AI Express // Latest edition // nr. 04 // Sep 17

The Build: Rust on the GPU, a security-audit skill, and Factory's $5B round

Thursday build. This week the toolchain itself moved: NVIDIA shipped native Rust on the GPU, Cloudflare open-sourced the security-audit skill behind its vulnerability harness, and a solo builder taught a 4B model to out-plan Postgres. Here is what is worth your time.

For Builders

For Builders

A 4B model beat Postgres on join-heavy queries after agentic RL

Rohan Bansal post-trained a 4B Qwen distill to emit pg_hint_plan hints, then trained it with a custom GRPO variant against live Postgres timings. Result: a 44.7% latency reduction across 113 join-heavy JOB queries, up from a base model that could not produce a plan for 99 of them. The write-up is as much about measurement as modelling: he built a four-container rig to kill page-cache noise, then quantified how often the reward signal fools itself, with a no-op candidate scored as a win up to 40% of the time. If you post-train on a noisy metric, read the calibration sections first.

Open-Source Pick

Open-Source Pick

Cloudflare's security-audit-skill turns your coding agent into an auditor

Cloudflare open-sourced the skill that seeded its fleet-wide vulnerability harness. It runs a six-phase audit: reconnaissance into a coverage ledger, coverage-led hunting by isolated agents, adversarial candidate validation where the agent that checks a finding is never the one that found it, structured findings validated against a schema, independent record verification, and target-neutral reports. Two zero-dependency Node validators keep the ledgers honest. Around 9.5k stars, MIT licensed.

For subscribers only

Dev Tools

NVIDIA brings GPU kernels to Rust, on two tracks

locked
Dev Tools

OpenSpec puts a spec in front of your AI agent

locked
Startup Spotlight

Factory triples its valuation to $5B on autonomous coding agents

locked
Design & Trends

QuiverAI bets the editable SVG is AI design's real output

locked
From the Lab

Ternary LLMs waste bits on zeros, and BITCOS packs them tighter

locked
From the Lab

Google makes design direction an explicit choice for UI agents

locked
Try This

Fine-tune a 459M NER model for about $9

locked
Hardware & Edge

Serving a 35B MoE from SSD, and Edge0's preview model

locked

The full brief, the tools, the GitHub radar and the context stay subscriber-only. Subscribers always read first, in full.

Read the whole thing next time.

Ten stories, the tools, the trends, zero noise - three times a week, before anyone else.

Join Telegram

free forever · no spam · unsubscribe in one click.