Search the archive

MAKING SENSE OF THE PERFORMANCE EXTENSIBILITY API

DEVELOPERS, REINVENTED

GitHub's CEO writes about how the developer role is being reinvented due to AI. These developers are moving through stages, from initial skepticism to becoming AI collaborators and strategists, focusing on delegation and verification rather than direct coding. AI will probably write a significant portion of code in the near future, leading to a transformation where their value lies in orchestrating AI agents and innovating beyond implementation.

NO, AI IS NOT MAKING ENGINEERS 10X AS PRODUCTIVE

AGENTIC CODING THINGS THAT DIDN'T WORK

Most of this dev's attempts at complex agentic workflows didn't stick, leading him to delete failed automations and simplify his approach. His most effective workflow ended up being 95% just talking to the AI through speech-to-text and providing more context, rather than using elaborate slash commands, hooks, or print mode features.

OPEN MODELS BY OPENAI

OpenAI has released two new open-weight reasoning models, gpt-oss-120b and gpt-oss-20b, designed for various use cases and deployment environments. These models are licensed under Apache 2.0, making them customizable and suitable for commercial use, with features like agentic task capabilities and full chain-of-thought access.

HARMONY

OpenAI uses the Harmony response format in its `gpt-oss` open-weight model series, which enables structured conversation outputs. The company has provided a Rust-based core with Python bindings for rendering and parsing the Harmony format.

CLAUDE OPUS 4.1

Anthropic has released Claude Opus 4.1, an upgraded version of Claude Opus 4 with improvements in agentic tasks, real-world coding, and reasoning capabilities. This new model achieves a coding performance of 74.5% on SWE-bench Verified and is great at precise code correction and debugging.

INJECTING JAVA FROM NATIVE LIBRARIES ON ANDROID

On Android, accessing advanced phone capabilities like Bluetooth or certain network features from Rust requires using the Java SDK via JNI because the NDK and Libc lack direct APIs. While JNI allows calling Java from Rust, traditionally, creating new Java classes from Rust was thought to be impossible. However, using `DexClassLoader`, Rust libraries can compile and inject necessary Java classes at runtime, making them more modular and easier for other developers to use.

HOW FAR CAN WE PUSH AI AUTONOMY IN CODE GENERATION?

Thoughtworks experimented with generative AI to autonomously build a Spring Boot application using strategies like reusable prompts and reference applications. While the AI could generate simple applications, issues came up with increasing complexity, including adding unintended features, making incorrect assumptions, and declaring success despite failing tests. Human oversight was essential for high-quality, maintainable code.

2025 STACK OVERFLOW DEVELOPER SURVEY

The Stack Overflow 2025 Developer Survey collected data from over 49,000 developers across 177 countries, focusing on technology usage, AI trends, and developer satisfaction. There is increasing AI tool adoption, with 84% of developers using or planning to use them, but also declining positive sentiment towards AI and distrust in its accuracy. Python's adoption has significantly grown, while Visual Studio Code and Visual Studio remain the top IDEs.

GENIE 3: A NEW FRONTIER FOR WORLD MODELS

Google DeepMind's Genie 3 is a new, real-time world model that generates diverse and interactive environments from text prompts, allowing users to navigate and interact within consistent, simulated worlds.

WHY IS GITHUB UI GETTING SO MUCH SLOWER?

The increasingly slow performance of GitHub's user interface is due to inefficient client-side routing and excessive DOM node rendering.

YOU CAN SPOT BASE64 ENCODED JSON, CERTIFICATES, AND PRIVATE KEYS

A trick to visually identify base64 encoded JSON is that they start with "ey", while certificates and private keys start with "LS" within files.

TAMBO

Tambo is a React package that lets devs build AI-powered apps with generative UIs by registering React components and tools for use by LLMs.

HOW WE MADE JSON.STRINGIFY MORE THAN TWICE AS FAST

V8 has sped up `JSON.stringify` (over 2x faster) through several technical optimizations. These optimizations are: a side-effect-free fast path for plain data, specialized string handling (one-byte vs. two-byte), SIMD instructions for string serialization, and an 'express lane' for frequently serialized objects. They also upgraded their DoubleToString algorithm and switched to a segmented buffer to avoid memory reallocation overhead.

COMPILE SVELTE 5 IN YOUR HEAD

Svelte 5's compilation process can be compared to vanilla JavaScript DOM manipulation. This article covers essential DOM concepts like the `template` element, event delegation, and reactive states using Svelte's new runes system ($state), showing how these translate to optimized JavaScript output. Through examples, it shows how Svelte 5 compiles components into functions that efficiently create and manage DOM elements.

LET'S WRITE REACT HOOKS FROM SCRATCH

This article goes over how React Hooks store and persist state within functional components. It begins with a simplified explanation using arrays and then goes into the actual implementation using a linked list within the virtual DOM's component tracking. Code examples are provided to show how `useState` could be implemented from scratch using both arrays and linked lists.

ONBOARDING FOR CODING AGENTS

When using coding agents like Claude Code, you should shift context from tool-specific files like `CLAUDE.md` to universal `README.md` files organized by domain for better onboarding. Code quality should be enforced through "Quality Gates," defined as environment constraints using tools like linters and type checkers. This approach allows AI coding agents to iterate and improve code until all checks pass.

READ THAT CODE!

It's important to read AI-generated code, despite the efficiency of tools like Claude Code. Otherwise, the codebase risks weakened architecture, and devs risk the loss of implementation knowledge. There needs to be a distinction between asynchronous tasks suitable for "auto-accept mode" and synchronous coding for core features requiring careful oversight.

EVERY VISUAL WORKFLOW TOOL IS JUST EXCEL FOR DEVELOPERS WHO GAVE UP

Visual workflow tools like Zapier are essentially Excel for devs who have lost confidence in their coding abilities. These tools, while seemingly easy, create complex dependencies, vendor lock-in, and limitations that ultimately make developers less capable and force them to pay for what they can already build.

OV

ov is a terminal pager for efficiently viewing text, especially tabular data and large files. It has functionalities like fixed headers, column highlighting, real-time updates with follow mode, and advanced search. The pager is customizable, allowing devs to configure keybindings, styles, and integrate with tools like `bat` and `delta`

DEPENDENCY CRUISER

Dependency Cruiser is a tool for validating and visualizing JavaScript, TypeScript, and related project dependencies against user-defined rules. It identifies rule violations in text or graphics and can generate dependency graphs in various formats.

SCREENCODER

ScreenCoder is a UI-to-code generation system that converts screenshots or design mockups into HTML/CSS code using a modular multi-agent architecture. It has visual understanding, layout planning, and adaptive code synthesis, allowing for customized modifications and easy tweaking of layouts and styles.

TOKENS ARE GETTING MORE EXPENSIVE

AI subscriptions are headed for a "short squeeze" despite the expectation that language model costs will decrease. The problem lies in the demand for the newest, most powerful models, which remain expensive, and the exponential increase in token consumption as AI agents become more capable. Unlimited subscription models are unsustainable, as users can exploit them to consume enormous amounts of tokens (which is why Claude is introducing weekly limits now).

PERPLEXITY IS USING STEALTH, UNDECLARED CRAWLERS TO EVADE WEBSITE NO-CRAWL DIRECTIVES

Cloudflare has discovered that Perplexity, an AI-powered answer engine, is using stealth crawling techniques to bypass website restrictions. Perplexity appears to be masking its crawler identity by modifying its user agent, rotating IPs, and ignoring robots.txt files after being blocked. This behavior violates established internet norms of transparency and respect for website preferences, so Cloudflare has responded by delisting Perplexity as a verified bot and implementing rules to block this stealth crawling activity.

DO LLMS IDENTIFY FONTS?

A live benchmark was created to test whether LLMs can identify fonts from dafont.com forum posts where users request font identification. The benchmark avoids contamination by only testing LLMs on new, unsolved forum posts and comparing their predictions to the eventual community solutions. GPT-4o-mini and Gemini-2.5-flash-preview-05-20 were both provided with images, titles, and descriptions, allowing up to five guesses each, but the results were surprisingly poor.

DRAWAFISH.COM POSTMORTEM — AUG 3, 2025 INCIDENT

The creator of DrawAFish.com, after getting Hacker News fame through vibe-coding and heavy reliance on Copilot, had a security incident involving username vandalism and offensive content due to a legacy password breach, lack of authentication in the username update API, and a JWT vulnerability.

CODE RELOADING FOR RACK APPS

This blog post explains how to implement code reloading in a Rack-based Ruby application, similar to what Rails provides, using Zeitwerk, the listen gem, and concurrent-ruby to automatically reload code changes without restarting the server.

KIMU

Kimu is an open-source, React, Remotion, and TypeScript-based video editor project with features like non-linear editing, multi-track support, live preview, and video export.

QWEN-IMAGE: CRAFTING WITH NATIVE TEXT RENDERING

Qwen-Image is a new 20B MMDiT image foundation model that can render complex text and help edit images precisely.

WHY OBSERVABILITY ISN'T JUST FOR SRES (AND HOW DEVS CAN GET STARTED)

OpenTelemetry (OTel) is a vendor-agnostic framework that introduces the idea of signals, primarily traces, metrics, and logs, which describe what your application is doing. It acts as a safety net for bugs, especially as systems scale.

WHAT IS THE N+1 QUERY PROBLEM AND HOW TO SOLVE IT?

The N+1 query problem is a performance bottleneck where a program makes one initial query and then N additional queries within a loop. The problem happens when fetching related data, resulting in excessive database traffic and slow application performance. The solutions include using a single query with a JOIN or fetching related data in batches using an IN clause, then building the relationship in code.

6 WEEKS OF CLAUDE CODE

Claude Code changed Puzzmo's engineering over the past six weeks, with a huge reduction in maintenance costs and an increased capacity for handling technical debt. It enabled rapid prototyping and sped up triage by offering immediate pull requests. Claude Code is most effective for developers with product and technical skills who have a "full-breadth developer" mindset.

LIVE CODING SUCKS

Live coding interviews are unfair because the stress induced by being watched reduces cognitive function, specifically working memory, which is necessary for solving coding problems. A study by Microsoft showed a large decrease in performance when candidates were observed during coding. Live coding measures stress response more than actual coding ability and may lead to the rejection of qualified engineers.

IF YOU'RE REMOTE, RAMBLE

For small remote teams, creating individual "ramblings" channels in their chat app allows each member to share thoughts and updates without disrupting main group channels. These channels function as personal journals or microblogs, allowing for better social connection and "water cooler talk" in the absence of in-person interactions.

LANGEXTRACT

LangExtract is a Python library by Google that uses LLMs to extract structured information from unstructured text documents based on user-defined instructions. It has source grounding, reliable structured outputs, and optimization for long documents with interactive visualization. LangExtract supports various LLMs, including cloud-based models like Gemini and OpenAI, as well as local open-source models.

TRACEROOT

TraceRoot is an open-source debugging platform that uses AI to help engineers resolve production issues faster by combining traces, logs, and source code. The platform has features such as real-time tracing and logging, structured logging for improved AI agent performance, and integration with tools like GitHub and Notion.

INTRODUCING CEREBRAS CODE

Cerebras is launching Cerebras Code, two new plans (Pro at $50/month and Max at $200/month) that offer access to the Qwen3-Coder, a recent open-weight coding model. These plans provide speeds up to 2,000 tokens per second with a large context window and no vendor lock-in.

A.I. RESEARCHERS ARE NEGOTIATING $250 MILLION PAY PACKAGES. JUST LIKE NBA STARS

AI researchers are being recruited like NBA stars and offered huge compensation packages. Tech companies like Meta are offering nine-figure salaries and perks to lure top AI talent, even engaging in personal recruitment efforts by CEOs. One 24-year-old researcher landed a $250M pay package to quit the startup he co-founded and join Meta's Superintelligence Lab.

PERSONA VECTORS: MONITORING AND CONTROLLING CHARACTER TRAITS IN LANGUAGE MODELS

Anthropic has identified "persona vectors" within AI models' neural networks that control character traits, allowing for control of personality. These vectors were discovered through the analysis of activity patterns associated with specific traits like "evil" or "sycophancy." They can be used to monitor personality shifts, mitigate undesirable traits during training (even acting as a "vaccine" by preemptively steering towards those traits), and flag problematic training data.

DOES THE BITTER LESSON HAVE LIMITS?

The "bitter lesson" in AI argues that general methods that use computation outperform human-centric approaches in the long run. However, the bitter lesson depends on high-quality, quantifiable data, which is often difficult to obtain in messy organizational environments. Also, prioritizing computation over efficient, purpose-built models might not always be practical or optimal in the real world.

I COULDN'T SUBMIT A PR, SO I GOT HIRED AND FIXED IT MYSELF

Frustrated by a long-standing search bug on Mintlify that he couldn't fix as a vendor, this dev joined the company and finally resolved the issue using an AbortController.

HTML IN CANVAS

The "HTML-in-Canvas" proposal introduces new HTML Canvas APIs for rendering HTML content into a canvas, addressing limitations in accessibility, internationalization, and performance when using canvas-based content.

TRY DEEP THINK IN THE GEMINI APP

Google is releasing "Deep Think," a more powerful version of Gemini 2.5 Pro, for Google AI Ultra subscribers and providing the gold-medal standard version to select mathematicians.

ANTHROPIC REVOKES OPENAI'S ACCESS TO CLAUDE

Anthropic revoked OpenAI's access to its Claude API due to OpenAI's alleged violation of terms of service by using Claude and Claude Code internally during development.

HOW CURSOR SERVES BILLIONS OF AI CODE COMPLETIONS EVERY DAY

Cursor is an AI-first code editor that uses AI models like GPT-4 and Claude to provide features like real-time code autocompletion, complex multi-file tasks, inline editing, automated code review through "BugBot," and background agents that can execute code in isolated cloud environments. These features are enabled by encrypting and sending code snippets to cloud servers for AI processing. Codebase indexing and semantic search also allows the AI to "understand" and answer questions about the entire codebase.

HOW GO 1.24'S SWISS TABLES SAVED US HUNDREDS OF GIGABYTES

Go 1.24 introduced Swiss Tables, a new map implementation that significantly improved memory efficiency for Datadog's services. While initially experiencing a memory regression from a separate mallocgc issue, Datadog discovered that its high-traffic services actually saw large memory reductions, up to 1 GiB saved in some cases, due to Swiss Tables' more compact storage format compared to Go 1.23's bucket-based maps.

STACK OVERFLOW DATA REVEALS THE HIDDEN PRODUCTIVITY TAX OF 'ALMOST RIGHT' AI CODE

A recent Stack Overflow survey found that while AI adoption among developers is rising, trust and favorability towards AI tools are declining. A major reason for this decline is the prevalence of "almost right" AI-generated code, which developers find frustrating and time-consuming to debug, negating potential productivity gains. This disrupts developer workflows and can introduce security risks.

I TRIED SERVO, THE UNDERCOVER WEB BROWSER ENGINE MADE WITH RUST

Servo is a web rendering engine built with Rust. It was a Mozilla project intended to potentially replace Gecko in Firefox, but Servo was later reorganized under the Linux Foundation after Mozilla layoffs. While not yet able to power a full-featured browser, Servo shows promise and continues development, with current efforts focusing on Shadow DOM and CSS Grid.