Search the archive

Try Scandit Agent Skills for free

OFFICECLI

OfficeCLI is an open-source, AI-friendly command-line tool that allows for easy creation, editing, and automation of Word, Excel, and PowerPoint documents without requiring any external Office installation, allowing AI agents to manage documents using commands.

OTARI

Otari is an open-source, OpenAI-compatible LLM gateway that allows users to manage API keys, enforce budgets, and track usage across over 40 providers through a single endpoint, which can be run either standalone or hosted.

MAPCN

Mapcn offers free and customizable map components for React, built on MapLibre and styled with Tailwind, making it easy to create beautiful maps.

A STARGATE FOR DATA

AI labs are projected to spend over $100 billion per year on data by 2030, as current advancements in AI capabilities show a shift from a compute-limited to a data-limited regime. The scarcity of high-quality data from the internet, combined with the urgent need for data collection across various private and public domains, shows the importance of treating data as a strategic asset necessary for economic and scientific progress.

A GLOBAL WORKSPACE IN LANGUAGE MODELS

Recent research by Anthropic has identified a novel set of internal neural patterns in LLMs, referred to as the J-space, which distinguishes conscious processing from unconscious activity. The J-space, developed autonomously during training, serves as a mental workspace that helps with deliberate reasoning and reporting of thoughts, showing flexibility in linking concepts and performing various tasks, although it only represents a small portion of the model's overall processing.

GLM 5.2 AND THE COMING AI MARGIN COLLAPSE (PART 1)

GLM 5.2 is such a cost-effective model that it shows the possible collapse of AI margins as smaller models make more sense to use for daily tasks. This may change the entire economics of AI and training models.

COMPANIES HIRE MORE AFTER AI ADOPTION

Companies that adopt AI experience a 10% growth in headcount within two years, primarily driven by high-intensity adopters.

SMALL AI MODELS GAIN TRACTION AROUND THE WORLD

Small AI models are becoming effective solutions for addressing healthcare and agricultural challenges in areas with limited infrastructure, offering localized services without the need for extensive data centers or broadband connectivity.

THE SOFTWARE ENGINEERING WAR

The ongoing conflict in software engineering between "builders," who prioritize rapid feature deployment and immediate user feedback, and "keepers," who focus on meticulous coding and system integrity, reflects a broader industry debate over the balance between speed and quality in development practices with AI in the mix.

GPT-5.6 SOL ULTRA WILL BE IN CODEX (1 MINUTE READT)

GPT-5.6 Sol Ultra will be available to use on Codex.

HUNTING A 16-YEAR-OLD SQLITE BUG WITH TLA+: IS DQLITE AFFECTED?

SQLite recently addressed a long-standing bug related to the Write Ahead Log (WAL) that could cause database corruption, present in the system for 16 years. The dqlite team investigated whether this bug affects their implementation by modeling SQLite's behavior in TLA+ and comparing it with how dqlite interacts with SQLite and concluded that dqlite is not affected due to its stricter locking mechanisms that prevent simultaneous writes and checkpoints.

AGENTIC TEST PROCESSES, LLM BENCHMARKS, AND OTHER NOTES ON AGENTIC CODING FROM GALAPAGOS ISLAND

The exploration of coding agents and their testing capabilities shows a disconnect between the efficiency of AI in automating processes and the reliability of its outputs, especially in coding and debugging tasks. While AI models can speed up workflows and increase code quality through good testing methods, they often struggle with generating reliable tests and identifying the true source of bugs.

POSTGRESQL AND THE OOM KILLER: WHY WE USE STRICT MEMORY OVERCOMMIT

Strict memory overcommit in PostgreSQL is necessary for preventing catastrophic out-of-memory (OOM) kills, as it allows the database to handle memory allocation failures gracefully without impacting all connections. A kernel bug in Ubicloud's product temporarily disrupted this configuration, showing the importance of careful management of memory limits to guarantee system stability while running PostgreSQL in production environments.

BETTER MODELS: WORSE TOOLS

Recent iterations of Anthropic's models, specifically Opus 4.8 and Sonnet 5, show a troubling trend of producing malformed tool calls when interacting with Pi's editing tool, as they frequently append extraneous fields that do not conform to the expected schema. This regression seems to be linked to the models' training environments, which appear to favor a more forgiving approach to tool calls.

WRITE CODE, NOT SPECS

Good code should be self-documenting, allowing for an understanding of its functionality without extensive specifications. Instead of maintaining both code and specifications, focusing on writing code that establishes patterns and relying on tests to validate functionality allows for greater trust and efficiency in development.

SOME NEW AGENTIC PATTERNS

At Prime Radiant, new "agentic" patterns have been implemented to enhance collaboration between agents and human implementers, resulting in improved tools and workflows. Notable examples include an agent named Nora, which acts as a team member rather than just an assistant, and a framework allowing multiple agents to collaboratively address tasks, leading to efficient problem-solving and feature development.

PLANNOTATOR

Plannotator is a browser-based tool that enables users to annotate and review plans, specifications, and code changes from various AI coding agents. It integrates with multiple platforms, supports real-time collaboration, and allows for easy sharing of annotated content with team members.

BRAMBLE

Bramble is a local-first, open-source password manager that encrypts and stores user data on personal devices without relying on cloud servers, allowing for secure peer-to-peer syncing and various authentication methods, including biometric options and passkeys.

STATE OF CLI CODING AGENTS, MID-2026

As of mid-2026, the terminal has become a favored platform for coding agents, with 35 actively maintained CLI agents available. The main players like Claude Code, Codex CLI, and Omp have come up as leading options, each having features such as memory capabilities, editing precision, and orchestration.

PERFORMANCE PER DOLLAR IS GETTING FASTER AND CHEAPER

The release of GLM-5.2 shows advancements in inference performance, with 2626 tok/s/node on AMD MI355X at more than twice the cost-efficiency compared to similar NVIDIA setups. Despite historical challenges with AMD's software support, improvements in model optimization are narrowing the gap.

JAMESOB'S GUIDE TO RUNNING SOTA LLMS LOCALLY

The guide shows detailed instructions and recommendations for setting up a state-of-the-art local machine learning system using specific hardware configurations to run advanced LLMs properly.

WHAT ORMS HAVE TAUGHT ME: JUST LEARN SQL

ORMs can provide convenience in working with SQL but often complicate data retrieval and management due to issues like attribute creep, inefficient queries, and dual schemas.

REBUILDING COGNITION'S AGENTIC MAPREDUCE

Cognition's approach to scaling agent tasks, using Agentic MapReduce, enables parallel processing of security findings by sharding codebases, verifying vulnerabilities, and using Git for communication and fault tolerance.

IF YOU'RE A BUTTON, YOU HAVE ONE JOB

The differences in how iPhones and Nothing Phones handle image rotation shows the importance of designing user interfaces that accommodate situational disabilities and improve usability by preventing unnecessary wait times for animations.

HOW OPENAI DELIVERS LOW LATENCY VOICE AI FOR 900M USERS

OpenAI serves WebRTC voice AI to 900M weekly users by splitting the stack into a stateless edge relay for packet routing and a stateful transceiver that owns the heavy WebRTC state, sidestepping the port-exhaustion and state-stickiness problems WebRTC hits on Kubernetes. The key method is encoding routing metadata into the ICE ufrag, a field already exchanged during setup, so the relay routes the first packet by reading the initial STUN request rather than doing a database lookup.

FLOW

Flow is FoundationDB's custom language extension for C++. It adds efficient actor-based concurrency using asynchronous primitives like Future, Promise, wait(), and ACTOR, while compiling down to native C++ for high performance and deterministic simulation. Flow lets developers write distributed asynchronous systems in a much simpler style than callback-heavy C++, with additional constructs like streams (FutureStream/PromiseStream) and choose/when for coordinating concurrent events.

THE SHORT LEASH AI CODING METHOD FOR BEATING FABLE

Expert developers should keep AI coding agents on a “short leash” by reviewing every proposed change, intervening frequently, committing often, and never letting the AI work unsupervised. It also recommends using AI as a code reviewer alongside human review.

THE PRIMARY PURPOSE OF CODE REVIEW IS TO FIND CODE THAT WILL BE HARD TO MAINTAIN (2 MINUTE READT)

The main goal of code review is not to catch bugs but to identify code that is difficult to understand and therefore likely to be difficult to maintain.

OMNIROUTE

OmniRoute is an open-source AI routing gateway that lets developers connect coding tools like Claude Code, Codex, Cursor, and Cline to over 237 AI providers (including 90+ free ones) through a single OpenAI-compatible endpoint. It automatically handles provider failover, cost optimization, and quota management.

VALMIS

Valmis is a cloud platform for building and deploying AI agents that automate business workflows across 100+ integrations using chat, scheduled jobs, webhooks, or app events. It emphasizes security by running agents in isolated containers so AI cannot directly access your API credentials or local files.

MORE DETAILS ON FABLE 5'S CYBER SAFEGUARDS AND OUR JAILBREAK FRAMEWORK

Fable 5 is being redeployed with stricter cybersecurity classifiers that block prohibited and high-risk dual-use cyber requests while allowing clearly defensive uses. Anthropic uses a Cyber Jailbreak Severity scale to rate jailbreaks by capability gain, breadth, weaponization, and discoverability.

HOW EXPENSIFY USES AGENT-DEVICE FOR MOBILE BUG EVIDENCE AND PROFILING

Expensify uses agent-device, a tool that gives AI agents “hands and eyes” on real devices by hiding platform fragmentation behind one API and returning the screen as a structured accessibility tree rather than a screenshot (richer and cheaper in tokens). They apply it to three repeated chores: an agent-device-evidence skill that reproduces a bug, distills the steps into a replayable script, and captures before/after recordings for PRs; automated Sentry span measurement across multiple runs; and one-prompt React profiling that returns component-specific render data.

CARPLAY IS ADDITIVE

Rivian should support Apple CarPlay as an optional feature because it gives drivers access to thousands of familiar apps without replacing Rivian's native interface.

THE FALL OF THE THEOREM ECONOMY

The hardest part of mathematical research is not proving theorems but discovering the right concepts and language that make those theorems almost inevitable.

INTRODUCING THE SAFARI MCP SERVER FOR WEB DEVELOPERS

Apple's new Safari MCP server lets AI coding agents connect directly to a local Safari browser to inspect the DOM, console, network, screenshots, performance, and accessibility.

ZEROFS

ZeroFS is an open-source filesystem that makes S3-compatible object storage behave like a high-performance, encrypted POSIX filesystem or block device, allowing applications to use cloud object storage almost like a local disk.

WHAT TO LEARN TO BE A REAL TIME GRAPHICS PROGRAMMER

To become a hireable graphics programmer, one must understand both the CPU and GPU sides of modern rendering, understanding APIs like DirectX12 and Vulkan, as well as mathematical concepts related to lighting and rendering techniques. Learning about Physically Based Rendering (PBR) is necessary, as it standardizes lighting representation across varying conditions, thus streamlining asset creation. Familiarity with C++ is needed for CPU programming, while knowledge of shader languages like HLSL and GLSL can be beneficial.

META'S AI STORAGE BLUEPRINT AT SCALE

Meta's storage architecture has evolved to meet the demands of modern AI workloads, addressing challenges such as maximizing GPU utilization and speeding up research velocity. By rebuilding the metadata subsystem and adopting a tiered caching approach with prefetching and on-demand hydration, its new BLOB-storage system improves performance, reducing data ingestion times and facilitating more efficient access for researchers.

CLIENT-SIDE LOAD BALANCING AT A MILLION REQUESTS PER SECOND

A client-side load balancer was developed to handle over a million requests per second of internal traffic, migrating routing decisions from a shared edge load balancer, Skipper, to improve latency and system resilience while reducing costs. It included features like occupancy-based bounded load and AZ-aware routing strategies to improve performance and efficiency.

MOST REWRITES SERVE THE ENGINEER, NOT THE BUSINESS

Most code rewrites cater to the preferences of engineers rather than the actual needs of the business, leading to a loss of valuable historical context and solutions that reside in the existing code. A rewrite should only be pursued when there are measurable issues, such as outdated dependencies or technical debt, rather than personal style choices.

DARK MODE WITH WEB STANDARDS

To improve user experience with dark mode on websites, developers should respect the user's system preferences while also allowing custom toggles for per-site settings.

REDEPLOYING CLAUDE FABLE 5

Access to the AI models Claude Fable 5 and Mythos 5 was briefly suspended due to newly imposed US government export controls, but it has now been restored following the lifting of these restrictions. Anthropic is collaborating with government partners to improve cybersecurity measures and create a consensus framework for assessing and addressing potential vulnerabilities and “jailbreaks” in AI models.

ZCODE 3.0

ZCode 3.0 introduces an optimized platform for easy collaboration using AI agents, enabling developers to plan, code, review, and deploy.

OPENWIKI

OpenWiki is a CLI designed to write and maintain documentation for codebases specifically for agents. Users can initialize the tool to generate documentation, update existing documentation, and customize various model settings.

SANDBOXING AN AI AGENT

The increasing autonomy of AI agents raises security concerns as they gain the ability to execute commands without explicit human approval, creating potential risks associated with private data access and untrusted content. Implementing sandboxes provides isolated environments for these agents, improving containment, resource management, and recovery processes while allowing for experimentation and parallel execution without risking the host system.

HOW WE KEEP GPUS RELIABLE ACROSS DATABRICKS AI

GPU failures during distributed training can primarily be categorized into crashed jobs, silent slowdowns, and numerical corruption, each posing distinct challenges for maintaining reliable performance. To address these issues, a comprehensive health check system at Databricks AI uses both proactive stress testing of workloads and continuous monitoring to detect and mitigate failures across the GPU infrastructure.

A TALE OF TWO MODEMS

Apple plans to equip the US version of the iPhone 18 Pro with Qualcomm modem hardware to support mmWave technology, while the international versions will rely on Apple's proprietary C2 modem, which doesn't have mmWave capability. This choice may prioritize carrier marketing over user experience, as Qualcomm modems offer faster speeds at the expense of battery life.

OF COURSE META PLATFORMS IS GOING TO BE A CLOUD

ARXIV'S NEXT CHAPTER: UPDATES ON OUR SPIN OUT FROM CORNELL UNIVERSITY

arXiv will become an independent nonprofit organization on July 1 after 25 years at Cornell University.