/ what friday looks like
browse archive →A quiet week on the model-release side, a noisy one on the agent-tooling side. Two papers worth a coffee, one open-source release that quietly does what closed labs charge for, and a sponsor we genuinely use. Let's go.
— Jarek, editor
The 1M context tier is the headline, but the more interesting line in the post is the new prompt-cache TTL — caches now stick for an hour instead of five minutes, which makes long-running agent loops noticeably cheaper. Benchmarks on SWE-bench and Terminal-Bench are competitive, but the practical win is for anyone who fits a whole codebase into a single prompt.
Sahar collected traces from a year of production agent runs and found ~70% of failures come from three places: tool-schema drift, silent truncation in retrievers, and prompt-cache invalidation no one notices. The post-mortems are the kind of thing you'd normally only see in a postmortem channel — generously shared.
Skips the academic eval suites and walks through a budget-friendly setup using Ragas, a handful of synthetic-but-realistic questions, and a small reviewer model. The honest framing is the best part: "if your eval set was hand-picked by the same person who built the retriever, you don't have an eval set."
30-minute teardown of the new WebRTC-based realtime API: how interruption events are sequenced, why the audio buffer is now server-authoritative, and where you still need client-side VAD. Useful even if you're on a different vendor.
A surprise mid-week drop. Apache 2.0, multilingual out of the box (incl. PL/CZ/JP at competitive levels), and benchmarks within striking distance of Llama 4 70B at a fraction of the VRAM. Runs at usable speeds on a single H100.
Long-running agents can now resume from any step without re-running upstream nodes. The DX is closer to a workflow engine than a chat loop, which is probably where this whole space is heading.
An open-source CLI that hooks into your Zotero library and your Arxiv RSS, then ranks each week's drops by how close they are to what you've read recently. Quietly excellent at surfacing the one paper you'd otherwise scroll past.
The annual "draw a clock" benchmark resurfaces. Comforting in the way only AI Twitter can be.
/ how we pick
Each week we read so you don't have to. The funnel:
Arxiv, GitHub trending, vendor changelogs, niche Discord/Slack communities, 50+ other newsletters, conference talks.
Anything that looks more than a press release or a tweet thread gets opened, read, and triaged.
Papers skim-read for the abstract + method. Tools tried, not just bookmarked. Releases checked against the actual changelog.
Each with one sentence of editor's context. If a week's quiet, you get fewer. We don't pad to hit a number.
/ the deal
/ faq
Most AI newsletters either rehash Twitter or pad a list with everything that moved. We do the opposite: 8 stories per issue, max, each with one sentence of editor's context on why it matters. If a week is quiet, the email is shorter.
Those are aggregators — broad, daily, optimized for not missing anything. We're weekly, opinionated, and built for people who ship: a developer's perspective on what's actually production-ready vs. demo-ware. Subscribe to both if you want — we won't be offended.
Yes — the archive opens with issue #001. Until then, the preview on this page is a faithful mock of the format and editorial voice.
Yes. Costs are covered by one or two clearly-labelled sponsor slots per issue. Never disguised as content. Never tracked across the web.
One click — either from the link at the bottom of every issue, or via the one-click unsubscribe button Gmail/Apple Mail now show natively.
Something not here? Email hi@aiweekradar.com.