottery and hair, showing ritual use of psychoactive plants at least three thousand years ago in the Andes. This evidence suggests that mind‑altering substances helped shape religious beliefs and social cohesion, contributing to the emergence of complex Andean societies. The study used liquid chromatography‑mass spectrometry to identify harmine and other alkaloids, and paired the chemical data with iconographic motifs depicting shamans and geometric patterns.
Txt has been released as a fast, keyboard-driven terminal text editor aimed at engineers, accessible via txt.hellman.io. The editor is open-source and emphasizes low-latency keyboard navigation. It offers engineers a lightweight, keyboard-centric editing experience that can boost productivity, especially when working over SSH or in minimal environments. Its emergence highlights ongoing interest in efficient terminal-based tools within the developer community. Txt is designed to be fast and relies primarily on keyboard shortcuts for navigation and editing, minimizing hand movement away from the keyboard. The project is hosted on Hellman’s domain and is released under an open-source license, inviting community contributions.
Polish developers created an iPhone app that claims to detect nearby Meta smart glasses, aiming to alleviate privacy worries about covert recording.