HN Flash

Hacker News. Daily summary. Top 20 stories.

2026-07-28

Kimi-K3 Releases on HuggingFace 7/27 (huggingface.co)

Article body wasn't reachable. HN discussion still summarized.

The thread debated whether inference pricing on a 3T model will reveal API subsidy margins.

1354 pts · 536 comments

Our position on open-weights models (anthropic.com)

Anthropic's CEO Dario Amodei clarifies the company does not advocate banning open-weights AI models.

Commenters split between seeing Anthropic's stance as reasonable and as hypocritical ladder-pulling disguised as safety concerns.

1039 pts · 1519 comments

AI companies are shredding rare books (Nitter)

AI companies buy rare books, scan them by cutting off spines, and shred the originals.

Commenters liken it to book burning, questioning the rarity of destroyed titles.

771 pts · 491 comments

How is the Bun rewrite in Rust going? (Tom Lockwood - Developer)

Analysis questions the costs and completeness claims of Bun's Rust rewrite by Anthropic.

Skepticism dominates: hidden costs, unidiomatic Rust, and unreleased tags undercut the triumphalist $165k narrative.

478 pts · 375 comments

7.1 Earthquake in Japan (data.jma.go.jp)

A magnitude 7.1 earthquake struck the Kumamoto region of Japan on July 28, 2026.

Anxiety runs high as damage mirrors 2016's deadly Kumamoto quake, with tsunami warnings and structural collapses reported.

476 pts · 88 comments

Kimi-K3 Technical Report [pdf] (github.com)

Article body wasn't reachable. HN discussion still summarized.

Split on whether open-weight models accelerate or decelerate AI progress; license terms trigger debate on enforcement.

385 pts · 171 comments

Netflix employee fired for sharing personal details in retreat trust exercise (California Post)

Article body wasn't reachable. HN discussion still summarized.

Skepticism of workplace trust exercises and criticism of Netflix's handling dominate the thread.

375 pts · 402 comments

A missing underscore sent innocent man to prison for 18 months (Ars Technica)

A one-underscore typo in a Skyrim username sent an innocent man to prison.

Shock that a judge convicted with no evidence beyond a wrong username.

351 pts · 208 comments

Benchmarking Opus 5 on SlopCodeBench (GitHub)

Article body wasn't reachable. HN discussion still summarized.

Opus 5 called incremental, prone to pedantic slop; Codex Sol preferred for coding.

345 pts · 90 comments

Decathlon Germany adds Wero payment option to decathlon.de website (Sporting Goods Intelligence)

Decathlon Germany now accepts Wero, a European payment system, on its website.

Most see Wero as a strong European competitor to Visa and Mastercard, similar to iDeal and Bizum.

332 pts · 221 comments

Judge Rejects Google's Attempt to DMCA Its Way Out of Being Scraped (Techdirt)

A judge dismissed Google's DMCA anti-circumvention lawsuit against data scraper SerpAPI.

Commenters split: Google's hypocrisy versus SerpAPI's unethical scraping methods.

329 pts · 136 comments

Using an open model feels surprisingly good (matthewsaltz.com)

Running an open model on a private endpoint feels freeing and cuts ties with providers.

Skepticism about the post being a Modal ad, with two camps on cost versus privacy of local AI.

303 pts · 125 comments

AI companies spend record sums on Washington lobbying (ft.com)

AI companies spent a record $2.22M and $3.53M on federal lobbying in H1 2026.

Two camps: those arguing lobbying is cheap transparency and those calling it a pipeline for corruption.

270 pts · 143 comments

Paged Out #9 [pdf] (pagedout.institute)

Paged Out #9 is a free, one-page-per-article hacker zine with technical and whimsical content.

Thread celebrates the zine's technical depth, with detailed breakdowns of obfuscated C and computable tiling articles.

269 pts · 30 comments

Canceling "Hey" (chadnauseam.com)

Canceling Hey subscription after DHH compared Roma people to wolves in a blog post.

Most commenters condemn DHH's post as bigoted, citing his history of far-right sympathies.

266 pts · 227 comments

A $500 RL fine-tune of a 9B open model beat frontier models on catalog review (Fermisense)

Fine-tuning a 9B open model with RL beats frontier models on catalog review at 40x lower cost.

Skepticism about the benchmark; agreement that fine-tuned small models make economic sense for narrow tasks at scale.

252 pts · 80 comments

Removing React.js from the codebase and adapting Htmx for UI interactivity (2023) (Misago Forums)

A forum developer replaces React.js with HTMX to eliminate code duplication and complexity.

Sharp camp split: HTMX as overcomplicated throwback vs. pragmatic fit for most sites.

248 pts · 182 comments

Apple Will 'Watch Everything Burn' When the AI Bubble Bursts (MacRumors)

Ed Zitron argues the AI bubble will burst, costing consumers via higher hardware prices.

Deep split between those praising Zitron's evidence-based skepticism and those dismissing him as a biased crank.

242 pts · 327 comments

Should you wash your solar panels? (incoherency.co.uk)

Washing a small solar farm yielded a 2%-5% power increase, worth about £60-£150 per year.

Safety debate: grounding requirements for transformerless solar systems and touch-potential risks.

237 pts · 239 comments

MAI-Cyber-1-Flash inside MDASH (Microsoft AI)

Microsoft announces MAI-Cyber-1-Flash, a cybersecurity model inside MDASH offering 96% CyberGym score at half the cost.

Skepticism over closed access and marketing language, with comparisons to open-weight alternatives and Microsoft's own product quality.

235 pts · 110 comments