AI & Tech Brief — August 22, 2026

TL;DR

Kagi — the paid search engine — added a setting to strip paywalled links from results, and it blew up on HN (1,100+ points). A small feature that clearly tapped a deep well of paywall fatigue.
Two of today’s most-discussed pieces are about the friction of AI text: a study finding AI help raises homework scores but then lowers exam scores, and a viral Claude Code skill that pipes Claude’s output through another model just to make it stop talking like a TED talk.
OpenAI cut GPT-5.6 Sol’s price (20% lower input, 33% lower output) and now lets API customers pin regional processing per request — incremental, but useful if you care about cost or data residency.

Key Stories

Kagi adds a setting to remove paywalled links from search results Buried in Kagi’s August 21 changelog (mostly a Stocks-widget and Assistant polish release) is one line: “we’ve added a setting for removing paywalled links from search results automatically.” That single toggle drove the top HN thread of the day (1,142 points). The discussion is less about the feature than about how exhausted people are with paywalls, SEO sludge, and the state of search in general — and how a paid engine can just decide to do this. A reminder that there’s real appetite for search that treats the reader as the customer. Sources: https://kagi.com/changelog · https://news.ycombinator.com/item?id=49388154
Study: AI boosts homework scores, then exam scores drop An Economist analysis (314 points on HN) of AI adoption in learning finds a consistent pattern: homework scores rise ~18% and completion time falls ~30%, but monthly exam scores drop ~20% within six months. In other words, AI is a great shortcut for the work that’s graded along the way and a poor substitute for actually learning the material — the gains don’t survive when the tool is taken away. One of the cleaner pieces of evidence so far in the “is AI helping students learn” debate. Source: https://news.ycombinator.com/item?id=49357530
“Claudette” / nobuzz: a skill to make Claude stop talking like a BuzzFeed article A Claude Code skill (/debuzz) that takes Claude’s last response and runs it through Google’s Antigravity CLI (Gemini) to translate it from millennial-clickbait into plain English — “load-bearing assumption,” “the third one is the most instructive yet,” and all. It has colleague / manager / director modes and prints the translation verbatim so Claude can’t re-inject the voice being removed. 285 points on HN; the thread is a cathartic pile-on about Claude 5’s house style. Yesterday it was “Vomit,” today it’s this — the market for de-Claude-ifiers is apparently robust. Sources: https://github.com/adnanakil/nobuzz/blob/main/README.md · https://news.ycombinator.com/item?id=49388752
OpenAI: GPT-5.6 Sol price cut + per-request regional processing Two API changelog entries from August 21. GPT-5.6 Sol now costs $4/M input and $20/M output tokens (20% lower input, 33% lower output), with promotional pricing locked through at least November 21, 2026. Separately, API customers with a Global-geography project can now select regional processing for an individual request via a prefixed domain — useful for data-residency compliance without restructuring the whole project. Source: https://developers.openai.com/api/docs/changelog
Felony charges for a citizen who deleted phone data at the US border The biggest non-AI story on HN today (816 points, 945 comments): a US citizen is facing felony charges after using a duress PIN that wiped his phone when border agents compelled him to unlock it. The legal question is whether on-the-spot wiping counts as destruction of evidence once a search has begun (probably yes) versus wiping before reaching the border (probably fine). The thread is a long, heated debate on the border-search exception, duress passwords, and why “just don’t carry the data across” is the only reliable answer. Not tech news per se, but squarely relevant to anyone who travels with a locked-down device. Source: https://news.ycombinator.com/item?id=49386895
“I’m becoming AI-blind” A short, widely-shared essay (384 points) arguing that heavy exposure to low-effort AI-generated text has trained the author’s brain to disengage the moment it detects the telltale signs — the over-eager framing, the “internal reasoning” tone, the pitch-every-detail-as-a-breakthrough cadence. Like banner blindness, but for prose. Pairs neatly with the homework study and the de-buzzing tools: the cost of cheap AI text isn’t just bad output, it’s readers who stop reading. Sources: https://cymerys.com/w/im-becoming-ai-blind · https://news.ycombinator.com/item?id=49386699

Quiet but Interesting

Rust Glancer: a Rust LSP built for ~100x less RAM — a 4-month solo project offering an alternative to rust-analyzer that targets under 100MB for reasonable projects and indexes once to disk so editor restarts don’t re-index. It trades rust-analyzer’s incremental, in-memory model (salsa + rowan) for a frozen, filesystem-offloaded analysis — slower per-keystroke, but far lighter. Notably built with heavy LLM assistance that the author is upfront about. 261 points. Sources: https://rust-glancer.github.io/blog/hello-world/ · https://news.ycombinator.com/item?id=49393052
Dan Luu: “There’s no reason for software to be slow anymore” — argues that coding agents have collapsed the cost of performance work by orders of magnitude, so optimizations that used to need a rare specialist (JIT compilers, workload-specific tuning) are now worth doing almost casually. Includes a neat experiment where an agent did workload-specific optimization of ripgrep against his own query history in minutes. 436 points. Sources: https://danluu.com/perf-opt/ · https://news.ycombinator.com/item?id=49395628
Sub-50ms text-to-speech — Nari Labs on how they got a TTS model to respond in under 50ms time-to-first-audio, a threshold that matters for making voice agents feel conversational rather than walkie-talkie. 153 points. Sources: https://nari-labs.com/blog/qwen3-tts-speed-cost-frontier · https://news.ycombinator.com/item?id=49389952

Skip

“Felony Bench” (723 points) — a novelty/entertainment site, not tech news.
“Kobo can run apps now” (563 points) — fun e-reader hacking, but niche.
“Canada suspends trade negotiations with USA” — developing political news, off-topic for this brief.

Sources checked: Claude Code changelog (quiet today; last release Aug 21), Claude release notes (quiet since Aug 6), Gemini CLI changelogs (quiet since v0.54.0, Aug 6), OpenAI API changelog (Aug 21: regional processing, GPT-5.6 Sol pricing), Codex changelog (Apple Messages plugin, Site co-editing, shared thread snapshots), Superhuman AI (top story is the Aug 21 Claude legal/SMB automation issue, covered yesterday), ByteByteGo (Schema Evolution, Aug 20), Dario Amodei (quiet), Sam Altman (quiet), DeepMind blog (quiet since Gemini 3.7 Flash, Aug 13), HN front page.