A short Saturday recap. This was a week where the open weights leaderboard shifted again, a significant researcher made a move that got 220 HN comments, a security story made me reconsider part of my ETL, and two unrelated things in the background of dev infrastructure are worth noting.
GLM-5.2 is now the leading open weights model
On June 17, Artificial Analysis updated their leaderboard and GLM-5.2 — from Zhipu AI — took the #1 spot among open weights models. The HN thread reached 734 points and 369 comments, which means it got real scrutiny rather than just upvotes.
I run a HuggingFace ETL for aiappdex.com that pulls models by pipeline tag and like count. GLM-5.2 didn't appear in my last refresh because it lands in the text-generation tag without a clean "chat" or "instruction-tuned" marker in the model card. This is the second time I've missed a highly-ranked model for that reason.
My planned fix: add a secondary pass for models above 2000 HuggingFace likes regardless of pipeline tag, then cross-reference against Artificial Analysis and LMSYS rankings monthly. Chinese open models seem consistently underrepresented in English-language coverage relative to their benchmark position. Worth tracking.
Noam Shazeer joined OpenAI
This landed June 18 with 251 HN points and 220 comments. Noam Shazeer co-invented the Transformer architecture at Google Brain, left to co-found Character.AI, and Google later spent approximately $2.7B to effectively re-acquire the team. Now he's at OpenAI.
I don't have a firm read on what this means strategically. What I do notice: OpenAI keeps winning the talent competition at the research layer. If you're building on any specific model provider's API and care about long-term architectural direction, this is the kind of signal to watch — not because it changes anything this month, but because it influences what gets built in 12-18 months.
For my own project I'm using Claude Haiku 4.5 for content generation and I'm not switching. But I track these moves because they shift where fundamental research happens next.
10,000 GitHub repositories distributing Trojan malware
The post at orchidfiles.com reached 620 HN points. The finding: roughly 10,000 repositories were actively distributing Trojan malware, typically disguised as game cracks, productivity tools, and video converters.
This matters to me directly. ossfind.com pulls repository data from GitHub to populate the OSS alternatives directory. My ETL checks license fields and star counts — using star count as a proxy for legitimacy, on the assumption that malicious repos don't accumulate thousands of real stars.
That assumption may be wrong. The orchidfiles post didn't detail the star counts on the affected repos, but the scale (10k repos) suggests some had at least nominal apparent popularity. I don't have a good solution yet. I've tightened my minimum threshold and I'm checking whether the repo owner has any activity history, but that's a heuristic, not a filter. I'll publish whatever I've actually implemented in 30 days.
RFC 10008: HTTP gets a QUERY method
A new RFC standardizing the HTTP QUERY method reached 301 HN points. QUERY is semantically identical to GET — safe, idempotent, cacheable — but allows a request body. This solves the longstanding awkward pattern of using POST for reads that need structured query parameters too large or complex for a URL.
My three sites are fully static Astro SSG right now, so this doesn't affect anything I'm running. But I've been thinking about adding a filter or search endpoint to the directories. If I do that, QUERY is the correct method. Browser and CDN support will take time to propagate, but RFC publication is where the clock starts. Worth noting now.
Lore: a new version control system
The top HN story on June 17 was Lore — 880 points, 491 comments — an open source VCS designed for large-scale monorepos. The README describes it as targeting billions of files and hundreds of thousands of contributors.
I run a five-workflow GitHub Actions monorepo covering three sites. It works. Git at this scale produces some large history artifacts I don't want and has slow blame operations on content files, but it's manageable.
I'm not switching from git. Lore is early-stage, has no ecosystem, and git's value is mostly the ecosystem — every CI provider, every deploy platform, every developer on the planet. A new VCS has to solve a problem severe enough to overcome that inertia. "Monorepo at Google scale" is a real problem but not mine.
What's worth watching: whether Lore's architecture ideas get absorbed into git tooling (the way Mercurial influenced several git features). That's usually how these contender projects matter over time.
Five things from a packed week. The open model rankings moved again, a major researcher changed teams, GitHub ETL has a safety problem I haven't fully solved, a new HTTP method got standardized, and a git alternative made noise without quite clearing the ecosystem hurdle.
Part of an ongoing 6-month experiment running three AI-curated directory sites. The technical claims here are real; this article was AI-assisted.
Top comments (0)