Intelligence Digest

The essay argues that relying on LLMs to write for you undermines the thinking process, as writing clarifies thought, and raises concerns about authenticity, disclosure, and loss of personal voice. It highlights the ethical and cognitive risks of using LLMs for writing, affecting authors, educators, and readers who value authentic voice and thoughtful discourse. Published on December 5, 2025 on bcantrill.dtrace.org, the essay notes that writing forces serialization of thought and that LLMs are poor writers lacking personal voice, as echoed in community comments.

Hackers withdrew approximately 4,000 BTC, worth about $320 million, from the Liquid Federation wallet, likely exploiting a recently patched rangeproof cache bug in the Elements blockchain. The theft represents one of the largest side‑chain losses to date, undermining confidence in the Liquid network and highlighting the risks of delayed disclosure of cryptographic fixes. The vulnerable rangeproof cache was fixed in commit c26d719c2… just last week; attackers may have monitored the public commit and exploited the bug before the fix was propagated.

Asahi Linux has announced official support for Apple's M3 chip, enabling Linux to run on the latest Apple Silicon Macs. This milestone expands the open‑source ecosystem on Apple hardware, giving developers and enthusiasts a viable Linux alternative on the newest Macs. The support includes initial GPU drivers and basic hardware enablement, though known issues such as sleep and HDMI functionality remain works in progress.

The A/I collective, known as Autistici/Inventati, announced its shutdown due to US government pressure, citing designation as a global terrorist organization. The shutdown highlights growing tensions between government security measures and digital rights, affecting activists who rely on the platform for secure communication. The U.S. State Department designated the A/I Collective as a Specially Designated Global Terrorist on August 26, 2026, prompting the group to cease operations.

The author describes a year‑long effort to add WebAssembly support to the Anubis proof‑of‑work gateway, targeting Chrome 66 for backward compatibility and redesigning the PoW difficulty scaling. Shows how open‑source projects can modernize with WebAssembly while preserving access for older browsers, and illustrates practical PoW difficulty tuning for anti‑scraping defenses. The implementation targets Chrome 66 (released 2018) to ensure legacy browser support, and the author notes that increasing the PoW difficulty by one makes a challenge 1024× harder in the worst case.

On September 5, 2025, Isar Aerospace’s Spectrum rocket achieved orbit on its second flight and successfully deployed its payloads, marking the first successful orbital launch from Western Europe. This achievement provides Europe with a new sovereign launch option, reducing dependence on foreign providers and strengthening the continent’s autonomous access to space for commercial, defense and scientific missions. The two‑stage liquid‑fueled Spectrum rocket transited MaxQ, completed MECO and stage separation, ignited its second stage, crossed the Kármán line at 100 km, jettisoned the fairing before reaching orbital velocity, and is capable of delivering about 1,000 kg to low Earth orbit.

Nitter and XCancel have resumed operation after receiving legal advice, restoring access to their alternative frontends for X (formerly Twitter). The projects announced the resumption via a GitHub commit and updated legal notices. The resumption highlights the ongoing demand for privacy‑focused, ad‑free ways to view X content amid legal pressure on third‑party frontends, and it preserves an important tool for users who wish to avoid tracking and account requirements. It also underscores the broader tension between platform control and open‑web access. The update appears in commit 1428b4c2 of the zedeus/nitter repository, which references legal advice and indicates that service will continue; XCancel’s site xcancel.com also posted a notice pointing to a CDC legal page (https://xcancel.com/cdclegal) confirming the resumption. Both services remain focused on privacy, offering JavaScript‑free, ad‑free browsing without requiring an X account.

OpenAI published a blog post detailing its work on building an automated AI researcher that can perform research tasks under human supervision, and discussed how such systems could aid AI alignment and safety. Accelerating AI research with automated agents could dramatically increase the pace of innovation, but also raises critical alignment and safety challenges that must be addressed to ensure beneficial outcomes. The system is described as a 'research intern' capable of carrying out well‑defined tasks that would take a skilled researcher a few days, operating under human direction, and OpenAI notes the high cost of human researchers (~$8,000 per day) as motivation for automation.

GrapheneOS announced it will overhaul its default Android apps—SMS/RCS, Gallery, and Keyboard—with privacy‑focused alternatives and will introduce a secure clipboard feature that blocks apps from reading clipboard data without explicit user permission. The changes are part of an ongoing effort to harden the OS while maintaining app compatibility. These updates significantly boost the privacy and security posture of GrapheneOS, giving users a hardened Android fork with less reliance on Google services and stronger defenses against clipboard‑based data exfiltration, which is increasingly relevant amid rising mobile threats. The move also signals the project’s commitment to continuously improve core components beyond the baseline AOSP offerings. The overhaul will replace the AOSP SMS/RCS app with a modern messaging app, update the Gallery app, and may substitute the AOSP Keyboard with alternatives such as FUTO; the secure clipboard will leverage Android’s IS_SENSITIVE flag and require explicit user approval before any app can access clipboard contents. GrapheneOS recently hired additional developers to accelerate these changes, indicating a faster rollout timeline.

The author shows how to create a working Python interpreter whose source code is only 1024 bytes long, using aggressive code‑size tricks and a minimal language subset. This feat illustrates how far interpreter design can be squeezed for educational and code‑golf purposes, highlighting techniques that may inspire ultra‑low‑footprint embedded scripting languages. The interpreter uses single‑character variable names, treats its own source as the data for function calls and loops, and omits features like whitespace‑based scoping to stay within the size limit.

The article reports that political interference in the US Census Bureau is undermining the reliability of federal statistical data, including delays to the 2025 American Community Survey and restrictions on noise infusion methods. Such interference erodes public trust in federal statistics, which are essential for fair congressional apportionment, federal funding distribution, and evidence‑based policymaking. Researchers and businesses rely on these data, so any bias or suppression can lead to misguided decisions across the economy. The article notes a ban on noise infusion techniques that protect respondent privacy, delays in releasing the 2025 American Community Survey, and calls for greater scrutiny of methodological changes that may be politically motivated. Commenters also warned that without independent forensic analysis, the full extent of data corruption might not be known until years later, possibly around 2029.

Terence Eden cites an Interisle report indicating that of the 85 million new gTLD domain registrations in 2025, about 8.5 million were added to blocklists by May 2025, suggesting an abuse rate of at least 10% and likely around 20%, meaning roughly one in five new gTLD domains are used for scams. The findings reveal a systemic abuse of the Domain Name System that threatens everyday internet users, undermines trust in domain registration, and pressures ICANN and registrars to strengthen anti‑abuse measures. The report notes 85 million new gTLD registrations in 2025, with 8.5 million blocklisted by May 2025, estimates a 10% floor abuse rate and a likely 20% rate, and cites ongoing ICANN discussions about the problem.

A researcher claims to have jailbroken OpenAI's GPT-6 Astra model within 24 hours of its release using an extended Task-in-Prompt (TIP) attack combined with four other undisclosed techniques, and privately disclosed the method to OpenAI. The rapid jailbreak highlights ongoing challenges in LLM safety alignment and demonstrates how adversarial techniques can evolve quickly after model releases, affecting AI security research and deployment practices. The attack builds on the ACL 2025 TIP paper, which hides harmful objectives inside benign tasks like solving a cipher or executing code, and was adapted because the original minimal TIP no longer worked on GPT-6 Astra.

NetBSD 9.5 has been released as the final update for the NetBSD-9 branch, marking the end of the series after which it will reach end-of-life. This release provides important security patches and hardware support updates, giving users a last chance to keep their systems secure before the branch is no longer maintained. NetBSD 9.5 includes the latest security fixes and driver updates for the 9.x series and is the last planned release; after this, no further updates will be provided for NetBSD-9.

The NetBSD gallery page showcases a collection of academic research projects that have used NetBSD as their operating system platform. This highlights NetBSD's ongoing relevance in academic research, demonstrating its portability, permissive licensing, and suitability for diverse hardware platforms. The gallery lists multiple projects; accompanying Hacker News comments discuss usability as a daily driver, licensing advantages, Microsoft's use of NetBSD on a new processor, and general skepticism, with no major breakthrough announced.

Simon Willison reports that OpenAI researchers are increasingly using coding agents and have launched a recursive self‑improvement initiative, with a chart showing median daily AI spend per researcher rising from near zero in February 2026 to about $600 by late August 2026, likely after gaining access to the subsequently released GPT‑6 Astra model. The trend illustrates how AI‑driven coding agents and self‑improvement loops are boosting research productivity, signaling a broader shift toward AI‑augmented R&D that could accelerate progress toward artificial general intelligence. The chart’s blue line rises slowly to ~$50 by April, plateaus around $150‑$165 in July, then climbs steeply to roughly $600 by late August 2026; internally used coding agents include OpenAI’s Codex CLI, and the acceleration coincides with an internal “RSI day” and presumed access to GPT‑6 Astra.

Simon Willison argues that attempting to rewrite a legacy system from scratch rarely succeeds because the old system continues to evolve while the new team lacks incentive for quality. This insight warns engineering teams that rewrite projects often create dual‑system complexity and wasted effort, highlighting the need for incremental migration strategies. He notes that the legacy system remains a moving target, its developers lose motivation, the new team faces unknown scope, and after delays the new system ships handling only a subset, leaving two systems in production.

The paper introduces a universal geometric framework that identifies shared structural properties across diverse embedding spaces and presents the vec2vec method for translating embeddings from unseen encoders while preserving their geometry. By providing a way to align and translate embeddings across models, the work could improve interoperability of vector databases, enhance representation learning, and raise new considerations for the security of embedding-based systems. The authors demonstrate that vec2vec achieves high cosine similarity between translated embeddings and target embeddings, leveraging manifold alignment techniques to capture universal geometry; the paper is authored by Rishi Jha, Collin Zhang, Vitaly Shmatikov, and John X. Morris and appears as arXiv:2505.12540.

A Reddit post argues that reproducibility in machine learning research is becoming irrelevant because experiments now require expensive hardware, industry demos are opaque and unverifiable, and researchers have incentives to hide non‑reproducible work; commenters debate the validity of these claims. If reproducibility erodes, trust in ML findings declines, hindering scientific progress and making it harder for academia and industry to build on each other's work, especially as hardware costs rise and AI systems become more pervasive. The author cites three reasons: (1) costly hardware and lab setups needed for physical AI experiments, (2) unverifiable claims from big AI companies about tool performance, and (3) personal incentives to withhold code to avoid competition; commenters note that high‑stakes projects like the Moon landing had tight internal validation, whereas most ML research lacks such checks.

Mador is an 80‑line JavaScript library that uses ES Proxies to create reactive DOM bindings via a simple read/write API, allowing developers to bind state to DOM elements with minimal code. It demonstrates how Proxy‑based reactivity can replace heavier frameworks for simple UI updates, offering a lightweight alternative for developers seeking minimal bundle size. The library exports a function that returns a [read, write] tuple; read registers a CSS selector and a callback that runs when tracked state changes, while write mutates the proxy state and automatically queues DOM updates.

The article examines how excessive doomscrolling on social media heightens anxiety and undermines productivity, and shares personal strategies to curb the habit. It highlights growing mental‑health concerns tied to digital consumption, showing how widespread doomscrolling can affect wellbeing and productivity for millions of social media users. The piece cites personal anecdotes—such as deleting most social media accounts, setting time limits, practicing mindfulness, and taking analog breaks—as methods to reduce doomscrolling.

Zach Kehs observed that, unlike physical buildings that would collapse under endless expansion, software can always become worse through added indirection or reduced performance. This insight underscores the unbounded nature of technical debt and software entropy, reminding developers that code quality can degrade indefinitely without active maintenance. The quote highlights that there is no physical limit to software deterioration, specifically citing the possibility of endless layers of indirection or performance loss.

An opinion piece on Neowin argues that Windows 11's special developer edition is a marketing misfire, offering little real benefit to developers. The article claims the edition is more about publicity than substantive developer tools. It highlights Microsoft's ongoing challenge to deliver meaningful developer-focused improvements in Windows 11, potentially affecting developer confidence and adoption of the platform. Critics worry that such marketing‑driven releases may divert resources from genuine tools like Dev Home, Dev Drive, and WinGet. The piece references Windows 11 Dev Home (including Dev Drive and WinGet integration), notes that Microsoft Dev Box is in maintenance mode with focus shifting to Windows 365, and suggests the special edition lacks concrete developer features.

D2, the declarative diagram scripting language, announced that it has transitioned to a non-profit organization to ensure long-term community stewardship. This shift signals a commitment to keep D2 open source and community-driven, reducing reliance on corporate sponsorship and enhancing trust among developers who use it for version-controlled diagrams. The non-profit status will govern the project’s finances and direction, while D2 continues to support SVG, PNG, GIF, PDF, and PPTX exports and offers features like auto‑formatting, syntax highlighting, and planned LSP integration.

In September 2026, a Minnesota federal court denied Elon Musk's request for a preliminary injunction that would have blocked the state's new law prohibiting AI-generated child sexual abuse material. The ruling highlights increasing state-level efforts to regulate AI-generated CSAM and may influence similar legislation elsewhere, while testing the limits of First Amendment defenses for synthetic content. Minnesota’s law holds AI companies liable for creating nudification images of children on their platforms, and Musk’s challenge argued the statute infringes free speech; the court found the law likely constitutional and declined to issue an injunction.

The Pentagon announced that its ban on using Anthropic's AI models remains in effect, contradicting earlier statements by Commerce Secretary Howard Lutnick that the administration had resolved issues with the company. The conflicting signals highlight ongoing uncertainty in U.S. government AI procurement policy, potentially affecting defense contractors and Anthropic's access to federal contracts. The Pentagon's stance persists despite a recent U.S. court ruling that deemed the ban unlawful and Lutnick's claim that the Trump administration had settled disputes with Anthropic.

A Reddit user argues that LLMs could achieve AGI within the next 12‑18 months according to OpenAI's definition, challenging common claims about their limitations. The discussion highlights shifting perceptions of AGI timelines and influences expectations for AI development, affecting researchers, policymakers, and industry stakeholders. The post cites OpenAI’s AGI definition (“highly autonomous systems that outperform humans at most economically valuable work”), mentions Google’s Project Astra as evidence of progress, and rebuts claims that LLMs cannot learn new weights, are merely next‑token predictors, or lack creativity.