Search the archive

AWS DESTROYED THE VALUE PROPOSITION FOR BEDROCK

AWS Bedrock sat as the neutral place between companies and model providers, but now it is just first-party Anthropic with fewer features.

GOOGLE CHROME IS KILLING ALL UBLOCK ORIGIN BYPASSES, MICROSOFT EDGE, OPERA TO FOLLOW

Chrome is looking to permanently drop Manifest V2 extensions and its bypasses.

EVERYTHING IS RECORDED NOW

Most work discussions are now being recorded by default due to businesses implementing AI.

Get the CISO's Guide to AI and Agentic Traffic

POLICY ON THE AI EXPONENTIAL

AI advances rapidly, outpacing slow policy-making, leading to potential risks in cybersecurity and job displacement. Anthropic recommends a regulatory approach like the FAA for AI, mandatory testing, and stronger security standards to address these risks. Additional policy areas include adapting macroeconomics and tax systems for AI-driven growth, reforming regulations to enhance AI's biomedical impacts, and ensuring AI aligns with democratic values globally.

DIFFUSIONGEMMA: 4X FASTER TEXT GENERATION

DiffusionGemma, a 26B Mixture of Experts model, uses text diffusion for faster text generation, offering up to 4x speed increase on GPUs by generating text blocks simultaneously. It targets speed-critical applications, providing advantages like low latency and bi-directional attention, and fits in high-end consumer GPUs when quantized. While it trades some quality for speed compared to standard models, it allows for efficient local inference and is optimized for NVIDIA hardware.

DON'T LET THE LLM SPEAK, JUST PROBE IT

Answers often already exist in LLMs' hidden state before they generate a single token. It's possible to skip generation entirely by grabbing the hidden state at the last prompt token, feeding it to a tiny multilayer perceptron, and calibrating the output. This results in a frozen model that acts as any classifier that can be written in English.

MEASURING LLMS' IMPACT ON N-DAY EXPLOITS

N-days are vulnerabilities that have already been publicly disclosed, but only patched on some devices. They can be more dangerous than zero-day vulnerabilities as the patches provide a roadmap to the bugs. The process of reverse-engineering a vulnerability from a patch has historically been slow and specialized work, but AI can accelerate and automate the process. This means anyone in the patch gap today faces a much larger threat than before.

MOATS NEED MODELS

The model, harness, workflow, and evaluation loop are no longer separate stack pieces but co-design surfaces that compound together. Defensibility comes from owning the full feedback loop rather than renting frontier capability that can be restricted, repriced, or reclaimed on the supplier's terms.

FABLE-5 SYSTEM PROMPT LEAK

The system prompt for Anthropic's Fable 5 model has been leaked. It comes to a total of around 120,000 characters. The entire leaked prompt is available in the thread.

FASTER CODE REVIEW WITH CURSOR'S BUGBOT

Cursor updated Bugbot to run over 3x faster, cost 22% less, and find 10% more bugs per review. Most runs now finished in under three minutes, shortening the review loop for developers.

PALANTIR'S KARP SAYS BUSINESSES ARE ‘UNHAPPY' WITH THE FRONTIER AI LABS

Palantir's CEO, Alex Karp, says the company's enterprise customers are unhappy with how frontier AI labs are operating. He claims these labs only care about burning through AI tokens to signal productivity. Accelerating costs are raising alarm among businesses and fueling efficiency concerns.

EU ORDERS META TO STOP BLOCKING RIVAL AI CHATBOTS ON WHATSAPP

The EU has ordered Meta to open up WhatsApp to chatbots from rival companies for free. Meta banned third-party AI chatbots from the WhatsApp for Business API in October last year. The EU said Meta was abusing its dominant position in the European messaging app by preventing competing assistants from using the WhatsApp API. Meta plans to appeal the decision, saying the order is a regulatory overreach that would grant some of the world's largest companies access to the API without paying.

real-world context across their organizations. Download the guide

OPENAI WEIGHS NVIDIA-BACKED LEASE FOR 10 GW OHIO DATA CENTER CAMPUS

OpenAI would control the computing equipment at the site, which is expected to begin operating in 2028, under a 20-year lease.

INTRODUCING RAMP APPLIED AI SOLUTIONS

Ramp Applied AI Solutions enables companies to deploy AI effectively by embedding engineers in finance teams to create custom solutions.

THE EVOLUTION OF AGENTIC SURFACES: BUILDING WITH CLAUDE MANAGED AGENTS

Claude Managed Agents streamline building production-grade agents by offering composable APIs with integrated infrastructure.

CODEX FOR BLACK HOLE SIMULATIONS

Astrophysicist Chi-kwan Chan used Codex to refine and test algorithms for simulating plasma and particle behavior around black holes.

APPLE'S ALL-NEW IMAGE PLAYGROUND PROMISES MORE THAN CARTOONS

Apple has significantly upgraded Image Playground in iOS 27, iPadOS 27, and macOS 27, adding support for photorealistic image generation through a new Apple Intelligence model running on Private Cloud Compute. Users can now create images in virtually any style, make targeted edits by describing or selecting areas to change, and use generated images for wallpapers, Contact Posters, Messages, and more. Apple is also continuing to label AI-generated content with hidden SynthID watermarks to improve transparency.

MAI-IMAGE-2.5 LAUNCHES AT NO. 2 FOR IMAGE EDITING ON ARENA

Microsoft launched MAI-Image-2.5, its strongest image model yet, now ranked No. 2 on Arena's image-editing leaderboard, surpassing GPT-Image-1.5 and Nano Banana Pro 2K. The model supports precise, localized edits — including object replacement, background cleanup, and face identity preservation — and is already powering features in PowerPoint and OneDrive. Developers can access it via Foundry and OpenRouter, with a faster, cheaper Flash variant also available.

LOVABLE: $500M ARR, 146 STAFF, 80% NON-TECHNICAL BUILDERS

Lovable reported $500 million in annual recurring revenue with only 146 employees, claiming 80% of its builders are non-technical users. The platform hosts over 50 million projects that receive 720 million monthly visits, with 8 in 10 users planning to monetize their creations. However, all data is self-reported and unaudited, raising questions about the platform's ability to maintain enterprise-grade support and whether apps built with natural language can scale effectively in production environments.

DESIGN SYSTEMS ARE OVER. PRODUCT CONTEXT IS THE WORK

Design systems as currently defined no longer match the evolving work needed, especially as AI reshapes product development. While the foundational components remain essential, AI requires broader product context rather than isolated design elements to maintain coherence at scale. The focus must shift from traditional design systems to comprehensive product context that includes the decisions, constraints, and reasoning that guide both human and AI outputs.

WRITE-FIRST DESIGN

Most design teams make decisions in Figma, where directions accumulate into plans without explicit agreement — but the author's team writes decisions before opening design files. Writing exposes flawed reasoning that mockups can hide, separates decisions from craft execution, and creates durable documentation that keeps teams aligned. The process uses written artifacts like discovery docs and design peer reviews as checkpoints, with decisions happening in written threads rather than meetings.

THE FLAW IS THE FEATURE

As AI makes polished, technically flawless work easier to produce, perfection alone is no longer enough to stand out. Research suggests people value creations more when they perceive human effort behind them, and that small imperfections can make otherwise competent work feel more authentic, relatable, and memorable. Rather than eliminating every irregularity, creators may benefit from preserving the subtle human touches that AI tends to smooth away.

UNLEASH YOUR IDEAS WITH ASCII

MonoSketch is a powerful open-source ASCII sketching and diagramming app that transforms ideas into visually stunning designs using ASCII characters.

DITHER IMAGE ONLINE

This online tool allows users to apply dither effects to images for free with real-time processing. Users can choose from various dithering algorithms, including Bayer, crosshatch, halftone, and contour patterns.

JAVASCRIPT 3D RIGID BODY PHYSICS

Crashcat is a JavaScript library for 3D rigid body physics simulations. The library provides physics engine capabilities for web-based 3D applications and games.

LAWFUL DESIGN

Legal requirements and design constraints should work together rather than against each other, with lawyers and designers collaborating early in the process instead of retrofitting legal elements after a design is complete. Examples like consent agreements and cookie banners show that the most ethical, low-friction, and legally defensible solutions are often one and the same when both disciplines are properly aligned. Legal phrases like "clear and conspicuous" are actually design specifications that call for thoughtful information hierarchy, rather than simply piling on more disclaimers or fine print.

USEFUL AI SKILLS AND WORKFLOWS FOR DESIGNERS

Designers can now encode expert thinking into reusable AI skills — curated collections from professionals like Jamie Mill, Marie-Claire Dean, and others covering product design, UX research, accessibility, and motion. Rather than one-off prompt libraries, these resources represent decision-making infrastructures that guide AI to follow specific ways of thinking through design problems. The real value comes from adapting these skills to fit individual workflows, team dynamics, and the specific challenges each designer faces daily.

BRANDON REDESIGNS HERITAGE PASTRY BRAND JUS-ROL TO COUNTER OWN-LABEL COMPETITION

Jus-Rol has refreshed its brand identity to better differentiate itself from own-label rivals, building on its heritage with a bolder logo, vibrant colour palette, and new sunburst graphic. The redesign improves shelf standout and product navigation while celebrating the fun and authenticity of home baking through simplified packaging and more natural food photography.

APPLE IS EMBRACING THE FANTASY OF AI PHOTO EDITING

Apple's updated Image Playground app can now generate photorealistic images and modify photos using text prompts.

‘SEE IN CMYK' WITH GOOGLE ARTS AND CULTURE AND THE EXPLORATORIUM

Google Arts & Culture and the Exploratorium collaborated with artist Stefanie Posavec to create "See in CMYK," an interactive project that uses AI to transform photos into colorful art made of custom icons.

Start with our all-access 30-day trial. SIGNUP HERE →

HOW BUILDING AN HTML-FIRST SITE DOUBLED OUR USERS OVERNIGHT

Rebuilding a utility company's failed React application into an HTML-first site using Astro instantly doubled its user completion rates overnight. Focusing on progressive enhancement, lightweight code, and backend data storage made the user experience much better for low-spec devices and poor connections.

INSTANT BY DEFAULT, MULTIPLAYER BY DESIGN

Plain is a developer platform that replaces the traditional request-response model with a sync-based architecture that maintains a live replica of the database within the browser. The system uses two specialized engines to handle different data types, using ElectricSQL for Postgres rows and Yjs for character-level concurrent document editing. By subscribing to live state instead of individual endpoints, every view stays updated automatically without the need for manual page refreshes.

POLICY ON THE AI EXPONENTIAL

Anthropic CEO Dario Amodei writes that rapid technological progress is currently outpacing legislative capacity, creating an urgent need for mandatory safety testing and government oversight of frontier models to prevent catastrophic risks in cybersecurity and biology. Addressing the potential for huge labor displacement and hyper-inequality requires a multi-faceted economic response.

THE UNWRITTEN LAWS OF SOFTWARE ENGINEERING

Software engineering relies on unwritten principles learned through experience that prioritize system stability and proactive risk mitigation. When production issues occur, the most effective approach is to perform an immediate rollback before debugging and to make sure that all backups have been verified through actual restoration tests. Engineers must also plan for the inevitable failure of external dependencies.

HOW TO BE GOOD AT RESEARCH

True research mastery requires moving beyond just imitation by deliberately training specific skills, starting with the selection of original problems rather than simply following current trends. Diversifying information inputs by studying foundational papers and reading full documents instead of summaries helps build the unique insights necessary for genuine innovation.

EXTEND UI

Extend UI is an open-source toolkit designed for building modern document applications with integrated file viewers and management features. The library includes specialized components for handling various formats like PDF, DOCX, and XLSX, alongside tools for e-signatures and file uploads.

APACHE BURR

Apache Burr is an open-source Python framework for developing decision-making applications, such as chatbots and agents, using a state machine architecture. The system allows devs to model complex workflows as graphs of simple Python functions, providing a clear structure for managing state and transitions.

HELIX DB

HelixDB is an OLTP graph-vector database written in Rust that unifies vector, graph, relational, and document data into one system to simplify AI memory management. It provides a CLI and SDKs for local development while offering a managed cloud service with high availability, ACID transactions, and auto-scaling.

SURPRISE! PAY $1,000

A development team seeking a faster alternative to GitHub Actions tested Blacksmith, a service offering a free trial with no credit card requirement. After exceeding the trial limits, the users unexpectedly received an invoice for over $1,000 because the service continued running rather than automatically suspending as anticipated.

THE IPAD WAS ON TAILSCALE

An investigation into why a WebRTC-based application failed exclusively on an iPad revealed a complex interaction between two separate software bugs. The first issue involved a hardcoded constant in the webrtc-rs library that generated data packets exceeding the maximum size allowed by Tailscale's VPN tunnel. Although these oversized packets were fragmented to fit the tunnel, Tailscale's policy of dropping all IPv6 fragments prevented the data from ever reaching the destination.

WHAT BUN CAN TELL US ABOUT AI, OPEN SOURCE, AND ANTHROPIC

Anthropic acquired Bun's maker Oven in December 2025, and the runtime has kept growing fast (npm installs up ~16X to 7.3M/month). However, two shifts stand out: AI now writes most commits (peaking above 80%), while human contributors have roughly halved, with most of the original team gone and external contributions falling sharply.