Search the archive

VULNERABILITY REPORTS ARE NOT SPECIAL ANYMORE

As of 2026, vulnerability reports are no longer considered special due to advancements in LLMs that can identify security issues similarly to human researchers.

GLM-5.2 - HOW TO RUN LOCALLY

Unsloth Studio is a web UI for local AI that allows users to run advanced AI models, such as the GLM-5.2, efficiently on various operating systems while providing features like model downloading, parameter tuning, and fast inference capabilities.

EVERYTHING A SENIOR ENGINEER NEEDS TO KNOW ABOUT WHAT'S INSIDE AN LLM

LLMs are made of Transformers, which use self-attention mechanisms to improve processing efficiency and address issues associated with traditional RNNs, such as sequential bottleneck and long-range decay of information. This post goes over various components of Transformers, including positional encoding, attention mechanisms, and the importance of architecture design in enhancing AI model performance.

OPTIMIZATION CATALOG. HOW 4 BYTES OF PADDING MAKE ARRAY CLEARING 49% FASTER

Adding 4 bytes of padding between a struct's fields in Go can improve the performance of array clearing, achieving up to a 49% increase in speed on Intel machines due to better instruction optimizations. When the array elements are misaligned, the performance suffers because the processor's optimized memory handling cannot be fully used.

P99 0 MS* AUTOCOMPLETE FOR 240 MILLION DOMAIN NAMES

Wirewiki is a website designed to inspect internet infrastructure, including domain names, and aims to provide fast and efficient autocomplete features for users. The site's autocomplete system uses client-side prefetching and an optimized API to deliver suggestions with minimal latency, achieving a 99% success rate of results ready before the user releases the key.

IN PRAISE OF MEMCACHED

Memcached is a simpler, high-performance alternative to Redis for caching, as it avoids issues related to persistence and complexity. Its stateless nature makes it a strong choice for dynamic web apps, provided query optimization is managed.

YOU SHOULD USE AI FOR REVIEWING CODE ESPECIALLY WHEN THE DIFF IS HUGE

Using AI to review large code changes can make it easier to focus on high-level issues rather than nitpicking line-by-line details. Reviewers should contribute unique knowledge, such as insights from recent discussions and overarching principles specific to the codebase, that neither the author nor the AI may possess.

DENO DESKTOP

Deno Desktop allows devs to create self-contained desktop applications from Deno projects, using a small binary that includes the Deno runtime and a web rendering engine. It supports a variety of web frameworks and features in-process communication for efficient backend and UI interaction.

ORCA

Orca is an advanced AI orchestration tool for devs to manage and run multiple coding agents in parallel through a user-friendly interface. It has capabilities such as terminal splits, design mode, and mobile companion applications, allowing users to monitor their agents and streamline workflows.

GLM-5.2 IS THE NEW BEST OPEN MODEL

GLM-5.2 has recently been released, with impressive performance metrics that position it as the strongest available open model. While it does great in specific areas, the model's higher costs and limitations, such as the lack of vision capabilities and tendencies towards overfitting on benchmarks, could make its practical utility worse in various tasks.

LLM-DRIVEN FEATURE DISCOVERY

LLM-driven feature discovery involves analyzing a dataset of chat transcripts to generate and cluster features related to model behavior. The method, which does not require access to model internals, identifies high-level features and predicts some aspects of the model's responses based on user input, although prediction accuracy is limited.

GLM-5.2 VS CLAUDE OPUS

Tech Stackups tested open-weights GLM-5.2 against Claude Opus 4.8 on a one-shot “build a 3D WebGL platformer from scratch” task: Opus was faster and shipped a cleaner, more correct game, while GLM-5.2 was much cheaper but rougher and, being text-only, couldn't visually self-check its work. The takeaway: use GLM-5.2 when cost and openness matter for text/logic work, and Opus when correctness, polish, and visual judgment matter.

I'M THE AGENT FOR CLAUDE NOW

An engineer at Aha! has successfully integrated AI tool Claude into their support workflow, improving efficiency by automating data retrieval and analysis from production logs and other tools.

THE TEXT IN CLAUDE CODE'S “EXTENDED THINKING” OUTPUT IS NOT AUTHENTIC

Anthropic's Claude Code encrypts its reasoning process into a signature, providing only summaries of its thinking rather than real insights, creating challenges for users seeking an audit trail of the model's decision-making logic.

TEN YEARS OF TERMS AND CONDITIONS

Over a ten-year period, the author documented 3,066,332 words from 821 terms and conditions agreements, showing the significant yet often overlooked burden of reading these lengthy legal documents.

TYPESCRIPT PERFORMANCE IN TANSTACK TABLE V9

TanStack Table V9 was initially very slow to verify and typecheck, but the team improved its performance by simplifying the library's internal code structure.

Register now

JSON-LD EXPLAINED FOR PERSONAL WEBSITES

JSON-LD is a structured data format that helps web crawlers understand the semantic structure of webpages, potentially leading to richer search previews and improved SEO. To implement JSON-LD, users can insert specific scripts within the head section of their HTML, defining elements such as the website, its owner, and various page types.

I STORED A WEBSITE IN A FAVICON

A favicon, typically serving as a small icon for a website, can function as a storage medium for data by embedding HTML content within its pixel values. This process involves converting the HTML into bytes and encoding them into the RGB channels of the favicon, allowing the browser to extract the information using JavaScript.

BUILDING RELIABLE AGENTIC AI SYSTEMS

The Preclinical Information Center (PRINCE) is a cloud-based platform developed to improve drug development in the pharmaceutical industry by integrating decades of safety study data through RAG and Text-to-SQL. It transforms traditional keyword-based searches into an intelligent assistant capable of handling complex queries and drafting regulatory documents.

THERE IS MINIMAL DOWNSIDE TO SWITCHING TO OPEN MODELS

Switching to open models has become less risky due to improvements in Linux, web applications, and open-source software. Although proprietary models still dominate in performance and usability, recent advancements in open models have narrowed the gap.

HEY, N00B, WE DIDN'T HIRE YOU TO COMPLETE TASKS

Completing tasks is less important than demonstrating your growth and contribution as a developer. Signals of higher performance include delivering quality work, effective communication, and continuous learning, rather than just finishing multiple tasks quickly.

→

RECALL

Recall is a local memory plugin for Claude Code that captures and summarizes session activities offline. It generates concise project summaries without requiring external servers or API keys.

LOUPE

Loupe is an iOS and iPadOS app designed to showcase the device fingerprinting capabilities of mobile devices by displaying real values from public iOS APIs. It categorizes the data it retrieves into three tiers based on accessibility: passive signals available to any app, signals that require user permission, and advanced signals through clever API usage.

TEMPORARY CLOUDFLARE ACCOUNTS FOR AI AGENTS

Temporary Cloudflare Accounts have been introduced for easier deployments for AI agents, removing the need for human sign-up processes that typically involve browser-based authentication and multi-factor prompts. Agents can now use the new `--temporary` flag with Wrangler to deploy projects quickly without prior account setup.

THE MINIMUM VIABLE UNIT OF SALEABLE SOFTWARE

The decision to build or buy software has changed, as LLMs allow for cheaper software development but still involve maintenance and oversight costs. A “zone of viability" is introduced where software must strike a balance between cost and complexity. It is more practical to buy software than to build it when prices are reasonable and rebuilding efforts are substantial.

TEN YEARS OF CLICKHOUSE IN OPEN SOURCE

ClickHouse, an open-source analytical database, has gained popularity over the past decade since its release in 2016, attracting over 2000 contributors. With a focus on modularity and accessibility, it serves as a resource for engineers to learn C++ and experiment with database optimization techniques. It has transformed from a private project to a widely adopted tool across various industries, successfully filling a niche for high-performance data analytics.

PLOY RAISES $27M TO TURN YOUR COMPANY'S WEBSITE INTO YOUR HARDEST WORKING EMPLOYEE

Ploy is an all-in-one marketing platform that automates website management and operations, allowing founders, agencies, and marketing teams to efficiently translate ideas into executed campaigns without the typical overhead.

WHEN I REJECT AI CODE EVEN IF IT WORKS

Despite the advantages of AI in coding, many engineers reject AI-generated code when they struggle to understand it, find its solutions overly complex, or realize it doesn't align with good engineering principles.

DEVELOPERS DON'T UNDERSTAND CORS

Many web devs don't have a clear understanding of CORS, leading to security vulnerabilities, as shown by Zoom's use of a localhost web server that bypasses CORS protections.

ASK FOR NO, DON'T ASK FOR YES

Instead of seeking permission for new initiatives at work, confidently present your plans with a deadline for feedback, allowing for progress while still keeping your supervisor informed.

HOW WE BUILT A PERSISTENT AGENT MEMORY LAYER ON ELASTICSEARCH WITH 0.89 RECALL AND ZERO TENANT LEAKS

A persistent agent memory layer built on Elasticsearch uses a multi-index system to improve long-term memory capabilities for agents, leading to high recall accuracy and tenant isolation. With separate indices for episodic, semantic, and procedural memory, the architecture efficiently manages various memory types while addressing challenges like memory decay and data contradiction through a consistent update process.

CELL-BASED ARCHITECTURE FOR RESILIENT PAYMENT SYSTEMS

American Express uses a cell-based architecture for its core payments ecosystem to improve resiliency and maintain high availability. This architecture isolates failures within independent units, allowing each cell to operate autonomously and contain issues without affecting the broader system.

SECURING THE FUTURE OF AI AGENTS

AI agents are improving productivity across various sectors, with projections suggesting they could generate $2.9 trillion in economic value in the US by 2030. To address the challenges posed by increasingly capable and misaligned AI, Google DeepMind created an AI Control Roadmap, focusing on security measures that go beyond standard alignment techniques.

EMACS, HOW IT ALL STARTED (FOR ME)

This journey to embracing Emacs as a primary text and code editor began with clumsy early experiences in programming, including misguided choices like using Microsoft Word for coding instead of a proper text editor. After exploring various editors like Notepad++, Sublime Text, and different programming languages, the author ultimately found that Emacs offered great support and functionality.

WHY WE'RE BULLISH ON LOOPS

Loops let agents autonomously complete tasks rather than relying on manual prompts from engineers. The four components for constructing a loop include a clear goal, sufficient context, evaluation mechanisms, and operational agents. Critics may view loops as a threat to engineering roles, but they actually represent a shift towards self-driving products.

I HATE COMPILERS

The complexities of compiler behavior can lead to unexpected nondeterministic outputs, even when given the same input bytes, which complicates the goal of achieving reproducible builds. As part of developing Anubis, a system for WebAssembly-based proof of work checks, challenges arose with making sure that output from the WebAssembly compiler was consistent across different environments and versions of tools. To address these issues, including variances in tool versions and address layout differences, a series of measures was implemented to stabilize the build process and verify reproducibility through checksum verification across different architectures.

FLUE

Flue is a framework designed for building autonomous agents that efficiently perform complex tasks using a programmable TypeScript environment. It allows developers to create agents that can maintain context, execute workflows, and operate securely in a sandbox environment, which facilitates real work without predefined sequences of actions.

AGENT-NATIVE

Agent-Native is an open-source framework designed for building interconnected applications where agents and user interfaces operate seamlessly together. The framework offers support for creating versatile agents that function as either headless APIs, rich chat experiences, or full applications, all while maintaining real-time synchronization and context-awareness.

I DISCOVERED A LARGE-SCALE MALWARE DISTRIBUTION ON GITHUB

A large-scale malware distribution campaign has been uncovered on GitHub, involving 10,000 repositories that contain Trojan malware disguised under various names by different contributors. These repositories consistently update their readme files to include links to a zip archive that contains the malicious software, while frequently deleting and re-adding the same commits, which complicates detection.

IS "AGENT READINESS" A REAL CATEGORY - OR JUST SNAKE OIL?

The concept of "agent readiness" has gained traction, with platforms like Cloudflare and Google introducing scoring systems to evaluate websites for AI agent compatibility, though much of this hype may not translate into tangible benefits. The term itself conflates two distinct challenges: discovering whether AI applications recommend content and ensuring usability for transactions once users arrive at a site.

I TOLD THEM FORCED CONSENT WAS UNLAWFUL. FIVE YEARS LATER, IT COST ELKJOP €1.8 MILLION

Elkjop was fined €1.8 million for violating GDPR and ePrivacy regulations by forcing customers to accept marketing communications in exchange for membership benefits. This ruling by the Norwegian Data Protection Authority reinforces that consent must be freely given, serving as a warning to companies that rely on coercive consent models.

ENTERPRISE-MANAGED AUTHORIZATION: ZERO-TOUCH OAUTH FOR MCP

The stable Enterprise-Managed Authorization extension for the Model Context Protocol enables organizations to streamline security and improve user experience by providing centralized, zero-touch single sign-on access to MCP servers via identity providers.

GOOGLE GEMINI CO-LEAD NOAM SHAZEER TO JOIN IPO-BOUND OPENAI

Noam Shazeer, one of the main authors of the Transformers paper that kicked off the LLM era and co-lead of Google's Gemini project, is leaving the company to join OpenAI ahead of its initial public offering.

THE OUTBOX PATTERN

The Outbox Pattern in microservices allows for independent processing of events, such as sending an email after completing a transaction, by recording messages in an outbox table to make sure the transaction can succeed without waiting for email delivery.

KUBERNETES IN THE AGE OF AI

Kubernetes has evolved from a container orchestration tool to a fundamental AI infrastructure platform, allowing organizations to efficiently manage generative and agentic AI workloads.