In Brief

Semble – Code search for agents that uses 98% fewer tokens than grep

Why this matters now: Semble (the code‑search tool) promises big token and latency savings for developer agents, potentially cutting costs and speeding agent loops in real projects today.

Semble is an open‑source index and retrieval server tuned for AI agents that claims to “return only the relevant code snippets, without grepping or reading full files” and to use roughly 98% fewer tokens than a grep‑then‑read approach, according to the project page. It fuses tree‑sitter chunking, static embeddings and BM25 signals to produce tiny, precise results in milliseconds, all on CPU.

“returns only the relevant code snippets, without grepping or reading full files”

The Hacker News thread was upbeat but cautious: token savings look real, but several commenters warned the real test is whether models will trust those snippets and stop rechecking context. Semble is a useful building block — especially for teams paying heavy token bills — but it isn’t a magic fix for end‑to‑end agent reliability.

Two EA-18 fighter jets collide at Mountain Home airshow, pilots ejected safely

Why this matters now: The Mountain Home incident (two EA‑18G Growlers colliding) is a rare, high‑visibility safety event that will prompt immediate investigation of airshow procedures and aircraft choices.

Two U.S. Navy EA‑18G Growlers collided during the Gunfighter Skies airshow at Mountain Home AFB; all four crew members ejected and survived, and organizers canceled the show, per local reporting. Video circulating online shows the jets entangling at low altitude, the crews ejecting, and a fireball on impact.

“Everyone is safe and I think that’s the most important thing,” organizers said in the immediate aftermath.

Beyond the obvious relief that no one on the ground was hurt, readers on technical forums latched onto the surprising success of the ejections — many noted how uncommon survivable, low‑altitude, entangled ejections are — and raised questions about whether two‑seat electronic‑attack variants are an appropriate display platform.

Ask an Astronaut: 333 hours of Q&A footage with astronauts

Why this matters now: Ask an Astronaut (the searchable ISS Q&A archive) turns hundreds of hours of astronaut answers into instant clips educators and curious listeners can use right away.

The creator stitched roughly 333 hours of astronaut Q&A into a searchable site that uses transcripts, phoneme alignment and embeddings so you can type a question and jump to the exact moment it’s answered, according to the project site.

“I wonder if all the questions kids could ask of astronauts have already been asked,” the developer said as the project’s origin story.

It’s a tidy example of archival media becoming instantly useful: teachers, podcasters, and parents can pull short, authoritative clips without combing long recordings. Users asked for improved speaker labels, multilingual support and an offline package — sensible next steps if the project grows.

Deep Dive

I turned a $80 RK3562 Android tablet into a Debian Linux workstation

Why this matters now: The RK3562 Debian image (for the Doogee U10 tablet) shows you can run full Debian 12 from an SD card on cheap Android hardware — no bootloader unlock and no changes to internal storage.

A one‑person effort posted on GitHub documents turning an $80 Doogee U10 Android tablet (Rockchip RK3562) into a Debian 12 Bookworm workstation bootable from an SD card. The repo bundles reverse‑engineered boot scripts, kernels, and userland packaging so you can remove the SD card and return to stock Android with no changes to the device’s internal storage, according to the project README. Hardware support is impressively broad for a hobbyist project: touchscreen, Wi‑Fi, Bluetooth, audio, sensors and even the Rockchip NPU for running quantized LLMs locally.

The practical charm here is twofold. First, it’s a privacy‑friendly, cheap path to getting a real Linux desktop on a tablet that ordinarily ships with locked Android tooling. Second, the author includes tools to run on‑device inference — benchmarks show a small model (Qwen3‑0.6B) benefiting from the NPU. That opens an accessible route for hobbyists to test local LLMs without cloud costs. Community reaction captured a hopeful tension: many readers posted variations of “exactly what I want from the iPad Pro,” while others pushed back on realism, asking whether 4 GB of RAM and partial 3D acceleration are enough for day‑to‑day work.

This isn’t a polished consumer product; the project is explicit about what remains partial (color tuning for cameras, battery gauge quirks, incomplete 3D). But it’s a practical proof‑of‑concept that low‑cost Android hardware can be repurposed into usable Linux machines without voiding warranties or wiping the original OS. For tinkerers, educators on a budget, or anyone interested in on‑device models, this repo is a valuable, runnable blueprint — and it nudges the conversation about device freedom forward in a concrete way.

Jank now has its own custom IR

Why this matters now: Jank’s custom IR (a Clojure dialect runtime) shows how a language‑specific SSA/CFG IR plus runtime tweaks can beat the JVM on tight benchmarks, illustrating practical gains from designing language‑aware compiler internals.

The author of jank (a Clojure dialect) replaced a direct lowering to LLVM with a higher‑level SSA/CFG intermediate representation that preserves Clojure semantics like var‑derefs and dynamic calls, then generated C++ and applied targeted optimizations. The result was dramatic: a simple recursive Fibonacci benchmark dropped from roughly 5,522 ms to 114 ms, outperforming the JVM baseline cited in the post. The write‑up walks through concrete wins — pointer‑tagging integers, removing unnecessary boolean boxing, fixing a nil indirection, metadata‑driven inlining and forcing Clang to inline hot arithmetic — and explains why LLVM alone wasn’t the right abstraction for these language semantics, according to the announcement.

This is valuable as a case study, not a general promise. The author’s choices trade some dynamic flexibility for aggressive native performance. Hacker News readers pointed out familiar tradeoffs: language‑level optimizations can reclaim huge performance that a generic backend misses, but the JVM’s decades of adaptive JIT work remain a high bar for general workloads. Still, the post is instructive: many of the biggest wins came from small, obvious runtime changes (nil handling, tagging) rather than exotic compiler tricks. For language designers and runtime engineers, jank’s story is a reminder that a domain‑specific IR can expose optimization opportunities that a one‑size‑fits‑all toolchain can’t.

Closing Thought

A theme today: thoughtful engineering beats hype. Whether it’s squeezing Debian onto an $80 tablet, trimming tokens for agentic code search, or building a language‑aware IR, the biggest wins come from matching tooling to the problem. Small, well‑targeted changes often unlock more practical value than sweeping promises.

Sources