The anonymous model topping OpenCode's charts is Zhipu's, and the weights drop tonight. Plus Debian's eight-way vote on AI contributions, GitHub's 95% false-positive cut, and the in-house agent raising 75% of Ramp's merged PRs.
Z.ai confirmed today that it, not some open-source lab, is behind Ox Alpha — the anonymous model that appeared on OpenRouter last week and quickly became the most-used model on the open-source coding agent OpenCode, well ahead of DeepSeek. Weights are coming. Debian's developers are voting on whether AI-assisted contributions belong in the project at all, choosing among eight competing proposals. And GitHub published real numbers from its own work evaluating LLMs before shipping them: a 95% cut in false positives on its secret-scanning system. And The Pragmatic Engineer got inside Inspect, the in-house coding agent that now raises 75% of fintech Ramp's merged pull requests. Four stories about who checks, who votes, and who builds before anyone trusts the machine.
A stealth model gets a name #
Z.ai confirms it built Ox Alpha and will release the weights
Z.ai confirmed today it built Ox Alpha, an anonymous model that surfaced on the marketplace OpenRouter last week and became the most-used model on OpenCode, a popular open-source coding agent. OpenCode's own dashboard ranks it first, nearly double second-place DeepSeek's usage. Community testers had already traced it to Zhipu's GLM line by matching all eleven of its tokenizer probes; Z.ai now confirms that and says it will open the weights. Developers running it for real coding work on Hacker News call it roughly Sonnet-to-Opus level, flag occasional infinite command loops, and suspect a compressed cut of a larger model. One commenter's caveat is worth keeping: it ranks below GPT-5.4 Nano on LiveBench even as it tops OpenCode's chart. Read any single leaderboard with care. It's a free, open-weight model with real adoption, not just a lab's own claim.
Rules for the code AI writes #
Debian votes on whether to ban or allow AI code contributions
Debian is running a General Resolution on AI-assisted contributions: developers choose among eight proposals, from an outright ban to a conditional allow requiring disclosure. Proposal A needs a 3:1 majority; the rest need a simple majority. Project lead Sruthi Chandran extended voting a week; turnout is thin, under 350 ballots so far. Gentoo and NetBSD already banned AI contributions. However this lands, it sets precedent for how open source treats code a coding agent writes.
GitHub's own LLM evaluation work cut false positives by 95%
GitHub shared how it evaluates LLMs before shipping them, using its secret-scanning system (catching leaked credentials) as the case study: discipline over benchmarks cut false positives 95% on an offline test set while holding recall inside its guardrail. The lessons: distrust production labels, since they reflect outcomes, not truth; keep test sets messy, not clean; use an LLM as a triage judge, not the final call. No model names or costs appear. Process is the point, not a scoreboard.
Build, don't buy #
Ramp's in-house agent now raises 75% of its merged pull requests
Ramp's internal coding agent Inspect now raises 75% of the company's merged pull requests, per a Pragmatic Engineer deep dive built on interviews with Ramp's CTO and Inspect's founding engineers. It runs in cloud sandboxes with access to internal data sources and verifies frontend and backend changes on the remote machine. Engineers can use any tool they want; most choose Inspect over Claude Code or Codex. Block, Stripe, and Shopify run in-house agents too. Build-versus-buy just got a scoreboard.
Also worth your time #
• The End of Programming — Paul Dix argues manual code-writing and review are ending, built mainly on May 2026's Bun-to-Rust rewrite (64 parallel Claude Fable 5 agents, 11 days, ~$165K in API cost) as evidence for a much bigger claim. Real HN currency (68 pts) and a live counter-thread, but the load-bearing evidence is three months old — reader beware before treating it as today's news.
Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one.
The New Way is human-curated — a person picks every story. The summaries are written with AI (Claude) and reviewed before we hit send.