- 01NEW MODEL·github.com · openai/gpt-oss · 21m ago
openai/gpt-oss-120b — first openly licensed reasoning model in 18 months
Apache 2.0, 117B params, native tool use, 131k context. Outperforms o3-mini on AIME 2024; underperforms on SWE-bench Verified. Integrates with vLLM 0.7+ — first-party chat template in the repo.
- 02RELEASE·nextjs.org/blog/next-16-1 · 1h ago
Next.js 16.1 ships Turbopack stable for App Router
Production builds cold-start drops 38% on a representative Next + Postgres app. webpack-only escape hatch available via `--webpack` for the next two minor versions. Migration guide walks through the three known server-component cache regressions.
- 03BREAKAGE·github.com · huggingface/transformers · 2h ago
transformers 5.0rc1: `AutoTokenizer.from_pretrained` rejects bare strings
Now requires an explicit `revision=` or `cache_dir=`. Half of the eval harness on HuggingFace fails. Pin to 4.55.x for the next week or two.
- 04EVAL·swebench.com/verified · 3h ago
SWE-bench Verified leaderboard shuffle — Sonnet 4.5 retires three entries
Five new submissions above 70%. Reproducibility mixed — three rows stop accepting new runs pending a container rebuild. Notebook in the supplementary ZIP.
- 05NOTE·blog · foo.bar · 5h ago
Why your RAG eval lies to you (a 4-paragraph rant from a maintainer)
The blog post everyone is sharing in Discord today. TL;DR — most public RAG benchmarks sample from a fixed query distribution that does not look like production traffic. Worth reading before you ship the next retrieval change.
ISSUE 042 · MON 23 JUL
Daily AI drops
for builders.
Five things that actually shipped today — model checkpoints, dev-tool launches, framework releases, eval benchmarks, and the occasional broken dependency — with enough context to decide by tomorrow morning whether to wire anything new into your stack.
- Five drops a day
- Every item sourced from primary docs
- 3 minutes to read
Why this exists
We don't chase the news cycle.
We chase release notes.
Every AI newsletter on the internet was written for the executive in the room. Patchwire is written for the engineer making an implementation decision at standup tomorrow — what changed, what works, what regressed, and where to start.
What we cover
- New model checkpoints with reproducible benchmarks
- Dev-tool launches, SDK release notes, breakage with workarounds
- Maintainer-direct posts — blog, changelog, mailing list
What we don't
- ·Funding rounds and executive interviews
- ·Opinion essays re-litigating the same scaling-law debate
- ·Press-release rehashes of features already shipped last week
- ·Vague “AI is changing everything” think pieces
- ·Cold-call partnership announcements from platforms you already use
Sample issue
What an actual Friday looks like.
A real issue — five drops, one optional eighth-paragraph rant. Each row links to a primary source we read in full before publication.
Sourcing
We read it before we send it.
Every drop traces back to a primary source — release notes, a GitHub commit linked from a maintainer, a benchmark paper we ran the eval on, or a direct post from the team that built the thing. If we can't read the source, it doesn't go in. Press coverage gets you halfway; release notes get you the rest of the way.
FAQ
What people actually ask.
Subscribe by email — there is no web signup. Reply to any issue and we read it.
Subscribe
Tomorrow's issue is one email away.
Email is the only channel — no apps, no accounts, no algorithm. Tap subscribe and we'll send a one-click confirmation. We don't start the daily issue until you tap it.
Check your inbox to confirm — one click.
Tap subscribe — we will drop you on the list and send a confirmation to your inbox in seconds.
Prefer email? Write to patchwire-2@polsia.app and we'll add you manually.