Search the archive

Learn more about evals →

AGENTIC CODING AND PERSISTENT RETURNS TO EXPERTISE

A clear division of labor exists in agentic coding where humans make the majority of planning decisions while the AI handles technical execution. Success in these tasks depends more on the user's domain expertise than their formal programming skills, allowing professionals in various non-technical fields to get results comparable to those of software engineers. Recent trends show a shift toward complex, end-to-end agentic work and a measurable increase in the estimated economic value of the tasks being completed.

MAKING BUDGET MODELS PUNCH ABOVE THEIR WEIGHT WITH A SMART HARNESS

Dirge is a highly efficient coding harness developed in Rust that improves budget model performance while still having a minimal memory footprint. Unlike traditional harnesses that focus solely on model capabilities, Dirge's supportive framework, with features like error correction, session continuity, and a memory management system help optimize model quality across different contexts.

HOW WE MADE CLOUD BROWSERS 3X CHEAPER AND FASTER

Browser Use's cloud browsers have been optimized to start quickly, remain isolated, and maintain low costs, now operating at $0.02 per hour with sessions starting in under a second. This was done by using Firecracker for lightweight virtual machines while implementing a control plane for dynamic scaling.

THE DIALOGUE DIVIDEND

Conversations with others often lead to greater problem-solving and understanding than solitary thinking can provide as they introduce a feedback loop that refines thoughts in real-time.

BUT YAK SHAVING IS FUN

Yak shaving refers to the process of engaging in a series of tangential tasks that divert from the original goal, showing how building projects from scratch can lead to unexpected challenges, discoveries, and joy.

INTRODUCING EVE

eve is an open-source agent framework that makes building, running, and scaling agents easier by providing built-in production capabilities like durable execution, sandboxed computing, and human-in-the-loop approvals. With a clear structure, agents can be defined through a directory of files that dictate their components, actions, and connections.

LORE

Lore is an open source version control system designed for scalability, particularly for projects with large binary assets like games. It has a centralized, content-addressed architecture that allows for efficient data storage and retrieval, with an emphasis on immutable revisions and low-overhead branching.

BUILDING TREX: CODE EXECUTION AND ARTIFACT GENERATION FOR AI CODE REVIEW

TREX is a code review tool that integrates code execution with AI-driven testing to identify bugs that may not be apparent in static analysis alone. It uses a multi-agent architecture, where a Greptile reviewer orchestrates specialized TREX agents that run tests in isolated environments, providing artifacts such as screenshots and logs to verify findings.

IMAGE COMPRESSION

Image compression techniques, including lossy and lossless methods, exploit characteristics of human visual perception to reduce file sizes. Formats like JPEG use a combination of chroma subsampling, discrete cosine transforms, and quantization for compression without noticeable quality loss, while GIF uses indexed color palettes and Lempel-Ziv-Welch algorithms to store limited color data.

BRAIN THE SIZE OF A PLANET: ARE LLMS THONKING TOO HARD?

An analysis of various LLMs in triaging security vulnerabilities showed that a council of four LLMs achieved an impressive 86.2% unanimous agreement on findings, while lower reasoning efforts consistently performed poorly. Although higher reasoning generally correlated with better results, there were exceptions where models struggled with complex tasks.

THE FOUNDER'S PLAYBOOK: BUILDING AN AI-NATIVE STARTUP

Anthropic has released a playbook of practical exercises and agentic workflows to navigate every stage of the business lifecycle for AI-native startups.

MAC-OCR

mac-ocr is a macOS command-line tool that uses Apple's Vision framework locally to recognize text from images and PDFs and create searchable PDFs.

CAN GZIP BE A LANGUAGE MODEL?

Gzip can be used as a language model through compression techniques that predict and generate text continuations by scoring potential outputs based on how well they compress with a given context.

EXA AGENT

Exa Agent is a new API that combines frontier language models with optimized web search tools to provide high-performance research, list-building, and data enrichment capabilities.

Give your agent an inbox →

DON'T RUN SQL MIGRATIONS IN TESTS: HOW I SPED UP THE TEST SUITE BY 2X

Large Go test suites often face slowdowns when thousands of SQL migrations are applied serially for every individual test. To address this, a content-addressing system identifies a "golden" database file by hashing the contents of all required migration scripts. If a matching database already exists, the system clones it instantly using a backup API, whereas missing versions are lazily generated and cached for subsequent runs.

BRITISH COLUMBIA, TIME ZONES, AND POSTGRES

British Columbia's shift to a permanent UTC-7 offset in 2026 shows the risks of storing future appointments in UTC-based database columns, which can lead to unexpected one-hour shifts in local time. These columns convert timestamps based on the rules available at the time of entry, so subsequent updates to time zone data can cause the stored UTC values to no longer match the user's intended wall-clock time.

SVGS AND PDFS CAN BOTH BE INTERACTIVE

Both SVG and PDF file formats have underused specs that allow for the integration of interactive layers and programmable scripts directly within a document. While SVGs can host self-contained applications using embedded JavaScript, PDFs use a specialized API that often struggles with compatibility across different viewing software.

MYTHOS AND FABLE CAN MAKE US ALL SAFER. SHUTTING THEM DOWN IS RECKLESS

The United States government recently used export-control authority to force the shutdown of powerful AI models like Mythos 5. This unilateral action is considered reckless by many security experts because these models are necessary tools for defenders to identify and patch vulnerabilities in global software infrastructure. Since comparable AI capabilities are already being released by international labs, domestic shutdowns do not eliminate risks but instead deprive local developers of the resources needed to counter emerging threats.

STOP USING JWTS

JSON Web Tokens should be avoided for managing user sessions because they are insecure and were never intended for long-term authentication. Traditional cookie sessions are a more reliable alternative, with better flexibility and protection against common security vulnerabilities. Truly stateless authentication is largely impractical, as secure token management usually requires a persistent data store to handle revocations and expiration properly.

THE WEB WE KNOW IS GOING TO DISAPPEAR

The evolution of technology has seen information move from physical disks and bulletin boards to the open Web and mobile apps, and it is now shifting toward centralized AI interfaces. This new model replaces traditional search-and-click habits with direct, conversational answers. As AI becomes the primary gatekeeper, websites are transitioning from being human destinations to serving as background infrastructure that feeds data into machine learning models.

Get started

NUBASE

Nubase is an open-source, AI-native backend platform that transforms AI-generated code into live applications fast. The system integrates eight core modules, including database, authentication, storage, and edge functions, into a single self-hostable service. MCP tools allow coding agents to directly manage infrastructure, deploy logic, and handle data modeling.

LIKEC4

LikeC4 is an open-source “architecture-as-code” platform that lets engineers describe software systems in a DSL and automatically generate live, always-up-to-date architecture diagrams from that code.

WHY IS META DESTROYING ITS ENGINEERING ORGANIZATION?

Meta has quickly dismantled its historically strong engineering culture by aggressively prioritizing AI initiatives, such as forcing engineers into data-labeling work, tracking employee activity, tying performance reviews to AI usage, and reducing autonomy that once made the company attractive. These changes have led to low morale, talent flight, organizational chaos, and even major security and reliability failures, including a high-profile Instagram account takeover incident allegedly linked to AI-generated and insufficiently reviewed code.

SPACEX LOCKS IN $60 BILLION CURSOR DEAL TO CLOSE GAP WITH RIVALS IN AI CODING RACE

SpaceX has agreed to a $60 billion all-stock deal to acquire Anysphere, the startup behind the popular AI coding agent Cursor. This move will improve the company's presence in the enterprise AI market and narrow the gap with industry leaders like OpenAI and Anthropic. It will also improve xAI's Grok.

CLAIREON

Claireon is an open-source Unreal Engine plugin with an MCP server inside the editor, allowing AI agents like Claude Code to search, inspect, modify, and automate nearly every aspect of Unreal development.

HUMILIATING IIS SERVERS FOR FUN AND JAIL TIME

This is a bug-bounty-focused guide to finding and testing Microsoft IIS servers, as IIS environments are often overlooked yet frequently expose misconfigurations, information leaks, and attack paths that can lead to sensitive data exposure.

MAKING HTTP REQUESTS FROM A CONTAINER THAT HAS NO CURL, USING BASH /DEV/TCP

Bash allows for manual HTTP requests and connectivity checks in minimal containers by using its internal /dev/tcp redirection to open and communicate over TCP sockets without needing external tools like curl or wget.

REVIEWS HAVE BECOME EXPENSIVE, REWRITES HAVE BECOME CHEAP

The tendency of AI to produce over-engineered code has made manual reviews more costly, so upfront planning is more necessary now.

Register now

EVERY FRAME PERFECT

The "every frame perfect" philosophy says that user interfaces should remain logical and explainable at any point in time, even during rapid transitions or animations. High-quality UI polish acts as a heuristic for code quality, building user trust by demonstrating that developers have invested time into every detail of the experience. Many applications currently suffer from visual glitches like white flashes, content layout shifts, and unsynchronized animations that occur between the start and end states of an interaction.

TIMESCALEDB COMPRESSION: HYPERCORE AND COLUMNAR STORAGE WITH UP TO 98% RATIO IN POSTGRESQL

TimescaleDB uses a hybrid row-columnar engine called Hypercore to achieve up to 98% compression for time-series data within PostgreSQL. This technology uses specialized algorithms like delta encoding, Gorilla XOR, and run-length encoding to optimize cross-row patterns that general-purpose databases often ignore. Data must be properly organized using specific parameters that group similar information into compressed batches of approximately 1,000 rows.

HOW DROPBOX USES MCP AND DASH TO CLOSE THE DESIGN-TO-CODE SECURITY GAP

Dropbox built an agentic system to close the “design-to-code security gap.” It pairs Dash (which indexes threat models), MCP (giving a review agent access), and a foundational model that compares the threat model against the pull request to flag gaps, surfaced right inside code review.

AI DEMANDS MORE ENGINEERING DISCIPLINE. NOT LESS

YES, WE STILL NEED ENGINEERS

While AI-driven prototyping accelerates initial software design, professional engineers are still needed for transforming these prototypes into secure, scalable, and production-ready systems.

COMPONENT PARTY

Component Party provides a comparative overview of various JavaScript web component frameworks by showcasing their unique syntaxes and features through clear, educational examples.

MESHERY

Meshery is an open-source, cloud-native management platform that simplifies the design and operation of Kubernetes-based infrastructure across multi-cloud environments. It has a visual, collaborative GitOps interface that allows users to manage deployments and infrastructure without relying solely on manual YAML configuration.

A BACKDOOR IN A LINKEDIN JOB OFFER

A developer was targeted on LinkedIn by a fake recruiter who requested a code review for a malicious GitHub repository. Investigation in a sandboxed environment showed a hidden backdoor designed to execute remote payloads automatically when project dependencies are installed. The scheme used stolen identities of a legitimate journalist and a software engineer to establish trust and lure the victim into running harmful scripts.

ANTHROPIC'S SAFETY SUPERPOWER

Anthropic's release of advanced models like Fable and Mythos prompted government-mandated access restrictions due to security risks, leading the company to adopt stricter data policies and defensive competitive safeguards. This strategic shift positions Anthropic to exert significant influence over AI development by aligning its commercial interests with internal control mechanisms.

A GUIDE TO AI INFERENCE ENGINEERING

LLM inference engineering can be split into two phases with opposite hardware bottlenecks. Prefill is compute-bound, processes the prompt to emit the first token, and is measured by time-to-first-token. Decode is memory-bandwidth-bound, generates each subsequent token sequentially, and is measured by tokens-per-second. This post walks through six tradeoffs on inference: batching, prefix caching, quantization, speculative decoding, parallelism, and disaggregation.

WHAT JOB INTERVIEWS TAUGHT ME ABOUT KUBERNETES

Kubernetes has become the industry standard for companies of all sizes because it offers organizational benefits like deployment uniformity, standardized hireable knowledge, and improved traceability.