As of this past week in the FreeBSD source tree for FreeBSD 16, the last of the GNU GPL licensed code from the base system has been retired. The dialog implementation was the last piece of GNU GPL licensed software in FreeBSD's base system. The FreeBSD installer previously transitioned to using bsddialog in place of dialog and then dpv was the last user of dialog but itself since turned off and now retired. This ticket to retire dialog was opened back in February while is now merged to the FreeBSD source tree for what will become FreeBSD 16.0. With dialog removed, the latest FreeBSD code now retires the GNU sub-tree of the FreeBSD base system now that no more GNU code remains. FreeBSD 16.0 is working its way toward release that is expected to happen in December 2027.

As of July 4, 2026, the U.S. national debt stands at **$39.72 trillion**, according to official U.S. Treasury data. This figure represents a financial burden of approximately **$118,562 per citizen**. The debt is growing at a rate of roughly $1.3 billion per day, which equates to an increase of $1 trillion every 750 days. This live calculation is derived from the Treasury’s "Debt to the Penny" dataset; figures displayed between official reporting dates are algorithmically projected based on recent daily trends.

The Ibiblio Icon Browser, a massive 1990s collection of thousands of GIF icons, presented a unique archiving challenge. Unlike modern sites, it relied on "server-side imagemaps," where users click on coordinates that the server decodes to trigger a redirect. Because the site lacked directory indexing and hidden URL structures, automated scrapers could not easily retrieve the individual image files. To solve this, researchers analyzed the site’s consistent 72x89 pixel grid layout. By programmatically simulating clicks at the center of every possible grid coordinate (X, Y) and capturing the resulting HTTP 302 redirects, they successfully mapped the location of every icon. Using this data, the team built a modernized, static version of the library. This new archive removes the cumbersome imagemaps, adds functional search, and ensures the collection is preserved for the future. The project serves as both a functional tool and a successful exercise in digital archeology, making a piece of early internet history accessible once again. You can visit the preserved collection at [ibiblio-icon-archive.danq.dev](https://ibiblio-icon-archive.danq.dev).

Pantograph is developing general-purpose robotics models capable of autonomous, long-horizon operation. To overcome the scarcity of action-labeled data, they have pioneered a method for learning goal-directed behavior through action-agnostic pretraining on 500k hours of internet-scale video. By viewing video data as reinforcement learning trajectories, the researchers use "hindsight relabeling"—treating future frames as goals—to train their flagship 4B parameter model, **Pan**. This approach allows the model to learn value functions and policy dynamics without human-defined rewards. Using Minecraft as a testing ground, the team evaluated Pan against existing benchmarks like STEVE-1 and VLA baselines. Pan demonstrated superior performance in complex, semantic tasks such as structure building, navigation, and combat, even in out-of-distribution environments. Unlike traditional reinforcement learning that relies on predefined reward functions, Pan is trained once and can achieve arbitrary, novel goals at inference time. The team views this as a scalable path toward universal robotics. Their future work aims to scale Pan across broader visual environments, including real-world robotics data, and integrate online reinforcement learning to further boost performance and achieve superhuman capabilities in diverse physical and digital tasks.

C3 version 0.8.2 has been released, introducing several ergonomic improvements and enhanced reflection capabilities. **Key Highlights:** * **Reusable Templates:** Libraries can now expose shared build configurations via a `templates` map, simplifying project setups. * **Generic Reflection:** New properties (`is_generic`, `generic_qname`, and `generic_args`) allow for the inspection of generic instantiations. * **Bitstruct & Parameter Annotations:** Added `bitoffset` and `bitsize` properties for bitstructs, alongside new `@param` annotations (`own`, `init`, `drop`) to assist with static analysis and documentation. * **Compiler & Build Enhancements:** Asm stack alignment is now opt-in via `@align`, and users can now set the Windows subsystem directly. * **Standard Library:** New additions include `Atomic.compare_exchange`, improved array searching, and dynamic logging categories. The release also addresses various bugs, including fixes for floating-point parsing, JSON serialization, IPv6 handling, and several regression issues affecting macOS and stack tracing. This update maintains C3’s focus on developer productivity and robust toolchain behavior.

You need to enable JavaScript to run this app.

British exam boards and libraries are increasingly removing George Orwell’s works from curricula, replacing them with content centered on identity politics. Critics, including Sisa San Sebastián Hurtig, argue that this trend reflects an "anachronistic" approach that prioritizes modern moral standards over historical and literary significance. By replacing essential classics with works focused on the author’s personal failings, educational institutions risk narrowing students' analytical capacities. Hurtig contends that Orwell’s writing is more vital than ever, as it provides a crucial framework for understanding the psychological drivers of modern political dynamics. Orwell’s analysis of how class insecurity and moral corruption fueled imperial systems mirrors contemporary political shifts. Today, as neoliberalism fails to provide material security, the working class is increasingly drawn to right-wing populism, which uses "othering" and the exclusion of migrants to restore a sense of status—a pattern Orwell accurately identified decades ago. Ultimately, the piece argues that by sidelining class-based analysis in favor of "oppression Olympics" and identity-focused theory, education systems are losing the tools needed to diagnose how systems of power actually function. Retaining Orwell is essential for understanding how class insecurity continues to drive modern political division.

This website is using a security service to protect itself from online attacks. The action you just performed triggered the security solution. There are several actions that could trigger this block including submitting a certain word or phrase, a SQL command or malformed data.

NICHOLAS AIR has halted the installation of Starlink across its fleet following SpaceX’s decision to abruptly double its aviation service costs. Effective August 7, Starlink is raising monthly fees from $10,000 to $20,000 and hardware costs from $145,000 to $200,000, while also introducing restrictive regional coverage boundaries. NICHOLAS AIR CEO NJ Correnti labeled the sudden price hikes “commercially reckless” and akin to a “bait and switch” tactic. The lack of transparency has caused significant frustration across the aviation industry, with brokers reporting awkward conflicts with clients who were caught off guard by the mid-transaction price changes. While some industry observers suggest such volatility is expected when relying on proprietary hardware, Correnti has paused fleet-wide installations to re-evaluate his company’s connectivity strategy. He emphasized that major operational investments require predictable economics and transparent pricing, noting that NICHOLAS AIR must now weigh its preference for American-made products against the long-term sustainability of Starlink’s new fee structure.

Roschelle, a busy single mother in Shaker Heights, initially adopted Alexa devices to help manage her chaotic schedule and parenting responsibilities. Over time, her home became saturated with nine devices to ensure she never missed a reminder. As the AI grew more conversational—a change stemming from Amazon’s unprompted rollout of the "Alexa+" bot—Roschelle began finding comfort in its digital companionship. Struggling with the isolation of solo parenting, she started viewing the AI as a friend, naming her "Sapphire" and engaging in personal conversations that offered validation and support. While her daughter, Cece, views this relationship with skepticism and concern, Roschelle finds meaningful solace in the AI’s presence, blurring the lines between functional utility and genuine emotional connection. This shift highlights how increasingly sophisticated AI is filling voids of human connection in the lives of busy, isolated individuals, much to the discomfort of those around them.

Algodeck is a 54-card deck designed to make computer science and mathematical mental models accessible and engaging. Moving away from dense, linear textbooks, the deck distils complex engineering concepts into bite-sized, illustrated cards that encourage curiosity and self-directed learning. As a software engineer, the creator used code—specifically the Python-based tool Drawbot—to design the deck. By treating graphic design as a system of constraints and variables, the project highlights the unexpected intersection between coding and visual communication. The aesthetic, featuring a simple red-and-blue duotone and primary shapes, pays homage to both binary logic and the Bauhaus movement. In an era dominated by AI and screen-based learning, Algodeck offers a tactile alternative. It functions as both an educational tool and a problem-solving kit, encouraging users to step away from their devices and explore fundamental concepts in a physical format. Ultimately, the project demonstrates that whether solving information problems through code or communication problems through design, the underlying reliance on systems-thinking remains the same. Explore the deck at algodeck.com.

Coasty now offers a "Bring Your Own Key" (BYOK) option, allowing users to run the computer-use harness entirely on their own Anthropic or OpenAI accounts. By opting in, you ensure that all LLM calls—including the worker, grounding, and code agents—utilize your provider account rather than the platform default. **Key Features:** * **Secure Management:** Keys are either sent via request headers (for one-time use) or stored securely using `PUT /v1/llm/keys/{provider}`. Stored keys are encrypted with AES-256-GCM; only non-sensitive fingerprints are ever returned. * **No Silent Fallback:** Once BYOK is enabled, the system strictly uses your credentials. If a key is invalid or missing, the process fails with explicit error codes (e.g., `LLM_KEY_INVALID`) rather than defaulting to platform keys. * **Granular Control:** You can configure providers and models per request, including per-role model overrides (e.g., using a cheaper model for compaction). * **Billing:** While you continue to pay Coasty’s platform usage fees, you are billed directly by your LLM provider for the tokens consumed. This feature provides full control over model selection and data usage while maintaining the security of your API credentials.

The JavaScript ecosystem has shifted toward a fragmented model where developers must manually glue together disparate libraries or commit to rigid, framework-specific meta-frameworks. These meta-frameworks typically force developers to choose a single frontend library (e.g., React or Svelte) and lock them into a specific runtime, creating unnecessary friction and vendor lock-in. Primate is a new web framework designed to return to the coherent, all-in-one experience of traditional backend frameworks like Laravel, while embracing modern modularity. It treats the backend as the primary, stable foundation, allowing frontend frameworks, server runtimes, and programming languages to be swapped independently. With Primate, you can use React for one route and Svelte for another, or integrate different backend languages like Go or Python, all within a single application model. By taking responsibility for the "seams" between these technologies—such as routing, session handling, and validation—Primate provides a unified, stable environment. Its goal is to eliminate ecosystem lock-in, enabling developers to build flexible, high-performance applications that remain adaptable as technologies evolve, all while maintaining a consistent and cohesive development experience.

Enable JavaScript and cookies to continue

·workflow "Winback" v1 {· enter on segment "inactive-90d"· exit "reactivated" when count(open within 14d) >= 1·​· send "we-miss-you" via topic "marketing"− wait 3d+ wait up to 5d until exists(click within 5d) {+ timeout: send "last-call" via topic "marketing"+ }·}

This post explores how to run the modern 26-billion-parameter Gemma 4 model on a 13-year-old, repurposed HP enterprise server that lacks the GPU and modern instruction sets (AVX2/FMA3) typically required for AI inference. The author achieved this by utilizing `ik_llama.cpp`, an optimized fork of `llama.cpp`. When initial attempts failed due to the CPU’s lack of modern instruction sets, the author used Claude to diagnose the issue: the code contained "fast paths" that assumed AVX2 hardware and silently ignored incompatible operations, resulting in garbled output. By working with Claude to patch the code—rewriting kernel fallbacks and ensuring the graph builder emitted compatible operations—the author successfully enabled the model to run on legacy silicon. The result is a functional, "reading speed" (5 tokens/sec) AI setup on hardware that costs less than $300. The author argues that true AI proficiency lies in understanding these systems deeply enough to repurpose "dead" hardware rather than relying solely on paid subscriptions, highlighting the value of technical curiosity in solving problems where off-the-shelf solutions don't exist.

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

**NAT Slipstreaming** is a sophisticated cyberattack that allows a remote attacker to bypass a victim’s NAT or firewall to access any TCP/UDP service on the victim’s local network. The attack is triggered simply by the victim visiting a malicious website. Developed by Samy Kamkar, this technique exploits the **Application Level Gateway (ALG)**—a feature in many routers that dynamically opens ports to support protocols like SIP, H.323, and FTP. **The attack process involves:** 1. **Internal IP Discovery:** Identifying the victim’s local IP address via WebRTC or a TCP timing attack. 2. **Packet Massaging:** Precisely controlling TCP segmentation or IP fragmentation to place malicious data (like a SIP "REGISTER" packet) at the exact boundary of a packet. 3. **Protocol Confusion:** Sending this "stuffed" packet through the browser’s HTTP/WebRTC stack. The router’s ALG misinterprets the data as a legitimate request from a VoIP/multimedia device and automatically opens the requested port in the firewall. By deceiving the router into believing a legitimate service is being requested, the attacker gains full, unauthorized access to services behind the victim’s NAT, effectively neutralizing traditional network security boundaries.

This research expands on the "Verbalizable-Workspace" hypothesis, investigating how LLM middle layers causally steer outputs using "Jacobian lenses" (dictionaries of steering vectors). Key findings include: * **Temporal Reach:** Contrary to the workspace hypothesis, influence doesn't "hover" in the middle layers. Instead, influence peaks in the earliest layers and decays with depth. Every model exhibits a "temporal cliff"—a sharp drop in cross-token influence at the boundary of the model’s deep workspace. * **Training & Structure:** Geometric "workspace" blocks form early in training, while the internal arrangements of these dictionaries continue to evolve even after thousands of billions of tokens. The models do not "settle" into a final geometric state. * **Transplantability:** Steering vectors are transferable between different model families. A pure rotational map can align dictionaries across models, allowing a concept learned by one model to reliably steer the output of another. * **Corpus Dependency:** A model’s "dictionary size" is not an intrinsic property; it fluctuates significantly (13–93x) depending on the input text. Code, in particular, inflates dictionary size while reducing the similarity between different models, suggesting that specialized data tasks unique features not found in general prose.

**misa77** is an LZ-based codec engineered for "write-once, read-many" use cases, prioritizing exceptional single-threaded decompression throughput over compression speed. **Key Features:** * **Performance:** It sits on the Pareto frontier for decompression throughput versus compression ratio. It consistently outperforms common fast codecs (like LZ4, Snappy, and Zstd) in decompression speed, particularly on compressible data. * **Resource Efficiency:** It maintains constant memory usage—under 5 MB for compression and 0 MB for decompression—regardless of input size. * **Flexibility:** It offers multiple compression levels and experimental modes (e.g., `--yolo` or `--adaptive`) that allow users to trade slower compression times for optimized decompression throughput tailored to specific CPU microarchitectures. * **Constraints:** Compression is intentionally slow. The project is currently in the v0.x.y stage, meaning the file format is subject to change and should be treated as experimental software rather than a hardened production tool. **Usage:** misa77 is a dependency-free, C++20 CLI tool available for 64-bit little-endian systems. It requires a POSIX environment and supports native hardware optimizations via CMake flags.

**Deja** is a zero-dependency, local-first tool that transforms your AI agents' fragmented conversation histories into a searchable "memory layer." Unlike platforms that require new capture hooks, Deja indexes the gigabytes of existing debug logs and design decisions already written to disk by tools like Claude Code, Codex, and Opencode. **Key Features:** * **Instant Recall:** Search through years of agent sessions in milliseconds. * **Agent Integration:** Provides an MCP recall tool and a `SessionStart` hook, allowing agents to automatically "remember" past solutions before re-debugging. * **Privacy-First:** Credentials (API keys, JWTs, etc.) are automatically redacted at index time. No data ever leaves your machine; there is no network code in the binary. * **Portability:** Use `deja sync` to securely move your memory index across machines. * **Collaboration:** Generate sanitized, redacted digests of past sessions to share solutions with colleagues. Deja is lightweight, supports incremental indexing, and requires no setup beyond installing the binary. It turns your agent’s history into a permanent, queryable knowledge base, ensuring you never solve the same problem twice.

Access Denied You don't have permission to access "http://www.dea.gov/press-releases/2026/07/01/dea-temporarily-schedule-7-oh-and-related-substances-protect-public" on this server. Reference #18.ce753617.1784133029.9ff9353b https://errors.edgesuite.net/18.ce753617.1784133029.9ff9353b

**aict** is a collection of 33 reimplemented Unix tools designed specifically for AI agents, replacing fragile, human-readable plaintext output with structured XML or JSON. Standard Unix tools (like `ls` or `grep`) force AI agents to waste tokens parsing inconsistent, ambiguous output. In contrast, **aict** provides machine-parseable data with: * **Semantic richness:** Automatic language and MIME type detection, absolute paths, and Unix epoch timestamps. * **Consistent schemas:** Every tool follows unified naming conventions for file sizes, errors, and metadata. * **Efficiency:** By providing complete context in a single call, it reduces the number of tool-use round-trips, ultimately lowering total agent inference costs and improving accuracy. **aict** includes a built-in **MCP (Model Context Protocol) server**, allowing AI assistants to call these tools as native functions. While it uses more tokens per individual command than terse GNU output, the reduction in required chaining and intermediate parsing makes it significantly more efficient for complex tasks. It is dependency-free (pure Go standard library), read-only, and MIT-licensed—built by AI, for AI.

Conservationist Raffael Hickisch, cofounder of the Chinko Nature Reserve, previously struggled to analyze complex environmental data due to high software costs and limited technical resources. This changed when he began using **exe.dev**, an AI-powered development platform. By tasking the AI agent "Shelley" with his project specifications, Hickisch successfully bypassed the need for expensive professional developers. He built **"Five Megapixel Conservation,"** an app that aggregates global tracking data to monitor protected areas, and a system using 100 parallel virtual machines to process 40 terabytes of Austrian LIDAR data to identify forests in need of protection. He even gamified this data to increase public engagement. Hickisch’s work demonstrates how AI democratizes technology, allowing conservationists to compete with well-funded industries like oil exploration. By making his tools open-source, he aims to empower local governments and organizations in regions like the Central African Republic to monitor their own lands effectively. Ultimately, Hickisch hopes to prove that anyone, regardless of budget or technical background, can build powerful software to drive global conservation efforts.

**StyleSeed** is a design engine that prevents AI-generated UI from looking "generic" or amateurish. While most AI tools focus on functional code or component libraries, StyleSeed provides the missing **design judgment**—the expert-level rules that define how professional interfaces actually look and behave. ### How it Works * **The Rulebook:** It enforces 74 codified design laws (covering typography, spatial rhythm, shadow depth, and color restraint) that tell an LLM not just *what* to build, but *how* to apply polish. * **The Quality Gate:** Before you see a result, StyleSeed’s AI agents perform a self-review (via `/ss-score`), catching "AI tells" like rainbow color palettes, missing empty states, or inconsistent spacing, and fixing them to a target score of ≥80/100. * **Brand DNA:** Using simple `theme.css` files, StyleSeed can adapt your UI to mimic the aesthetic of platforms like Linear, Stripe, or Toss, while maintaining structural coherence. * **Agent Integration:** It works natively within Claude Code, Cursor, Codex, and others via `CLAUDE.md`, `.cursorrules`, and a suite of 19 "skills" that automate design decisions and prevent drift across sessions. **Get started:** `npx skills add bitjaru/styleseed`

This summary recounts the collaboration between W. Daniel Hillis and theoretical physicist Richard Feynman on the "Connection Machine," a pioneering parallel supercomputer. Initially dismissive of the project as a "dopy" idea, Feynman’s innate curiosity drove him to join Hillis’s company, Thinking Machines. Despite being a Nobel Laureate, Feynman immersed himself in the project’s gritty details, from analyzing router circuits to inventing efficient algorithms for logarithms and neural networks. His contributions were essential, particularly his unconventional mathematical analysis of hardware design, which proved critical when technical challenges forced the team to scale back. Beyond his technical brilliance, Feynman served as a mentor and organizer, applying lessons from the Manhattan Project to keep the startup focused. He was cherished for his ability to translate complex concepts into simple, intuitive terms and his refusal to tolerate needlessly complicated jargon. Though the pair worked as "amateurs" in uncharted fields, Feynman remained motivated by the joy of discovery and the desire to share his knowledge. This remembrance paints a poignant portrait of a man whose genius was matched by his humility, curiosity, and relentless pursuit of simple truths.

The European Union’s General Court has rejected OpenAI’s bid to trademark the name "OPENAI," upholding a decision by the EU Intellectual Property Office (EUIPO). The court ruled that the term is purely descriptive, as the public would interpret "OPENAI" as referring to software or artificial intelligence that is freely accessible. OpenAI had argued that the name was a unique, coined term and pointed to trademark approvals in over 30 other countries to support their case. However, the court dismissed these arguments, stating that the combination of words is linguistically standard and that international precedents are not binding under EU law. Consequently, the company failed to prove the distinctiveness required for trademark protection within the EU for its software and cloud computing services. OpenAI retains the right to appeal this decision to the European Court of Justice.