daily.dev · Popular · 实时热榜

AI coding tools are moving from flat subscriptions to metered, usage-based pricing. GitHub Copilot went from unlimited access to premium request multipliers to full token billing starting June 1, 2026, and is retiring annual plans. Anthropic restricts via rate limits rather than repricing, with weekly caps on Claude Code since August 2025 and metered-from-first-token pricing for Enterprise. Cursor has repriced repeatedly, with an email reportedly sent to subscribers announcing that Auto pricing

Domain-Driven Design's core ideas—understanding the problem domain, model-driven design, knowledge crunching, and Ubiquitous Language—become more important, not less, as AI agents take over implementation. The domain model's value lies in the shared team understanding it builds, not the artifact itself; an AI-generated spec no one reads solves nothing. Precise, consistent naming (Ubiquitous Language, Bounded Contexts) sharpens prompts and prevents agents from conflating distinct concepts. Teams

An architectural argument that multi-agent AI systems should be designed by modeling the domain first — its rules, state, and invariants — rather than starting with agent workflows. Business logic embedded in prompts (e.g., 'never approve claims above €10,000') is fragile because models can misread or ignore natural-language rules. Encoding those constraints as domain errors instead gives agents a feedback loop: when an action is rejected, the agent can reason, retry, or request human approval.

CSS is getting a new Class Prefix Selector (.prefix-*) that lets developers target all classes sharing a hyphenated prefix, such as .btn-primary and .btn-secondary, without listing them individually or adding a base class. The proposal, originally pitched by Lea Verou in 2024, was resolved just two weeks ago at the CSS Working Group F2F meeting in Berlin (August 2026) and now exists only in spec text for CSS Selectors Level 5. It won't match the empty string (class="foo-" is not matched by .foo-

A comedic breakdown of how coding interviews have evolved at frontier AI labs like OpenAI and Anthropic, contrasted with everyday hiring at non-tech companies. Reddit-sourced interview loops reveal grueling take-home projects (webhook delivery systems, web crawlers, in-memory databases), line-by-line code interrogation, and system design rounds testing LLM inference APIs and profiler internals, all done without AI assistance. Meanwhile companies like Nordstrom, Walmart, and Kellogg's have adopte

Go 1.27 shipped with the usual pile of language tweaks: generic methods, struct literals that accept any valid field selector as a key, and function type inference that now works across all assignment contexts. There’s also a 30% cut in small-object allocation costs, encoding/json/v2 , post-quantum signatures via crypto/mldsa , a native uuid package, and a goroutine-leak profiler that’s now generally available. Solid release. Nobody’s arguing about any of that. What people are actually talking a

A creator benchmarks ScriptC, an experimental tool from Vercel Labs that compiles TypeScript directly to a debuggable C backend or LLVM IR (and then a static binary), against Node.js and Bun. Across minimum memory usage, million-object allocation, and server startup time tests, ScriptC dramatically outperforms both: 2.2MB vs 40MB (Bun) vs 70MB (Node) minimum memory, 88MB vs 176MB vs 259MB with a million objects, and 2.3ms vs 15.8ms vs 56ms startup time. ScriptC achieves this by compiling TypeScr

A satirical recap of a wild summer in which AI models allegedly disproved several long-standing math conjectures, including the Jacobian conjecture and a 30-year-old graph theory problem, and made progress on the Riemann hypothesis. Highlights include OpenAI shipping ten solved problems with formal Lean proofs, Anthropic's unreleased model (prompted by Bun creator Jared Sumar) running 60 sub-agents and burning 31 million tokens to boost a proof-adjacent metric, and Terence Tao warning about a cr

A deep dive into two popular collections of AI agent 'skills' (markdown-based instruction files): Matt Pocock's skills repo (200k+ GitHub stars) and Lauren 'Potato' Tan's Pstack skills. The reviewer tests standout skills like unslop (removes AI writing tells and adds human voice), grilling/grill-me (relentless interview-style planning), blast-radius (impact analysis before changes), show-your-work, wizard, writing-for-agents, and wait-what. Includes live before/after comparisons showing how unsl

Explores what LINQ actually is—a query pattern paired with provider-specific translation—rather than a universal query language. LINQ to Objects and LINQ to XML feel seamless because they operate on similar tree/collection math, while LINQ to Entities (EF Core) leaks abstraction because it bridges C#'s imperative model with SQL's declarative, set-based one, sometimes silently falling back to expensive client-side evaluation or generating suboptimal SQL. A practical framework is offered: use LINQ

A study by independent developer Théo Ducreux, using his project ValidateHTML, scanned the 5,000 most-visited web domains and found that 87.2 percent violate HTML or CSS specifications, with only 12.8 percent fully valid and just 2.6 percent producing zero errors or warnings. Over a third failed accessibility checks, with 20.4 percent missing alt text and 41.6 percent missing ARIA labels. The most common issue, found in over 59 percent of sites, is incorrect tag nesting, which Ducreux attributes

A personal reflection arguing that AI-generated content and vibe coding drain mental energy rather than provide it, citing ThePrimeagen's public frustration with vibe coding on stream (hating the code it produces, feeling guilty for not matching hype, and wanting to return to 'tradcoding') alongside a mention of a YouTube video exploring AI coding's mental health effects.

Godot 4.7.2 is a maintenance release with 57 fixes from 39 contributors, addressing bugs in core, editor, GUI, input, networking, navigation, rendering, and platform support. Notable fixes include preventing multiple main threads, improved mouse input performance on Windows with high polling rates, and updated entropy source handling in mbedTLS. No known incompatibilities exist with the prior 4.7.1 release, and users are encouraged to upgrade.

Somehow, in 2024, the internet is having a real argument about dark mode toggles. Not dark mode itself. The toggle . Specifically: should it have two states or three? Lea Verou, who works on web standards and clearly has opinions people listen to, says two is enough. Her logic: a persistent, always-visible toggle can flip between light and dark, defaulting to system preference on first visit and storing an override in localStorage after that. Adding a third “System” option, she argues, just adds

A revisit of Kent Beck's Extreme Programming Explained, thirty years after XP's creation, arguing its five values, fourteen principles, and 24-26 practices remain relevant amid AI-driven software development. The piece breaks practices into 'more critical than ever' (whole team, team continuity, energized work, real customer involvement), 'controversial' (pair programming, code-and-tests-only, test-first programming), 'common sense', and 'still relevant?' categories, weaving in commentary on AI

A deep dive into caching for system design covering how caches work internally, where copies of data can live (in-process, Redis/Memcached, CDN, browser, database), and the tradeoffs of each layer. It walks through Redis internals (hash table resizing, approximated LRU/LFU eviction via sampling, lazy and active expiry), Memcached's slab allocation and rebalancing history, HTTP caching semantics (no-cache vs no-store, ETags, stale-while-revalidate, stale-if-error), CDN tiered caching, and why dat

A general advice piece arguing that becoming a better programmer isn't about shortcuts, tools, or AI, but about consistent habits: going beyond tutorials to learn fundamentals, experimenting with working code, stepping outside your comfort zone, taking ownership of hard problems, contributing to open source, surrounding yourself with skilled developers, and teaching what you learn. The overarching message is that steady habits compound into real growth over time.

An overview of Tiger Style, the coding philosophy behind TigerBeetle, the financial transactions database, exploring its priority order of safety, performance, then developer experience. Covers explicit limits on queues and buffers, fixed integer widths over architecture-dependent types like usize, avoiding dynamic allocation after startup, heavy use of assertions, batching, cache-conscious data layout, single-threaded transaction processing to avoid contention, minimizing data copying in hot pa

An engineering leader shares why AI-based code review tools like Cursor's BugBot and custom Claude code review skills became too slow, noisy, and expensive (hitting $1000/day in one case) for a team producing thousands of PRs a quarter. The fix that worked: converting team-specific coding rules from markdown guidelines into custom linters, which run in seconds, are deterministic, can run in-editor and on pre-commit hooks, and stop agents from ever pushing bad code. Several example custom lint ru

Hi Persons, i am making a Rust version of Skyrim with Bevy, inspired by the OpenMorrowind. I want to make it and i will, this will help modders and gamers play this amazing game on Linux, android, windows and i am aiming for Mac too. But recreating the engine isn’t enough for me, i tough to my self. Why not making an modern launcher, make it a mod manager, so gamers can search, install and organize mods. But this is not enough, why not convert the game files? Let’s do that too, convert the files