Search the archive

CHOOSING AN AI MODEL: ONE PROMPT, 11 MODELS, VERY DIFFERENT RESULTS

A range of AI models, including DeepSeek, Qwen, and Kimi, were tested for generating web pages, showing large variations in performance and credit costs. The best-performing models for a static coffee shop website were DeepSeek V4 Flash, which consumed only 2.4 credits on average, and GPT 5.6 Sol in low effort mode, which averaged 141 credits. In contrast, higher-tier models like Claude Opus 5 showed a lot of credit usage with results that didn't necessarily justify their cost.

HOW COMPACTION WORKS IN PI

Compaction in coding agents like Pi is a process used to summarize and manage the growing conversation history that can exceed the model's context window during log coding sessions. By creating a smaller representation of prior messages, compaction makes sure that the conversation can continue while preserving relevant information, preventing errors related to context overflow.

UNDERSTANDING IS THE NEW BOTTLENECK

As AI agents increasingly take on coding tasks, humans still need to understand the code being generated to actively participate in the creative process for collaboration purposes. Effective techniques for improving this understanding include creating structured code explainers, utilizing interactive environments to intuitively grasp systems, and developing shared spaces for team discussions.

BLOG ABOUT THINGS YOU DON'T UNDERSTAND YET

Blogging serves as a learning tool, where the author consistently gains new insights through the writing process. This approach not only enriches the author's understanding but also offers value to readers, as it forces clear communication and thoughtful engagement with complex topics.

INTRODUCING GEMINI 3.7 FLASH

The new Gemini 3.7 Flash model significantly enhances coding and development capabilities by improving accuracy and efficiency in software engineering and web development tasks while offering it at a reduced cost compared to its predecessor. This latest version also incorporates better safeguards, making it more robust for handling complex workflows in knowledge-intensive fields and enhancing the developer experience with improved adaptability and reduced manual oversight.

DEEPSEEK HARNESS DEVELOPER PREVIEW: EVERYTHING IS A PLUGIN

DeepSeek Harness is now in developer preview, allowing developers to create agent harnesses using a modular plugin system where every capability, such as models, tools, and scheduling, can be swapped or extended. The platform supports multiple runtime modes and ensures complete traceability of agent actions through session logs.

TEXT AI WATERMARKS WILL ALWAYS BE TRIVIAL TO REMOVE

The European Union AI Act, which takes effect in August, mandates that all AI-generated text be watermarked to make sure it can be identified as artificially generated. Despite proposed solutions like Google's SynthID and potential Unicode tricks from other providers, the methods for watermarking text can still be easily circumvented.

ACCELERATING GPT-5.6 SOL ULTRAFAST WITH OPENAI

Cerebras and OpenAI have introduced a new service tier, GPT-5.6 Sol Ultrafast, which increases AI processing speeds up to 750 output tokens per second without sacrificing quality. It is powered by Cerebras' Wafer-Scale Engine, which eliminates memory bandwidth bottlenecks by packing 44 GB of SRAM on-chip to keep model weights local. This allows organizations to use AI more effectively in critical scenarios, such as addressing production outages rapidly or responding to cyberattacks.

EMPTY SHELVES OR LOST KEYS? RECALL IS THE BOTTLENECK FOR PARAMETRIC FACTUALITY

LLMs such as Gemini 3 and GPT-5 often encounter factual errors not due to a lack of encoded knowledge, but because they struggle with recalling this stored information. By using a knowledge profiling framework, analysis shows that while encoding is nearly saturated in these models, recall is still a hard challenge.

NP-OVERRATED

Despite the common belief that NP-hard problems are practically unsolvable, advancements in algorithms and heuristics have demonstrated that efficient solutions can often be found for a majority of cases.

ONE WEEK OF BUILDING AND REVIEWING CODE WITH LLM AGENTS

A week-long experiment using LLM agents for backend coding and review revealed insights into productivity, error detection, and the evolving role of human oversight in software development processes.

ORDINARY ABUNDANCE

Modern conveniences and technologies that once inspired awe have become commonplace in daily life, reflecting humanity's relentless pursuit of progress and the legacy of past innovations.

HOW (SOME) CHINESE AI PRACTITIONERS VIEW MODEL DISTILLATION

Chinese AI practitioners are increasingly recognizing the controversial practice of model distillation, which involves using data from superior "teacher" models to improve the capabilities of smaller "student" models, as a technical strategy that is misunderstood and stigmatized.

Join live (and bring your questions) on August 19

HOW TAILSCALE HELPED FIND THE SQLITE WAL-RESET BUG

Tailscale faced a lot of challenges due to a recurring database corruption issue within its SQLite implementation, which persisted for six months and affected its service reliability. After extensive investigation and collaboration with SQLite developers, its team identified and resolved a 16-year-old bug in SQLite, improving their backup and recovery processes to prevent future incidents.

HOW WE BUILT A SOFTWARE FACTORY TO DRIVE ASTRO'S GITHUB ISSUE COUNT TO ZERO

The Astro team automated GitHub issue triage with isolated AI subagents running in GitHub Actions that reproduce bugs, diagnose root causes, and ship preview releases for reporters to verify. They drove open issues from 200+ down to about 30, with zero expected within a month. The pipeline works as a state machine driven by issue labels, and agent failures are treated as signals of underlying problems in the codebase like opaque abstractions or missing tests.

HTML OVER WEBSOCKETS: REAL-TIME SPAS WITH BARELY ANY JAVASCRIPT

Building a Single-page Application usually requires JavaScript for rendering views, connecting to an API for JSON data, and managing two independent codebases. However, HTML over WebSockets is an alternative where the server sends pre-built HTML directly to the client. This method simplifies architecture by consolidating rendering logic on the server, reducing complexity, and improving real-time capabilities with a persistent bidirectional communication channel.

AI IS REMOVING THE MIDDLE CLASS OF SOFTWARE ENGINEERING

AI is accelerating the rate of software development, leading to an environment where poor engineering practices proliferate unchecked, resulting in complex codebases that no one fully understands. This shift creates a stark divide in the industry, where great engineers gain increased value and job security, while those lacking the necessary skills struggle to remain employable.

BUILD WIDE, SHIP NARROW

The traditional engineering workflow of tightly planning features before coding has evolved, allowing for broader initial builds and iterative improvements through AI-assisted development. By focusing on demoing work before code reviews and breaking down tasks into smaller, independently reviewable pull requests after building, engineers can improve productivity and address feedback more efficiently.

Get started securing agent access

GROK BOT: A NEW KIND OF COLLEAGUE

Grok Bot is an AI tool designed to act as a virtual teammate that can handle various tasks by interfacing with user applications, learn workflows over time, and collaborate with other bots. It operates 24/7, allowing users to assign multiple projects and manage task completion efficiently without the need for constant oversight.

INTRODUCING GROK 4.6

Grok 4.6 improves capabilities for long-running agents and ambitious interactive projects, showing improved performance across various coding and knowledge work benchmarks. With a focus on safety and usability, Grok 4.6 is designed to transform product ideas into functional applications while incorporating self-checking and verification features.

INTRODUCING DELTA

Delta is a new multiplayer coding environment that allows developers and agents to collaborate in real-time, keeping their code and conversations interconnected. With features like DeltaDB, comments on code are synchronized and remain relevant as the code evolves.

PROGRAMMING LANGUAGE ADOPTION PATTERNS AT META

Meta's programming language usage shows that Rust and TypeScript have experienced exceptional year-over-year growth rates, with TypeScript showing an increase of +1,200% in recent months. Meanwhile, languages like Haskell and Java have been steadily losing market share.

CRACKS IN THE AI THESIS

Business adoption of AI shows mixed results, with Anthropic leading in market share but slower uptake of its latest model, Fable 5, compared to OpenAI's offerings. Despite growing competition from open source models, American companies continue to increase their overall AI spending, although there is a limit to how much they are willing to pay for performance improvements.

INSIDE AN APPLIED AI COMPANY

An Applied AI company focuses on providing complete solutions for complex workflows, improving accuracy and efficiency through tight integration, continuous evaluation, and a unique operational structure that bridges engineering and customer engagement.

AGENTIC ENGINEERING IS JUST EVERYTHING WE HAVEN'T BEEN DOING

Effective implementation of AI tooling requires addressing existing issues within teams rather than relying solely on automation.

WHY TINY JPEGS LOOK DIFFERENT IN CHROME

Tiny JPEGs can appear differently in Chrome due to a rendering optimization that uses partial IDCT scaling, which allows the browser to decode only low-frequency data when the image is displayed at a small size.

MIGRATING A LARGE FLOW MONOREPO TO TYPESCRIPT

Yelp successfully migrated its large codebase from Flow to TypeScript over three years by gradually converting one package at a time, making sure of type integrity throughout the process.

Register now

WHERE AN AI WATERMARK CAN HIDE IN PLAIN TEXT

Anthropic announced it would start embedding an imperceptible watermark in the text generated by its AI model, Claude, though the specific method of implementation remains undisclosed. While the file-level watermark can be bypassed easily, the text watermark is theorized to exist in a statistical bias in word choice, allowing it to survive copying and pasting, but strategies such as canonical regeneration and prose rewriting could strip this watermark.

5 CSS PROPERTIES YOU SHOULD KNOW FOR BETTER TEXT DESIGNS

To improve web typography, five CSS properties can improve text design: background-clip allows for images to fill the text, while vertical-align and align-content help in vertically aligning elements. Additionally, box-decoration-mode helps with uniform styling of broken text edges, letter-spacing can create unique animations, and text-combine-upright helps with vertical text layout for East Asian typography.

THE DANGERS OF POSTGRES SUBTRANSACTIONS

Accumulating excessive subtransaction IDs in PostgreSQL can lead to performance issues across the cluster, affecting throughput and preventing new read replicas from accepting connections due to overflowed running transaction states. To mitigate these risks, it is recommended to keep transactions short and monitor subtransaction cache and activity, while potential solutions on the PostgreSQL side could involve increasing the cache size or implementing a new snapshot methodology.

ON COMMENTS

For the past decade, a unique syntax highlighting configuration has emphasized code comments by coloring them bright orange to encourage better documentation practices. While comments can often clutter code with unnecessary information, useful comments, like TODOs, correctness arguments, and rationale for specific choices, are invaluable for informing future developers and clarifying the reasoning behind the code.

CREATING HOPE

Creating a culture of hope within engineering organizations is necessary for overcoming a persistent state of emergency, where teams feel overwhelmed and demoralized by constant pressures and shifting priorities. This is better than defaulting to a "salvation" mindset, which seeks quick fixes like adopting microservices without understanding the context or true needs.

SOURCEBOT

Sourcebot is a self-hosted tool designed to help users understand their codebase through features like complex question answering, code search, and navigation across multiple repositories.

CORSAIR

Corsair is a unified integration layer for agents, allowing secure connections to various applications while maintaining control over permissions and actions. It implements different approval modes for interactions, making sure that sensitive operations, such as sending emails or making API calls, require user confirmation.

SHATTERING THE DREAM - WHEN A JOB OFFER BECOMES A ZERO-DAY ATTACK

The Operation Dream Job campaign, linked to the Lazarus group, is targeting the defense sector worldwide by distributing modified PDF viewers that execute malicious payloads when opened. The attackers exploit zero-day vulnerabilities, including CVE-2026-68820, to deploy advanced malware like the Troy backdoor and FudModule rootkit, while using compromised web servers for command-and-control operations.

STOLEN THOUGHTS

This paper shows that encrypted chain-of-thought traces returned by OpenAI, Anthropic, and Google APIs can be replayed into weaker, jailbroken sibling models to recover the stronger model's hidden reasoning in plaintext. The authors also decoded 315,320 reasoning blocks from public agent trajectories and found 704 distinct privacy artifacts, including API keys, passwords, access tokens, and PII, showing that these supposedly hidden traces can create a serious data-leak surface.

WHY GO IS AN IDEAL LANGUAGE FOR AI-ASSISTED SOFTWARE ENGINEERING

The Go programming language is an ideal choice for AI-assisted software engineering, as it prioritizes code readability, maintainability, and a cohesive platform that supports both human and AI collaboration through its design principles.

A SHELL EXCLAMATION MARK IS NOT FOR YELLING. BE LAZY

Event designators in shell commands allow users to save time and avoid repetition by referencing and modifying previous commands and their arguments without the need for excessive keystrokes.

I PUT GITHUB COPILOT BEHIND A MITM PROXY. HERE'S WHAT I FOUND

Analyzing GitHub Copilot's network traffic reveals its mechanisms for context processing, user interaction logging, and implications for privacy and data management.

JOINING CURSOR

After over five years at PlanetScale, Fatih Arslan is leaving to join Cursor, driven by a passion for developer tooling and the innovative potential of Cloud Agents.

Reserve your spot today