ClickHouse scaled its internal observability platform, LogHouse, to over 100 petabytes by developing SysEx, a specialized system for ClickHouse-to-ClickHouse data transfers. SysEx, unlike OTel, preserves native ClickHouse types and removes intermediate conversions, resulting in a 20x increase in event volume handled with significantly less CPU usage. They use HyperDX, a ClickHouse-native UI, for seamless data exploration, correlation, and root-cause analysis, while still using Grafana for existing dashboards.
There are different types of race conditions in Go concurrency. This article goes over how to handle them. It differentiates data races (concurrent access to shared data with at least one write) from race conditions (unpredictable order leading to incorrect state), then explains several approaches to handling race conditions, including mutexes, compare-and-set operations, and idempotence.
HoneyBook engineers used Cursor to upgrade Storybook from v6 to v8 in just 2 hours, a task that would usually take a developer about a week. They broke the upgrade into small, manageable tasks with clear success criteria: first establishing a minimal working Storybook v8 setup with a dummy story, then systematically migrating individual stories using Storybook's migration guide as context.
Good salary negotiation involves shifting the mindset to embrace negotiation as a normal business practice instead of having feelings of discomfort or moral objection. Don't give a number first. Instead, thoroughly research and focus on the overall compensation package beyond just salary.
2025.06.23·
complexsystemspodcast.com·45 min read·TLDR Web Dev
The React team's push for frameworks, while intended to improve performance, lacks nuance and has led to confusion and dismissiveness towards traditional SPA approaches. There also needs to be better React documentation on Server Components.
"Levered beta" might be a winning strategy in the age of rapidly improving AI, where being early and loud matters more than building a superior product. In markets driven by trends and quick advancements in AI models, companies focused on distribution and claiming mindshare can thrive, even with mediocre products. These companies are essentially betting on future AI improvements to validate their early market entry and aggressive marketing, positioning themselves as the default choice when the technology finally catches up.
In this article + video + repo [16], we'll show you how to build a tabbed in-app feed where users can mark notifications as read, archive them, and switch between views. You'll use Next.js and shadcn/ui to build the Notion-style interface, and the Knock feeds API and JavaScript client to get real-time notifications. View the tutorial
Nxtscape is an open-source, privacy-focused browser that integrates AI agents to automate tasks locally, with a user experience similar to Chrome. It allows users to connect their own AI providers or run local models.
Phoenix.new is a new online coding agent for Elixir and Phoenix development. It provides a root shell within an ephemeral virtual machine, allowing the agent to install packages and run programs without affecting the user's local machine. This agent understands real-time collaborative applications and includes a headless browser to interact with and test its own front-end changes.
SnapQL is an AI-powered desktop application designed to explore and query PostgreSQL databases quickly. It generates schema-aware queries using AI and supports any PostgreSQL database while keeping credentials local.
Startups often fall into p-hacking traps due to pressure for quick results. Common p-hacking practices include ignoring corrections for multiple comparisons, shifting metrics after seeing results, and prematurely stopping experiments. To avoid p-hacking, startups should pre-register hypotheses, use appropriate corrections, and avoid peeking at results before the experiment's completion.
This is an in-depth technical overview of Google's Tensor Processing Units (TPUs). TPUs prioritize scalability and energy efficiency by using systolic arrays and Ahead-of-Time (AoT) compilation via the XLA compiler. The article goes through the TPU architecture from the single-chip level with its TensorCores and Matrix Multiply Units to multi-chip configurations, including trays, racks, pods, and multi-pod setups.
mbake is a Makefile formatter and linter that automates style consistency, .PHONY management, and syntax validation in Makefiles through configurable rules, CI/CD integration, and a plugin architecture.
Anthropic has announced support for remote MCP servers in Claude Code, letting users integrate their favorite tools and data sources for a personalized coding experience.
Null pointer exceptions are a common and costly problem in software that can be prevented through validation, null-safe patterns like the Null Object design pattern, and defensive programming practices.
Mark Zuckerberg's AI hiring spree has now turned to Daniel Gross, the CEO of Ilya Sutskever's startup Safe Superintelligence, and former GitHub CEO Nat Friedman. Meta tried to acquire Safe Superintelligence earlier this year, but its efforts were rebuffed. Zuckerberg started negotiating with Gross soon after those talks ended. Both Gross and Friedman will join Meta and work on products under Alexandr Wang. Meta will get a stake in NFDG, a venture capital firm run by Gross and Friedman.
X users will soon be able to make investments and trades on the social media platform. The company is also exploring the introduction of an X credit or debit card. Elon Musk aims to change X into an 'everything app' - a one-stop shop for messaging, payments, and shopping. X Money, a digital wallet and peer-to-peer payment service, is launching later this year. The push into financial services will open X up to regulatory challenges, such as compliance with licensing and money laundering regulations.
The US Food and Drug Administration has approved a new HIV-prevention medication called Yeztugo. The drug nearly eliminated the spread of HIV in clinical trials. It is the first in a new class of antiretrovirals that block HIV from infecting and making new copies of itself inside the immune cells it targets. It costs $14,109 per injection, so health insurers will likely decline to cover the treatment in favor of other forms of PrEP, which can cost as little as $30 per month.
A SpaceX Starship rocket exploded during a ground test late Wednesday. The explosion destroyed the upper stage that was slated to launch on the next Starship test flight. Preliminary data suggests that the incident was caused by the failure of a high-pressure nitrogen tank inside Starship's payload bay. SpaceX had set June 29 as a tentative launch date for the next Starship test flight, but that won't happen now. The extent of the damage at the site is unclear, so it is unknown how long the test site will be out of commission. Videos of the explosion are available in the article.
Delve ran 4 ads across 2 TLDR newsletters and brought in 66 leads, with a meaningful chunk from enterprise. The result: $1M in attributed pipeline and a 52x ROI.
Software development tools are getting killed off and replaced by agentic coding. The typical agentic coding workflow doesn't involve using traditional IDEs - it is console-based. This lets developers work faster, be more ambitious, and have more fun. Developers should start learning how to use these tools or start thinking about other career options.
Kubernetes is a massive force multiplier, but it has a steep learning curve. A decade after its release, there is still a lot of churn inside the ecosystem and people stepping on well-documented landmines. This post looks at what can be added to Kubernetes to make the great tool even more applicable to more people and problems.
DARPA is enlisting researchers to find out ways to conduct high-level mathematics research with an AI 'co-author'. It has created a grant-making program called Exponentiating Mathematics to speed up to pace of progress in pure math. Pure math is a domain where visionary theoreticians make audacious observations about how the world works. Progress in the field can take decades or centuries to solve.
There's an increasingly sharp distinction between corporate use of open source software and people's cooperative use of it. While open source licenses allow corporate use of open source software, it is qualitatively different than the personal cooperative sphere of open source. This means that the social rules are different. Existing open source licenses, practices, and culture don't draw this distinction. The current system for corporate use of open source is an increasingly bad deal for the open source people involved.
The Base Commerce Protocol offers a glimpse into a new payment architecture that is transparent, programmable, and structurally more efficient than what we've relied on for decades.
The best engineering organizations are the ones where normal software engineers can consistently make progress, deliver value to users, and move the business forward.
OpenAI has changed the pricing structure of the business version of ChatGPT to include a credits system that clients can use to upgrade to more advanced tools and add more features.
Early testers are calling GPT-5 "materially better" than GPT-4, though Sam Altman gave no specific launch date for the new model beyond summer. Altman floated advertising possibilities but drew a hard line against letting payments influence responses, suggesting ads might appear outside the model's output stream.
MiniMax's second-generation video AI model, Hailuo 02, features major upgrades in both performance and price. It uses an architecture called Noise-aware Compute Redistribution that improves training and inference efficiency by a factor of 2.5. The architecture handles long video sequences differently depending on the stage of training. Hailuo 02 has three times more parameters and four times more training data compared to its previous version. A video generated with the model is available in the article.
The first comprehensive model of LLM serving economics reveals why current approaches to scaling inference hit walls faster than expected, as AI companies race to serve token-intensive reasoning models and agents. Network latency, not bandwidth, creates the primary bottleneck that prevents companies from simply adding more GPUs to increase capacity. Algorithmic breakthroughs like speculative decoding, which delivers double the speed at no additional cost, continue to reshape the economic landscape as providers struggle to match surging demand.
Recommendation and search systems are shifting from item IDs to rich "Semantic IDs," generative retrieval, and multimodal embeddings, enabling cold‑start coverage, long‑tail discovery, and unified search‑recs architectures that scale efficiently.
Traditional large language model (LLM) systems often rely on sequences of GPU kernel launches and external communications calls, which results in underutilized hardware. This post discusses how a team created a compiler that automatically transforms LLM inference into a single megakernel, which eliminates launch overhead, enables fine-grained software pipelining, and overlaps computation with communication across GPUs. The end-to-end GPU fusion approach reduces LLM inference latency by 1.2 to 6.7 times.
This document lists major changes made to the Model Context Protocol (MCP) specification since the previous revision, 2025-03-26. Some of the changes include the removal of support for JSON-RPC batching, the added support for structured tool output, and the clarification of security considerations and best practices in the authorization spec. A link to the complete list of all changes is available.
StochasTok randomly decomposes tokens during training: instead of always seeing "strawberry" as one unit, models encounter it split as "straw|berry," "str|awberry," or even "s|t|r|a|w|b|e|r|r|y," learning the internal structure humans naturally perceive. Models trained with this method achieve near-perfect accuracy on character counting and multi-digit math while maintaining performance on standard benchmarks.
An end‑to‑end variational encoder that augments semantic speech tokens with automatically learned prosodic features, removing hand‑engineered pitch inputs and yielding more natural continuations in human preference tests.
Israeli developer Maor Shlomo recently sold his six-month-old, bootstrapped vibe-coding startup, Base44, to Wix for $80 million cash. His eight employees will collectively receive $25 million of the $80 million as a 'retention' bonus. Base44 grew to 250,000 users in six months. It generated $189,000 in profit in May even after covering high LLM token costs. The startup grew mostly through word of mouth.