Search the archive

WHERE DID THE OLD WEB GO? WE FOLLOWED 657,607 LINKS TO FIND OUT

An analysis of over 657,000 links from a restored 0.mk database reveals that approximately 76.7% of the pre-2015 links are no longer accessible.

Grab your copy

HOW TO EVALUATE LLMS BEFORE PRODUCTION

Evaluating LLMs for production involves understanding the transition from initial benchmark performance to real-world effectiveness, as production conditions often include ambiguous data, inconsistent labels, and edge cases that benchmarks do not capture. Successful evaluation requires a focus on product goals, a testing process similar to integration testing, and the use of error analysis to address specific performance challenges.

BALANCING FLEX ITEMS WITH FLEX-WRAP

Flexbox layouts can sometimes result in orphan items when used with wrapping, which often prompts developers to rely on media queries to avoid this issue. A new CSS feature, `flex-wrap: balance`, helps create more visually appealing layouts by making sure items wrap evenly, alongside the introduction of `flex-line-count` to set a minimum number of lines for wrapped items.

YOU NEED TO FIND PRODUCT-MARKET FIT AGAIN (SORRY)

Companies must continuously seek product-market fit, especially in the face of emerging competitors and technological advancements. The creation of PostHog Desktop shows this process.

AGENTIC ENGINEERING FOR RELIABLE CODING AGENTS

To implement agentic engineering effectively, begin by clearly defining success conditions for tasks, incorporating the simplest deterministic checks to monitor outcomes, and making sure focused feedback is provided to the agent during execution. Limit access and permissions, record all actions and decisions for accountability, and continuously refine the workflow based on past failures to build a trustworthy coding environment.

code against known benchmarks. See the repo.

APPLE INTRODUCES M6 AND M5 ULTRA FOR A BIG LEAP IN PERFORMANCE AND AI COMPUTE

Apple has launched the M6 and M5 Ultra chips, marking advancements in performance and AI capabilities, with the M6 being the company's first 2 nm chip. These new chips have better CPU and GPU configurations, increased unified memory bandwidth, and improved efficiency.

WHIP

Whip is a fast coding-agent harness developed in Go, designed for efficiency and streamlined interactions with various models and tools. It integrates best practices from multiple frameworks, supports open-source model discovery, and provides features such as an interactive user interface and background subagents.

LATTICEDB

LatticeDB is an embedded property-graph database that integrates vector similarity, full-text search, and graph traversal capabilities into a single local engine with a fast query layer. It allows for efficient querying of connected data, supports multiple programming languages like Python and JavaScript, and has great performance metrics.

HOW AI IS ACTUALLY REMAKING SECURITY

AI is changing cybersecurity by making the reliance on "security by obscurity” useless, as attackers can now use AI's capabilities to efficiently and extensively probe systems for vulnerabilities. This shift requires security teams to focus on reducing exposure, improving their analytical advantages, and fostering resilience to combat the new landscape of cyber threats.

PYTHON'S PRE-DECLARED CONSTANTS ARE KINDA WEIRD

Python has six pre-declared constants (True, False, None, __debug__, Ellipsis, and NotImplemented) that exhibit distinct behaviors. Among these, True, False, and None are unique keywords, while __debug__ is a special identifier that cannot be assigned to, contrasting with the built-in nature of Ellipsis and NotImplemented, which can be shadowed by globals.

WHEN STR.LOWER() IS A SECURITY VULNERABILITY IN PYTHON

The use of `str.lower()` in Python's StringPrep implementation poses a security vulnerability due to its reliance on the Unicode data version that comes with the Python interpreter. The recent remediation involves creating new exceptions to align `str.lower()`'s behavior with Unicode 3.2.0.

EXOWILL

An AI "great helper" should learn and adapt to an individual's long-term goals while ensuring data sovereignty, empowering users to shape their own technological aids.

COMPRESSION IS PREDICTION

Compression techniques and LLMs fundamentally share the same goal of predicting and reducing data, as both rely on understanding and using probabilities to efficiently represent information.

THE MOVE TO PYTHON 3 BEGINS!

EVE Online is transitioning from Python 2.7 to Python 3 to improve performance, improve development tools, and future-proof the game's codebase.

ONE GO BINARY, ONE YAML FILE, ONE SQLITE DATABASE: WHY I WROTE MY OWN MONITORING TOOL

A lightweight monitoring tool called Gjallar was developed to oversee a variety of services without the complexity of traditional monitoring platforms, using a single binary, a YAML configuration file, and an SQLite database.

START BUILDING

HOW TO DESIGN AN ANIMATION

Animation improves user experience by clarifying interactions and conveying brand personality, but creating it can be a complex task that requires breaking down intricate motions into manageable components represented mathematically. Designers can develop animations that feel natural and responsive to user input by using tools such as graphs, easing functions, and physics simulations.

YOUR EXECUTABLE IS A SQLITE DATABASE

The concept of replacing the traditional ELF executable format with SQLite as an executable format is explored in this article, introducing a new prototype called SELF, which functions as a structured executable and linkable format. This shift allows for simplified querying and modification of executable files using SQL.

HOW TO REPORT A BUG SO IT ACTUALLY GETS FIXED

Effective bug reporting includes owning the bug, creating a reproducible environment, and providing detailed documentation, all while assuming initial responsibility for the issue. By following a structured process that includes version bisecting and thorough communication with maintainers, the chances of a bug being addressed are improved.

AI CODING WILL PREVENT EXPERTISE

The growing reliance on AI coding tools may negatively affect the development of true expertise among junior developers, requiring them to possess expert-level skills to effectively use AI.

STOP USING CONVENTIONAL COMMITS

Conventional Commits prioritizes commit type over the more critical component, scope, resulting in a misleading format that fails to communicate essential changes effectively. The article advocates for a shift towards scope-prefixed commit messages, emphasizing that this approach aligns better with the actual needs of developers and contributors.

works

KERN

Kern is a lightweight, rootless container runtime that allows for fast, kernel-enforced execution of workloads, achieving cold starts in approximately 3.5 ms without a daemon.

PICOMQ

PicoMQ is a framework designed for durable, real-time streaming over HTTP, utilizing S3-compatible object storage. It provides tools for easy installation, node management, and stream operations through its command-line interface and Docker integration.

WHAT LANGUAGE ARE AGENT SKILLS WRITTEN IN?

The percentage of newly created agent skills written in languages other than English rose from 13.0% to 16.3% in the first half of 2026, with Chinese leading among these non-English skills. The growth of non-English skills shows a shift in the AI development landscape, influenced by increasing contributions from countries like India, Brazil, and others.

HOW EUROPE IS KILLING MAKERS AND MICRO-ENTREPRENEURS

New packaging regulations introduced by the European Union threaten the viability of makers and micro-entrepreneurs by imposing disproportionate costs and administrative burdens. The regulations could lead to job losses and stifle innovation, forcing many micro-businesses to consider selling exclusively in non-EU markets instead.

REINFORCEMENT LEARNING FOR LLMS: THE COMPLETE GUIDE

Reinforcement Learning (RL) has been integral to the development of LLMs, addressing complex issues such as alignment, reasoning, and safety while being a fast-moving area of AI research. This overview covers foundational concepts of RL, its application in LLM training, various algorithms used, including policy gradient methods and recent advancements like Group Relative Policy Optimization (GRPO).

steps to improve. Score your docs

AN UNLIKELY EXPERIMENT

The use of the "unlikely" macro in programming allows developers to optimize their code by providing the compiler with information about which execution paths are less likely to be taken.

ANGER, ANXIETY AND AGENCY

Anger in the workplace is often counterproductive, as it does not improve situations and can negatively impact those around you, compared to more constructive emotional responses like curiosity.

GLASSBOX

GlassBox is a client-side tool that reveals the various signals a website can collect from a browser.

WHY CODE VERIFICATION MATTERS MORE THAN EVER IN THE AGE OF AI

AI made writing code cheap, but verifying it correct/safe is now the bottleneck, so verification should be a layered, risk-adjusted stack.

Register now

THERE'S NO REASON FOR SOFTWARE TO BE SLOW ANYMORE

The cost of software optimization has drastically decreased, enabling more developers to implement performance improvements that previously required specialized skills. This allows for dynamic custom software tailored to specific workloads. As the barrier to entry for coding and testing optimizations lowers, it becomes increasingly feasible to get performance gains with minimal effort and time.

AI CHIP ARCHITECTURES

Advancements in AI chip architectures have led to NVIDIA GPUs, Google's TPUs, AMD Instinct GPUs, and Cerebras' Wafer-Scale Engine, each with distinct designs catering to specific computational AI tasks. The key differentiators include NVIDIA's focus on programmability and scalable architecture, Google's emphasis on a single-purpose design optimized for matrix multiplication, AMD's commitment to memory capacity and open standards, and Cerebras' innovative wafer-scale integration for ultra-fast memory access.

HOW I FIND PROBLEMS TO SOLVE AS A STAFF ENGINEER

Finding worthwhile problems to solve as a staff engineer involves actively listening to the challenges faced by team members rather than waiting for management to identify opportunities. By absorbing feedback, allowing potential problems to accumulate, and connecting seemingly unrelated issues, engineers can uncover needs and develop impactful solutions that align with the broader organizational goals.

AI AND INFRASTRUCTURE ENGINEERING

The integration of AI into infrastructure engineering is improving efficiency by automating routine tasks and lookups, freeing engineers to focus on higher-level decision-making, but it may also lead to diminishing expertise in foundational skills.

ON TEACHING AI HOW YOU WORK

A collection of eleven skills has been developed to streamline the process of teaching AI how to interact with design and engineering tasks by consolidating key lessons and practices into a single repository. This framework allows users to apply these skills efficiently, ensuring consistency in design intent and output while minimizing the need for repeated explanations.

Claim 50% off

MAKA

Apache Maka is an incubating project that provides a local-first AI agent workspace where user interactions and data are stored locally, allowing for the execution of tasks while maintaining control over session data.

PROLIFERATE

Proliferate is an open-source AI IDE that enables users to run various coding agents, such as Claude Code and Codex, in parallel within isolated workspaces. It supports self-hosting and has features like reusable workflows, integration with multiple tools, and the ability to manage different tasks simultaneously in a structured manner.