Dev · Friday, 21 August 2026 · Issue 83 · 13 stories

Go 1.27 SIMD, TypeScript 7 self-hosts

A 7.5x speedup on encryption without leaving portable Go code: that's what the new simd packages deliver. And TypeScript 7 just passed its toughest test: compiling itself.

Languages & Frameworks

Go 1.27 ships generic methods and experimental portable SIMD

Go 1.27 lands August 2026 with a pair of long-requested language upgrades: generic methods can finally declare their own type parameters, and struct literals accept any valid field selector. The headline is the first experimental "simd" and "simd/archsimd" packages, bringing portable SIMD to the standard library. The toolchain also parses GCC-style response files and turns stdversion vet on by default. All of it stays behind the Go 1 compatibility promise.

go.dev ↗15 min read at source

Go 1.27's portable simd packages: 7.5x faster ChaCha20 encryption

Go's stdlib ChaCha20 lacks SIMD on amd64, so this writer reached for the new portable `simd` and `simd/archsimd` packages landing in 1.27. The payoff: a 7.5x speedup on his ChaCha20-BLAKE3 scheme, using portable code instead of hand-tuned per-architecture assembly. Worth a read if you've ever envied Rust's auto-vectorization from Go.

kerkour.com ↗3 min read at source

TypeScript's own compiler now builds on the TypeScript 7 engine

The TypeScript repo now compiles itself with TypeScript 7. That merged PR is the biggest dogfooding yet for the Go-based native compiler, running the language's entire codebase through the very rewrite it targets. It's a serious stability signal: Microsoft won't ship the Go compiler as ordinary tsc until the language can build its own code. Large projects eyeing the 8-12x speedup now have a reference migration.

github.com ↗3 min read at source

Rust 1.98 brings fast-math floats and itoa-grade int formatting

New algebraic f32/f64 methods let the compiler reorder and vectorize float math the way -ffast-math does in C, without the undefined behavior. Integer types gain a format_into method whose buffered output now benchmarks on par with the itoa dependency, giving teams a standard replacement. The release also fixes an undefined-behavior edge between ManuallyDrop and Box. All live on stable via rustup today.

blog.rust-lang.org ↗3 min read at source

Tools & Platforms

Zellij 0.45: run Zellij inside Zellij without the keybind fight

Zellij 0.45 ships first-class handling for running inside another Zellij session: zoom the nested instance or pass keybindings straight through on focus, ending the lock-mode dance. It also implements the Kitty graphics protocol alongside Sixel, so images in panes survive scroll, resize, and stacking. Rounding it out: command-by-command scrolling via OSC133, a redesigned default UI, and a mobile web client.

zellij.dev ↗5 min read at source

SpacetimeDB 2.0's taunting benchmarks don't survive an honest technical review

SpacetimeDB 2.0 landed with a meme-y launch mocking rivals and benchmarks too good to be true. Reading closely, those numbers are neither fair nor honest: the product occupies a genuinely different database niche, which makes it tempting to build rigged comparisons against incumbents. The review digs into what's real vs. what's theater.

strn.cat ↗14 min read at source

AI-Assisted Development

AirTag traces 1,000-book mystery order to Amazon's AI scanning hub

404 Media hid an AirTag inside an order of roughly 1,000 books bought anonymously via Biblio, then tracked it to Amazon's VGT3 facility in Las Vegas. Worker threads there confirm the site destructively scans large volumes of books. It is the strongest proof yet that price-insensitive bulk book orders, long suspected in the trade, quietly feed AI training pipelines. Sellers now have a concrete signal for spotting anonymous AI-data buyers.

simonwillison.net ↗1 min read at source

Engineering Practice

How id packed its whole game catalog onto Quake's $9.95 shareware CD

Fabien Sanglard forensically dissects Quake's 1996 shareware CD, built to do far more than hold the 22 MiB game: id encrypted its entire back catalog on the disc so gamers could call an 800 number and unlock any title. The scheme collapsed fast, crackers broke the encryption just 39 days after launch, leaving id stuck with 150,000 unsold CDs. A sharp postmortem on DRM that punished the exact customers it was courting.

fabiensanglard.net ↗6 min read at source

How Shopify took mobile E2E test stability from 50% to 98%

Shopify's mobile E2E suite got so flaky it was blocking more good PRs than bad ones, so they pulled it from pull-request checks entirely. Rather than keep patching Appium, they wrapped it in an opinionated, builder-style API that forces an assertion on every step and uses computer vision to find elements the way a user does, not by crawling the view hierarchy. Stability climbed from 50% to 98%. The lesson: stop bailing out a framework that keeps manufacturing the same failures, and rebuild the layer above it.

shopify.engineering ↗7 min read at source

Inside TigerBeetle: zero-allocation design that kills tail latency

A deep-dive deconstructs why TigerBeetle, the Zig-written financial ledger, treats the OS kernel and memory fragmentation as the real bottleneck, not the query planner. It statically pre-allocates every byte at startup, locks the allocator, and runs a single-threaded loop with direct I/O and Viewstamped Replication. The payoff: hundreds of thousands of transactions per second with predictable sub-millisecond tails. Worth a read for anyone building high-throughput systems.

ixuvo.com ↗10 min read at source

PagedAttention loans OS virtual memory to LLM KV caches, doubling GPU serving

The KV cache, not the weights, is an LLM's scarcest resource, and naive serving reserves huge contiguous chunks per request, most of it idle. PagedAttention gives the cache OS-style virtual memory: fixed-size blocks shared across requests and reused as they fluctuate. Fragmentation vanishes, wasted gigabytes shrink roughly 64x, and one GPU serves two to four times more users.

thegustafson.com ↗15 min read at source

Kambhampati's position paper: stop calling chain-of-thought "thinking"

Subbarao Kambhampati makes the case that labeling LLM intermediate tokens as "reasoning" or "thinking" traces is not a harmless metaphor but a dangerous one. It misleads users into reading human-like cognition into stochastic token prediction, warps how we evaluate and trust models, and steers research in questionable directions. His call: treat those traces for what they are, not as a window into a mind.

arxiv.org ↗1 min read at source

GitHub's outage: your autoscaling policy is blind to the sidecar it scales

The GitHub outage traces partly to an autoscaling policy that watched the host service's load but never the Istio sidecar's concurrency limits, so the sidecar saturated and couldn't spin up. One system's saturation can hide while another metric looks healthy, as Slack proved back in 2021. Autoscaling is bespoke operational engineering per service, not a check-the-box feature, and it only surfaces under load testing everyone should be doing.

surfingcomplexity.blog ↗4 min read at source

You’re caught up.

That was all 13 stories in the 21 August issue.