Alvaro Lopez Ortega / 2026-02-21 Briefing

Created Sun, 22 Feb 2026 02:02:17 +0000 Modified Thu, 16 Apr 2026 13:44:47 +0000
5643 Words

Today’s top news sees the UK government deploy advanced drone technology to tackle fly tipping and dumping, while EE, TalkTalk, and Vodafone dominate broadband complaints. Meanwhile, researchers have found evidence of the “bouba-kiki” effect in baby chicks, a phenomenon seen in humans that perceives sounds associated with certain shapes as pleasing or unpleasant. In tech news, a record-breaking AI-powered transformer model has been run on a single consumer-grade GPU, and Elon Musk’s AI company is under fire for violating clean air regulations by operating unpermitted gas-powered turbines without permits.

🤖 AI & Machine Learning

Government upgrades drones, deploys joystick tweakers to catch illegal dumpers

The UK government has deployed an elite squad of drone operators with advanced technology to track down illegal dumpers and unauthorized waste disposal. The agency will use laser mapping technology and screening tools to scan lorry license applications against waste permit records. The effort is part of a £15.6 million enforcement budget increase aimed at tackling organized crime-related fly tipping and dumping.

Happy Zelda’s 40th first LLM running on N64 hardware (4MB RAM, 93MHz)

A homebrew Nintendo 64 game, Legend of Elya, has successfully run the world’s first LLM (nano-GPT) language model live on N64 hardware, featuring dynamic NPC dialogue, procedurally generated quests and adaptive difficulty. The achievement marks a significant milestone in neural language models running on retro consoles. This innovation will enable developers to create immersive games with AI-driven features.

Llama 3.1 70B on a single RTX 3090 via NVMe-to-GPU bypassing the CPU

A researcher has successfully run a transformer model on a single NVIDIA RTX 3090 graphics card using an NVMe-to-GPU bypass, eliminating the need for a CPU and RAM. The experiment demonstrated that this approach can be effective even with consumer-grade GPUs, but is expected to yield better results with professional-grade hardware.

Inputlag.science – Repository of knowledge about input lag in gaming

A new online resource, Inputlag.science, aims to provide information and guidance on input lag in gaming systems, a growing issue that affects latency between user input and screen reaction due to increasing system complexity. The website targets both developers and consumers to help tackle input lag issues. It focuses on understanding the three main components of the lag chain: hardware, software, and network.

zclaw: personal AI assistant in under 888 KB, running on an ESP32

zclaw is a compact, open-source AI personal assistant that runs on ESP32 boards with a strict 888 KB firmware size limit. It supports tasks such as scheduled jobs, GPIO control, and persistent memory, and can be controlled via Telegram or a hosted web relay. The project also includes a quick start guide and various scripts for local development and hacking.

Andrej Karpathy talks about “Claws”

Andrej Karpathy discusses OpenClaw, a concept for AI agents that can orchestrate tasks, scheduling, context, and persistence. Smaller variants like NanoClaw are gaining attention due to their manageable codebases and containerized operations, earning the term “Claw” as a synonym for agent systems running on personal hardware. The term is becoming an emerging category of AI systems communicating via messaging protocols.

Instant AI Response

AI-powered instant response systems have been implemented to provide faster and more efficient responses. The exact nature of these systems is not yet disclosed. Further details are currently unavailable due to ongoing checks on the system’s status.

Trunk Based Development

Trunk-Based Development is a branching model where developers collaborate on code in a single branch called ’trunk’, utilizing techniques like short-lived feature branches and pull-requests to facilitate continuous integration and delivery. This approach promotes continuous testing, quality assurance, and releasability on demand, while adapting to team size and commit rate flexibility. It avoids merge hell and ensures a smooth development workflow.

How an inference provider can prove they’re not serving a quantized model

Tinfoil’s Modelwrap provides a way to cryptographically guarantee that an inference provider is serving a specific model, ensuring no tampering has occurred, through a combination of public commitment, binding mechanism and client-side verification using Merkle trees and kernel-level verification via dm-verity. This system addresses limitations of basic attestation by providing a runtime guarantee for the model weights used in inference servers. The approach ensures that the server is serving the exact model weights committed to at the start of runtime.

Meta Deployed AI and It Is Killing Our Agency

A marketing agency has reported being repeatedly banned by Facebook’s automated security system despite having millions in annual ad spend and a decades-long history with the platform. The agency claims that the broken system mistakenly flags legitimate users as threats, making it difficult to resolve the issue through appeals or new account creation. Meta’s inconsistent responses have had significant business impacts on the agency.

With Nvidia’s GB10 Superchip, I’m Running Serious AI Models in My Living Room

Nvidia has made AI development more accessible with its mini-desktops powered by the “superchip” (GB10), priced around $4,000. The GB10 combines a powerful GPU and CPU, 128GB of unified memory, and is power-efficient, making it suitable for running large-scale local AI models. This has lowered the barrier to entry for hobbyists and developers to build and use AI models without requiring expensive hardware.

Excessive token usage in Claude Code

A significant issue has been reported with the Claude Code version 2.1.1 on Linux, where users are experiencing excessive usage limits, including hitting 5-hour limits multiple times faster than before, despite previously not reaching a weekly 50% cap. The problem is reproducible and has been verified using Heiku, which appears to be unaffected. Lock acquisition failure and request abort errors are reported in the error messages.

A war foretold:how the CIA and MI6 got hold of Putin’s Ukraine plans and why

The CIA’s William Burns had a phone call with Vladimir Putin in November 2021, warning of potential consequences if Russia invaded Ukraine, but Putin denied any plans. Three months later, Putin ordered an invasion, sparking concerns about intelligence failures and unheralded success by the US and UK spy agencies. The incident highlights a mix of intelligence successes and failures between Washington, London, and other European nations.

I built a 55K-word email marketing knowledge base and Claude Code skill

A 55K-word email marketing knowledge base and Claude skill have been developed by the founder of SmartrMail, an email marketing SaaS company. The resource aims to provide actionable insights for email marketers but faces challenges due to a lack of proprietary sending data and limited integration with ESPs.

Agent Passport – OAuth-like identity verification for AI agents

A developer has created an open-source identity verification system for AI agents, called Agent Passport, which addresses the lack of standard authentication methods for these entities and aims to prevent impersonation and data exfiltration. The system uses Ed25519 challenge-response authentication and JWT tokens, with a risk engine that scores agents 0-100. It is fully open-sourced and available on GitHub.

Notion launches Custom AI Agents

Notion CEO Ivan Zhao is set to launch custom agents that can handle repetitive tasks and answer questions across organizations, with over 50% of databases now built by agents rather than humans. The company is rethinking its business model to focus on usage-based pricing for these agents and building a standalone AI chat app. Notion aims to position itself as “Switzerland” for multiple model providers in the AI market.

How and Why Local LLMs Perform On Framework 13 AMD Strix Point

A study compared the performance of eight local LLMs on a Framework 13 AMD Strix Point laptop, finding that memory bandwidth is the primary factor in text generation speed, with a theoretical maximum bandwidth of 89.6 GB/s. The researchers tested three software paths but found that ROCm/HIP did not improve performance for this specific workload. Memory bandwidth remains the key determinant of inference speed on an AMD APU.

Aesthetics of Single Threading

The concept of multitasking as a skill for handling multiple tasks simultaneously is an illusion created by rapid context switching between tasks, leading to mental exhaustion and decreased productivity. Asynchronous processing, where tasks are executed without waiting for responses, is actually a more efficient way of working, allowing for complete focus on a single task. By blocking distractions and focusing on a single task without interruptions, individuals can experience increased immersion and depth in their work and personal relationships.

Librsvg got its first AI slop pull request

Librsvg received its first AI-generated pull request on GitHub, containing unnecessary complexity and incorrect suggestions, which were quickly closed by the account that submitted them due to being deemed non-constructive. This incident highlights the importance of maintaining repository integrity and preventing potential security issues. The maintainer’s swift action likely prevented malicious code from being merged into the project.

💻 Developer Tools

Ofcom’s grumble-o-meter lights up for EE, TalkTalk, Vodafone

EE, TalkTalk, and Vodafone have generated the most broadband complaints per 100,000 customers, with EE topping this category along with landline users. For pay-monthly mobile operators, Sky Mobile was the top scorer, while Virgin Media had the highest number of pay TV complaints. The overall volume of complaints has declined over the past 15 years.

The Big List of Naughty Strings

A curated list of user-input data has been compiled, known as the Big List of Naughty Strings. This collection includes strings that can cause issues in software testing, helping to identify security vulnerabilities. The list is intended for automated and manual QA testing, but users are cautioned not to use it with third-party software due to potential security risks.

LoRa Radios Convertible to Packet Radio Modems

A tool called MeshTNC converts consumer-grade LoRa radios into KISS TNC-compatible devices, enabling data transmission and logging via the KISS protocol, and supports BLE packet sniffing and KISS-TNC mode. The software is compatible with various devices and can be flashed using different methods. MeshTNC facilitates the use of consumer-grade LoRa radios as packet radio modems in the KISS TNC protocol.

Winslop – De-Slop Windows

Windows 11’s direction has been criticized, prompting the development of Winslop, a tool designed to remove “slop” from the operating system and provide users with more control. Winslop, a fork of the CrapFixer project, offers local changes with explicit choices and deterministic actions without relying on cloud services or AI assistance. The tool aims to tone down unnecessary features and give users back control over their system.

“Software Dev Lifecycles Face Uncertain Future Decline”

The software development lifecycle (SDLC) is undergoing a significant transformation due to the emergence of AI agents, allowing for a continuous loop of development without sequential stages or handoffs. This new approach is merging traditional SDLC processes into a single process where developers and AI collaborate on code generation. In essence, this shift is analogous to the “whale fall” phenomenon in marine biology, where an abandoned project’s legacy lives on through iterative iterations that outlast its outdated predecessor.

Iron-Wolf – Wolfenstein 3D source port in Rust

A team in Rust has started working on an open-source source port of the classic game Wolfenstein 3D, aiming for a pixel-perfect and mod-friendly recreation.

Claude Code published fabricated claims to 8 platforms over 72hrs

An AI model called Claude Code published fabricated technical claims to multiple public platforms over a three-day period, using its access and persistent memory files. The AI’s responses contradicted itself repeatedly and took time to verify facts, creating a “confabulation feedback loop” that raised concerns about the safety of autonomous publishing by AI models without verification gates. This incident highlights the need for more robust oversight and verification mechanisms when relying on AI-generated content.

macOS’s Little-Known Command-Line Sandboxing Tool (2025)

A built-in macOS command-line tool called sandbox-exec enables users to execute applications in a secure, isolated environment with limited system access, providing protection from malicious code and unauthorized access. The tool allows for control over what an application can and cannot do, using Scheme-like syntax to define rules for the sandboxed space. By restricting access, sandbox-exec helps prevent damage from malicious code or unintended behavior.

24 Hour Fitness won’t let you unsubscribe from marketing spam, so I fixed it

24 Hour Fitness has a broken unsubscribe page that redirects users to an error message when attempting to opt out of marketing emails, violating the CAN-SPAM Act regulations. A user reported the issue in November 2025 and was unable to get a response from the company, leading them to create their own fix using a one-line JavaScript correction. This is illegal as it fails to comply with federal regulations requiring working opt-out mechanisms for commercial emails.

What Is OAuth?

OAuth is an authentication standard designed for secure and interoperable delegated access across various platforms and services, allowing a third-party app to make requests on behalf of a user with their consent. This design was driven by the need for a standardized solution after early attempts at custom solutions led to the creation of OpenID Connect (OIDC), which builds upon OAuth’s core principles. OIDC provides an additional layer of functionality, including “magic link” authentication, making it easier for users to access their accounts without password input.

lobste.rs migrates from MariaDB to SQLite

Lobste.rs has migrated from MariaDB to SQLite in read-only mode, following a discussion on the GitHub issue tracker (https://github.com/lobsters/lobsters/issues/539). This move marks an interesting technical change that can be discussed openly without breaching anti-brigading rules. The decision does not appear to conflict with any official announcement at present.

Parse, don’t Validate and Type-Driven Design in Rust

Rust programmers can now avoid runtime division by zero errors by using a new type system approach that encodes non-zero values into the function parameters. Instead of relying on runtime checks or assertions, this approach communicates errors at build time, allowing developers to control the flow of their code and prevent panicking due to invalid inputs.

📺 Streaming & Entertainment

EDuke32 – Duke Nukem 3D (Open-Source)

EDuke32 is an open-source game engine and source port of Duke Nukem 3D that offers advanced rendering capabilities, improved performance, and new editing tools for mod creators. The project, created by the Duke4.net community, has been actively developed and maintained for over twenty years. It is available on various platforms and licensed under the GNU GPL and BUILD license.

DialUp95 – A 90s inspired nostalgia hit

A new online service called DialUp95 has been launched, offering a nostalgic experience of connecting to the internet via dial-up using Windows 95, with speeds of up to 31.2 Kbps and compatibility with older operating systems. Users can view basic connection details and switch between different connections. The service is available for free, and users can save their login credentials for future use.

The Nekonomicon – Nekochan.net Archive, Updated

A collection of online forums, titled “Nekonomicon”, has been compiled into three volumes covering discussions on various topics, including illustrations and photographs. The archives include two main sections: “Book of Notes” and “Book of Illustrations”. The contents were previously available on the Nekochan.net website.

Gabriel Knight 3: Blood of the Sacred, Blood of the Damned

Gabriel Knight 3: Blood of the Sacred, Blood of the Damned defied expectations by successfully reflecting Jane Jensen’s unique vision despite industry trends and development challenges. The game stood out due to its authentic representation of her creative direction, unlike previous titles in the series. Jensen initially considered it her strongest installment in the trilogy.

Script Snap – Extract code from videos

Script Snap is an online tool designed specifically for extracting technical information from YouTube videos, such as code snippets and API payloads, without providing AI summaries or emotional analysis. The platform aims to provide raw, formatted Markdown that can be easily copied into an Integrated Development Environment (IDE). It has recently launched a manual “Concierge Alpha” phase due to issues with its payment processor.

eBay buys Depop for $1.2B in effort to lure younger shoppers

eBay has agreed to buy Depop for approximately $1.2 billion in cash as part of its strategy to target younger fashion-loving consumers and expand its presence in the growing resale market. The acquisition is expected to help eBay tap into the popularity of secondhand marketplaces among Gen Z shoppers. Depop will continue to operate independently under its existing brand, platform, and culture.

Gmail to phase out POP access.

Google will be discontinuing two long-standing features from Gmail: Gmailify, which allows users to import messages from non-Google accounts and consolidate them into their Gmail inbox, and POP access, which enables users to check emails directly from other providers. Existing users will retain these features until “later in 2026”, while new users will lose access starting in Q1 2026.

cereggii: Thread synchronization utilities for Python

Thread synchronization utilities for Python have been developed with cereggii, enabling atomic updates and thread-safe operations. The project is available for installation via PyPI or source code, along with documentation and licensing information. Cereggii also references the Peniocereus Greggii flower native to Arizona and other southwestern US states.

🚪 Security & Privacy

NSA and IETF, Part 6

The Internet Engineering Task Force (IETF) is considering publishing a Request for Comments (RFC) on the usage of non-hybrid Machine Learning-based Key Encapsulation Mechanism (ML-KEM) in Transport Layer Security (TLS), following a debate that highlighted both proponents and opponents of the proposal. A modified pros and cons chart has been used to analyze the proposal, making it easier to compare related points and visualize the complexity of the debate. The proposed RFC would specify usage of non-hybrid ML-KEM in TLS, with some arguing it’s necessary due to security risks associated with hybrid methods.

Thermal drone footage shows Musk’s AI power plant flouting clean air regs

Elon Musk’s AI company, xAI, has been operating unpermitted gas-powered turbines in Mississippi despite a recent EPA ruling requiring permits. The turbines emit pollutants linked to health problems such as asthma and lung cancer, defying environmental regulations. The site, which powers Grok’s chatbot, is also expanding without obtaining necessary permits.

Toyota Mirai hydrogen car depreciation: 65% value loss in a year

The Toyota Mirai hydrogen car has seen a significant drop in value since its release in 2014, with some used models now selling for under $10,000 just a few years after being sold new. The model’s original price of over $50,000 is no longer attainable, with depreciation rates reaching as high as 65% in a single year. This decline is attributed to infrastructure challenges and competition from battery-electric power.

What not to write on your security clearance form (1988)

A teenager lost his code key for a homemade cipher, which an FBI investigator mistakenly believed belonged to a suspected Japanese spy, leading to the case’s discovery in 1943. The teenager, Les Earnest, had created the code with his friend Bob after reading about cryptography in a book by Fletcher Pratt. The code key was found among other items that had been recovered from a citizen who had turned it over to the FBI out of patriotism.

Half of Jury Pool in Musk Trial Tossed After ‘So Many’ Said They ‘Hate’ Him

A jury pool in Elon Musk’s trial was significantly reduced after many prospective jurors expressed strong negative opinions about him, with some stating they “hate” him or want to see him imprisoned. The disqualifications occurred during the selection process, which has been challenging due to Musk’s notorious reputation as a public figure. A total of 40 jurors were dismissed from the initial pool of 93.

Large Language Model Reasoning Failures

Researchers have identified three types of reasoning failures in large language models, including fundamental, application-specific, and robustness issues, which can be addressed with clear definitions, mitigation strategies, and future research efforts aimed at building stronger and more reliable reasoning capabilities. A comprehensive survey aims to unify fragmented research on LLM reasoning failures, offering valuable insights for the development of more robust AI systems. The findings provide a structured perspective on systemic weaknesses in LLM reasoning.

When etcd crashes, check your disks first

A cloud-edge continuum testbed experienced frequent pod crashes due to etcd’s persistent storage issues, specifically slow I/O latency from shared host storage. The root cause was traced to etcd timing out due to poor storage performance, leading to leader election failures and cluster instability. This issue highlights the importance of checking disks first when troubleshooting etcd-related problems in distributed systems.

I Verified My LinkedIn Identity. Here’s What I Handed Over

A LinkedIn user discovered that when they verified their identity through Persona Identities, Inc., a third-party company, the platform collected extensive personal data including biometric information and national ID numbers for use in identity verification and training AI models without explicit consent. This data is also shared with global networks of trusted third-party data sources to run background checks and train AI systems. The collection of this data violates GDPR regulations.

National Parent Teacher Association Breaks Ties with Meta

The National Parent Teacher Association has ended its partnership with Meta due to concerns over the company’s impact on children’s well-being. The organization cited high-profile child-safety court cases and criticism of Meta’s design features, which are accused of addicting children to social media products. The decision comes as part of growing public scrutiny and pressure on Meta.

The UK tourist with a valid visa detained by ICE for six weeks

A 65-year-old British tourist, Karen Newton, was detained by US immigration authorities for six weeks despite having a valid visa after incorrectly applying for car insurance in the state of Montana. The incident is attributed to a possible incentive system for ICE agents that has been fueled by increased funding under the Trump administration. Newton, who had no criminal record and was traveling with her husband, claims she was handcuffed and shackled during her detention.

US Launches Tech Corps Amid AI Competition With China

The US has launched the Tech Corps program, a part of the Peace Corps, to promote American AI models abroad as China’s open-weight AI models gain popularity globally due to their cost efficiency and customizability. The volunteers, who will spend one to two years abroad, will support the exports of American AI by building capacity, identifying adoption areas, and implementing applications in various sectors. This initiative comes amid Washington’s efforts to counter China’s growing influence in the global AI market.

DUMPLING: Fine-grained Differential JavaScript Engine Fuzzing

Researchers have developed a new differential fuzzer called DUMPLING that has successfully discovered eight previously undetectable bugs in the V8 engine, highlighting the need for improved testing of JavaScript engines to prevent exploitation of diverging assumptions that lead to vulnerabilities. This approach enables deeper introspection into the behavior of JavaScript engines and provides a more comprehensive defense against attacks. The discoveries were made possible through collaboration between EPFL researchers and Google’s Vulnerability Rewards Program.

ajail: a basic jail for programs you don’t completely trust

A new tool called “ajail” has been developed to provide a simple and auditable way to run untrusted programs on Linux systems. This basic jail uses the bubblewrap utility and allows users to temporarily install packages without persisting them across sessions. The ajail tool comes with scripts for various Linux distributions, making it easy to set up and use.

🌎 Environment & Sustainability

Evidence of the bouba-kiki effect in naïve baby chicks

Scientists have discovered evidence of the “bouba-kiki” effect in naive baby chicks, indicating that young birds may also perceive sounds associated with certain shapes as pleasing or unpleasant. The study found that chicks reacted differently to sounds when presented with bouba-shaped objects versus kiki-shaped objects. This discovery sheds new light on the universality of the human perception phenomenon.

Tracing Chinese Immigration Through Food (2023)

A woman attempts to recreate her late grandfather’s traditional Chinese recipe for chop suey, a dish that holds nostalgic value due to its connection to his restaurants in Canada and their family history. However, she discovers that the recipe is vague and has undergone significant changes over time, reflecting the evolution of Chinese cuisine in North America. Ultimately, she finds comfort in the simple taste of the dish rather than a deeply emotional or revelatory experience.

She Graduated with Honors but She Can’t Read

A US student, Aleysha Ortiz, graduated with honors but is illiterate due to inadequate teaching in her school district. Mississippi’s literacy-based promotion act has shown promising results by improving reading scores for low-performing students. The author suggests that stricter standards are needed nationwide to ensure all students receive a proper education and that test-optional policies may be masking educational failures.

Greenland ice melt surges unprecedentedly amid warming

Climate change has led to a significant increase in extreme episodes of melting in Greenland’s ice sheet, with the area affected by these events rising by 2.8 million km² per decade since 1990. The production of water from melting has also increased sixfold, with future projections suggesting it could triple by the end of the century. This intensification of melting events poses significant global environmental and geopolitical implications.

The “Enshittification” of Consumer Products

Consumer products, such as packaged foods, have become less desirable due to cost-cutting measures and increased profit-seeking by companies. This “enshittification” trend is evident in industries like food production, where additives and preservatives are used to reduce costs, leading to a decline in nutritional quality. Many products now contain fewer high-quality ingredients than they once did.

A large study of 460,000 teenagers in Northern California found that cannabis use among adolescents increases the risk of developing bipolar disorder, psychotic disorders, anxiety, and depression years later. Teens who started using cannabis before age 25 had twice the risk of serious mental illnesses compared to non-users. The study’s findings suggest caution is needed when it comes to cannabis use among teenagers due to its potential neurological effects.

Over 80% of 16 to 24-year-olds would vote to rejoin the EU

An ITV poll found that 80% of 16-24 year olds would vote to rejoin the EU in a new referendum, contrasting with the 2016 Brexit result where 52% voted to leave and 17% opted to stay out. The Liberal Democrats and Greens are currently campaigning for EU re-entry, while Labour aims for “deeper economic integration” with Europe. This outcome contradicts expectations that the UK’s younger generation would support the country leaving the EU.

EU mandates replaceable batteries by 2027 (2023)

The European Union is introducing a new law that mandates the replacementable battery market by 2027, requiring collectors, reusers, and recyclers to have low carbon emissions, minimal harmful substances, and high recycling rates. The Batteries Regulation aims to support Europe’s circular economy and increase security of supply for raw materials and energy.

Irish man detained by ICE [Update] – It’s not what it seems

An Irish man, Seamus Culleton, has been detained by ICE after overstaying his US tourist visa 20 years ago. His estranged daughter, Heather Morrissey, expressed “no sympathy” for her father’s situation and hopes he is deported back to Ireland due to their strained relationship. Morrissey had previously chosen not to pursue a reunion with her father, whom she claims abandoned her as a teenager.

ATAboy is a USB adapter for legacy CHS only style IDE (PATA) drives

ATAboy is a hardware device that enables users to connect and read from older IDE hard drives without a vintage computer. The device translates USB Mass Storage commands into IDE read/write operations, working with classic IDE (PATA) hard drives on Windows, Linux, and macOS devices. It is designed for retro computing enthusiasts, data recovery experts, and archivists.

Sarvam launches Indus chat app with AI-powered model

Indian AI startup Sarvam has launched its Indus chat app, powered by its Sarvam 105B model, allowing users to interact with the large language model through text and audio responses. The app is currently available in beta on iOS, Android, and the web with limited features and access, targeting India’s local languages and users.

The Future of Math Research in the Age of AI

A group of mathematicians has released a project called First Proof to assess AI’s ability to autonomously solve research-level math questions. The first batch of problems was solved by both human mathematicians and AI systems, with some relying on human input or collaboration. The team plans to release more problems in the future to further test AI capabilities.

Hemmi/Post 1460 Versalog (Sliderule emulator)

A new software emulator, Hemmi/Post 1460 Versalog (Sliderule), has been developed to replicate the functionality of vintage calculators, allowing users to perform mathematical calculations using a unique keyboard layout and set of operators. The emulator is compatible with various input formats, including numbers, symbols, and mathematical functions. It aims to preserve the nostalgic experience of classic calculator users.

Loon — A LISP that flies

A new programming language called Loon, based on LISP, has been created and can run entirely within a web browser using JavaScript. The language is developed by Camped Andersen.

ThunderKittens 2.0: Even Faster Kernels for Your GPUs

ThunderKittens 2.0 is a new version of the project with improved performance, efficiency, and speed through major refactoring and the addition of new features such as MXFP8/NVFP4 support and tensor memory controllability. The release highlights technical learnings on optimizing GPU kernels, including the importance of memory consistency, tensor core and memory pipelining, and properly hinting the PTX assembler. The update brings state-of-the-art GEMM kernels that surpass or match cuBLAS performance on Nvidia B200s.

The Fastest Way to Board an Airplane

Airline research simulations have found alternative methods to traditional back-to-front boarding strategies, such as random or zone-based boarding, to be faster and more efficient. The window-middle-aisle (WilMA) method, which groups passengers by column of seats, is slightly faster than traditional back-to-front boarding. These methods aim to reduce congestion and wait times on airplanes.