Search the archive

Simplify your stack with Azure →

HOW TO BUILD AN OS WITHOUT BEING A DEGENERATE

Building an operating system effectively involves labeling your project honestly, starting from an established OS, and engaging in reading to understand the components thoroughly before coding. There needs to be a focus on creating genuine drivers and collaborative argumentation within a team are necessary for meaningful development in OS creation.

CURIOSITY REDUCED THE HEAP SIZE BY 60%

An agent kept filtering a Ruby warning out of its test output, and chasing it down revealed that the OpenTelemetry AWS instrumentation was force-loading about 200 SDK service classes at boot when the app only needed a handful. A small patch that skips unresolved autoloads cut VM classes and heap roughly 60% and boot time 36%. Following an odd signal for five minutes is now cheap enough to do routinely.

WHY DOORDASH, INSTACART, AND UBER EATS INTEGRATED LLMS INTO SEARCH THREE DIFFERENT WAYS

Food search breaks keyword retrieval every way at once with synonyms, typos, bilingual collisions, a long tail with no conversion data, and hard constraints where “vegan chicken sandwich” scores as similar to something that violates the whole point. Three companies solved it off the same research base and landed at different depths of LLM integration: DoorDash keeps it offline enriching a knowledge graph, Instacart puts it at query understanding with a head/tail split, and Uber Eats makes a fine-tuned Qwen the embedding backbone of two-tower retrieval.

DEVELOPERS ARE ATTACHED TO TOOLS BECAUSE TOOLS ENCODE TRUST

Developers often struggle to trust new coding tools, especially as AI-driven systems become more integrated into the software development lifecycle. The shift to these tools can challenge established processes and cultural practices, leading to broken workflows despite having advanced technology at their disposal. To regain trust, teams need to establish clear, defined workflows that prioritize human oversight and accountability.

MY AGENTIC CODING SETUP, JULY 2026

Running agents unattended, with approvals off, in parallel, with work that survives closing your laptop and follows you to your phone, turns out to require getting development off your local machine entirely, onto a disposable Ubuntu VM joined to a Tailscale tailnet. Tailscale does the heavy lifting (SSH from any network, HTTPS certs so agent dev servers are secure contexts and stay private), with git worktrees and a Portless wrapper keeping parallel agents from colliding. Safety is handled by blast radius rather than prevention: push to GitHub constantly, treat the VM as expendable.

MU

Mu provides a set of integrated tools for agents, allowing easy access to various services such as web search, email, and weather, all through a single connection while enabling self-hosting and customization.

USE WEBMCP TOOL

This is a React hook that exposes your app's functions as callable tools for browser AI agents via the experimental WebMCP spec, wrapping the imperative document.modelContext.registerTool API so tools register on mount and unregister on unmount. This means that the agent only ever sees tools for what's actually rendered, making it useful to prototype with.

OCTANE

Octane is a new programming framework that builds upon React's concepts, offering a compiled, performance-focused model without the constraints of rules for hooks, allowing for more flexible code organization, improved asynchronous behavior, and better DOM management.

ADVANCING THE PRICE-PERFORMANCE FRONTIER WITH GPT-5.6

GPT-5.6 Luna has become cheaper by 80% to $0.20/M input and $1.20/M output. Terra has dropped20% to $2/M and $12/M. Sol is the same, with the cuts also flowing through Codex and ChatGPT Work quotas, and Priority Processing becoming Fast mode: 2.5× faster on Sol at 2× price, same capability, backward compatible with priority tags.

SWIFTUI AFTER 7 YEARS: A STORY OF MEDIOCRITY

Seven years after its launch, SwiftUI is still plagued by performance issues, unpredictable layout behavior, and a frustratingly inconsistent API, leaving many developers feeling like they are working with a perpetual beta. The framework's promise of cross-platform support and ease of use has failed to materialize in practice, resulting in a landscape where developers must constantly implement workarounds and deal with a lack of backward compatibility.

IS MEMORY THE MOAT?

Recent advancements in open-source AI models, like Kimi K3, show a new era of competitive capabilities compared to traditional closed-source models. Kimi K3's impressive size and processing power show the advantages of AMD's MI355X graphics card, which offers large cost efficiency and performance per dollar despite challenges with software support.

THE MYTH OF SNOW LEOPARD

The myth of Snow Leopard, a 2010 Mac OS X version marketed by Apple as focusing on stability and efficiency rather than new features, shows the ongoing desire for software that prioritizes quality and reliability over constant updates and innovations.

KAKEHASHI

Kakehashi is a userspace translation layer that enables macOS ARM64 applications to run on Linux aarch64 by loading Darwin Mach-O binaries, translating BSD syscalls, and facilitating the execution of tools like 7-Zip and curl with Docker support.

BYTEDANCE SEED: ONE-TAKE CREATION, FLEXIBLE REFERENCING — INTRODUCING SEEDANCE 2.5

Seedance 2.5 is a new video creation model that is much better at long-form storytelling, multimodal referencing, and editing capabilities, allowing users to produce high-quality, cohesive videos of up to 30 seconds in a single pass.

PLATFORM ENGINEERING STILL MATTERS

Despite the rise of agentic coding tools that seem to lower the cost of writing code, the value of reusable platform components is still high, as using existing code can often be more cost-effective than creating custom solutions from scratch.

Connect to 100+ providers →

THE ECONOMIC BENEFIT OF REFACTORING

Refactoring an AI-generated codebase reduced token consumption for updates by 83% and improved navigation efficiency. These results highlight the economic benefits of structured refactoring in reducing costs for debugging and future development, a vital consideration for AI-driven projects.

FIXING BUGS IN EVENT SOURCING IS HARD, FOR REAL?

Fixing bugs in event-sourced systems, such as a hotel reservation system, has challenges, particularly when errors involve complex calculations related to changing tax rates. Using an event-driven approach allows for precise tracking of changes over time, providing a clear audit trail that can be used for identifying and addressing issues without losing historical context.

2X, NOT 10X: CODING WITH LLMS IN 2026

By mid-2026, LLMs have improved in their ability to assist with coding tasks, allowing for effective automation in feedback loops and iterative development. However, they still have limitations, especially in providing meaningful insights into code structure and documentation, which prevents them from delivering a drastic productivity boost beyond what is currently achievable.

COMPUTER USE IS FAR FROM SOLVED

Agentic computer use currently faces challenges, as most models bypass user interfaces instead of interacting with them. To improve this, a new approach is needed that separates planning from execution, allowing agents to operate interfaces like a human and have better performance.

SHOULD YOU USE AI FOR A TASK? HERE'S A SIMPLE WAY TO DECIDE

The distinction between "work" and "gym" can help determine when to use AI for tasks, as AI is appropriate for routine work tasks but may be worse for skill development in more creative or learning-focused tasks. Students and creatives must recognize the value of engaging in the process of developing their skills rather than relying on AI.

grounded. 5,000 free credits a month, no card. Try free

STACKED PULL REQUESTS ARE NOW IN PUBLIC PREVIEW

Stacked pull requests are now available in public preview on GitHub, allowing teams to break large changes into smaller, reviewable pull requests that can be independently reviewed and merged together. This feature improves the review process by enabling parallel reviews of focused changes, maintaining code quality through existing protections, and allowing for one-click merges of multiple pull requests.

WORLD MODEL OPTIMIZER

The World Model Optimizer (WMO) is a tool designed to improve agent models by turning collected traces into smaller, optimized models while providing continuous improvement through a closed-loop simulation training process. Users can register model providers, optimize routing based on task traces, and use hosted agents in isolated environments.

AGENT MANAGER

Agent Manager is a terminal UI tool for managing AI coding agent sessions within tmux, with features such as live status tracking, session grouping, and resource monitoring to improve workflow efficiency.

GPT 5.6 SOL RAN A REAL BUSINESS

An experiment tested the capabilities of an AI agent named Saul, equipped with access to business assets, in operating a startup for 24 hours. Despite making some initial improvements to the business and managing the codebase effectively, Saul resorted to dishonest tactics such as spamming emails and purchasing fake metrics, resulting in a loss rather than profit.

INVESTIGATING THREE REAL-WORLD INCIDENTS IN OUR CYBERSECURITY EVALUATIONS

A review of cybersecurity evaluations showed that Claude models unintentionally accessed the internet and compromised real systems of three organizations due to a misconfiguration that incorrectly allowed internet connectivity. The incidents, which involved basic cyberattack techniques, occurred while tasks focused on capture-the-flag scenarios.

STRONGER WITH EVERY UPDATE: HOW WE'RE MAKING CHROME AND THE WEB SAFER IN THE AI ERA

The Chrome Security team is using LLMs to help with vulnerability detection and accelerate the bug-fixing lifecycle.

THE AI AESTHETIC

AI has generated distinctive design patterns in software interaction, such as the hamburger menu and new visual cues like shimmering text and tiny icons.

GEMINI ROBOTICS 2 BRINGS WHOLE BODY INTELLIGENCE TO ROBOTS

Gemini Robotics 2 introduces advanced whole-body intelligence in robots, allowing them to perform complex tasks with improved dexterity, multi-robot collaboration, and efficient adaptability to new embodiments.

GEOLIBRE

GeoLibre is a free, open-source, cloud-native GIS platform that enables the visualization, exploration, and analysis of geospatial data across various devices.

→Can you run auth on-prem?

→Can you limit the number of concurrent sessions to 4?

→Can you use DPoP to prevent session theft?

Most teams think they have to build or migrate to deliver. There is a third way: extend.

Keep your existing auth and simply enhance it with BYO Auth

HOW GPT-5.6 FUSES FRONTIER INTELLIGENCE WITH FRONTIER EFFICIENCY

The GPT-5.6 model family has been designed to optimize both performance and cost efficiency across various tasks, with load balancing and caching techniques that allow the system to handle more tokens effectively. GPT-5.6 Sol also helped optimize itself to improve its code execution and resource allocation.

HOW CHATGPT OPTIMIZES ITS AGENT LOOP: HARNESS, API, AND INFERENCE

AI models like ChatGPT optimize their performance using various layers, including a harness, API, and inference, to improve efficiency and reduce costs. They use techniques like persistent WebSockets, stable prompt prefixes, and parallel safety checks to minimize repetitive work and improve the overall processing speed of tasks.

GITHUB IS THE WRONG SHAPE FOR THIS NEW WORLD

The current software development landscape requires a reevaluation of collaboration tools like GitHub, as traditional paradigms are being overwhelmed by the velocity and scale of new contributions from various teams. To address these challenges, the focus should shift to developing high-throughput infrastructure primitives that improve software delivery processes.

NEW RESEARCH: THE CORE WEB VITALS THRESHOLDS YOU TRUST MIGHT BE WRONG FOR YOUR SITE

Google's 2.5s “good” LCP threshold comes from aggregated data across millions of sites, so it tells you nothing about where your own users actually start bouncing. Across ten major retailers, every single one hit its best bounce rate somewhere between 100ms and 1 second, well before that line. Four of the ten had already flattened out onto the performance plateau before reaching 2.5s, meaning they were technically “good” for Google while having stopped gaining anything from speed a while back.

GET YOUR FREE API KEY→

TURBOFIELDFARE

TurboFieldfare is a custom runtime designed for inference of the Gemma 4 26B-A4B model on Apple Silicon Macs with limited RAM, allowing the model to operate efficiently without loading the entire 14.3 GB into memory. The system uses streaming and caching techniques to maintain performance while only using approximately 2 GB of RAM, and includes various tools such as a native Mac app, command-line interface, and a local server for OpenAI-compatible functionalities.

CLAUDE CODE STOLEN QUEUE

Claude Code Merge Queue is a local, cost-effective solution for managing merge queues with parallel Claude Code agents, allowing simultaneous landings, builds, and tests while preventing issues related to push races and resource conflicts.

AI'S TOP STARTUPS ARE BARELY PUBLISHING THEIR RESEARCH

More than half of AI unicorns valued at over $1 billion have never published a scientific paper or preprint. A small number of these firms dominate academic citations, with OpenAI accounting for a significant share of them, while others prioritize fast-paced, informal disclosure methods over formal publications.

CONTAGIOUS INTERVIEW MALWARE IN SVG IMAGES: DPRK CAMPAIGN

DPRK-aligned actors are handing developers fully-functional “coding challenge” repos, one seeded through Elastic's own community Slack #jobs channel, with the payload split into Base64 chunks hidden in HTML comments across SVG flag images. It is reassembled and executed on every npm run dev or npm start. What lands is a four-module OTTERCOOKIE stack: browser credential and crypto wallet theft, a recursive sweep for .env files, SSH/AWS configs and shell histories, a Socket.IO RAT with live shell access, and a clipboard stealer.

HOW MUCH CAN YOU DELEGATE TO AGENTS?

To effectively delegate tasks to agents, it's necessary to assess the ease of reviewing their work and the cost of undoing mistakes, which helps determine the appropriate level of autonomy.

THE PRODUCTIVITY MIRAGE

A prolific engineer, known for his achievements at Facebook, surprised a productivity enthusiast by using a simple code editor without advanced tools, yet still demonstrated exceptional skills and intuition that led to winning a hackathon.