Incredibly fast autocomplete with in-memory trie

p99 0ms autocomplete for 240 million domain names
6 minutes by Ruurtjan Pul

Wirewiki's autocomplete feels instant by fetching suggestions on keydown, before the user finishes pressing a key. The server uses an in-memory trie for popular domains and a compressed disk index for the rest, responding in under 15ms for 99% of requests. The main remaining bottleneck is network distance, since a single European server adds too much latency for users further away. Multiple geo-distributed servers would fix this but are hard to justify for a niche tool.

Free security tools for AI-powered coding
sponsored by Endor Labs

Security teams talk risk. Developers talk function. AURI by Endor Labs is the common language between them, grounded in shared context and evidence. AURI by Endor Labs brings security into the architecture of agentic coding, with free tools for developers and AI agents to detect vulnerabilities, block malware, and fix security bugs. Learn more and try it yourself.

The cost YAGNI was never about
4 minutes by Kent Beck

YAGNI means build things when you need them, not before. Building too early costs you twice: you lose the ability to make better choices once you have real information, and you spend money before you earn it back. AI tools make it tempting to build speculative structure for free, but that makes the problem worse, not better. The cost was never in the typing.

A story of delayed AWS pipelines
7 minutes by Vladimir Romashov

A single Terraform data source for AWS organization info secretly makes three API calls, not one. These calls share a rate limit across an entire AWS organization, so dozens of teams running pipelines at once exhaust that limit fast. This caused roughly 1 in 10 pipelines to stall for up to 15 minutes. A fix now shipped in terraform-provider-aws v6.29.0 makes the extra calls optional.

How LLMs actually work
23 minutes by 0xkato

Large language models turn text into integers, look up learned vectors for each, then pass everything through many stacked layers. Each layer lets tokens share information through attention, then processes each token independently through a feed-forward network. Residual connections and normalization keep deep stacks stable and trainable. The model ends by scoring every possible next token and repeating the process one step at a time.

curl and CVE dispute
7 minutes by Daniel Stenberg

The curl project became a CVE Numbering Authority, giving it full control over assigning CVE identifiers to its own security issues. Not every bug deserves a CVE, and curl deliberately skips them for issues so unlikely that raising an alarm would waste effort across the industry. A reporter recently disputed this approach over a wildcard certificate hostname bug, escalating to MITRE.

And the most popular article from the last issue was:

• Stop using Conventional Commits