Editorial intro
Two motifs run through today’s picks: systems that solve constrained real‑world problems (elevators, NICs, deployable agents) and the boundary between flashy demos and production realities (AI prototypes, streaming huge models). Short reads first, then a longer technical unpack of an attention‑grabbing engineering proof.
In Brief
Elevators (Why your lobby button sometimes feels ignored)
Why this matters now: Elevator scheduling choices — from SCAN algorithms to modern RSR scoring — materially change wait times for millions of office workers and should influence building controls and retrofit decisions today.
Algorithms behind vertical transport are quietly sophisticated. The write‑up on elevator strategies walks from classic disk‑scheduler analogues like SCAN/LOOK to modern, multi‑car coordination where a central scheduler scores cars by ETA, load, and anti‑bunching penalties. Destination Dispatch systems — where you type your target floor before boarding — are popular, but the piece highlights a counterintuitive result: in many typical office flows they can worsen wait times because they pin down flexibility the scheduler needs. As the author puts it,
"The elevator did hear you, it just has a lot to think about."
Practical takeaway: there’s no one best algorithm — building traffic patterns (lobby-to-floors morning rush vs. inter-floor midday) determine which approach wins. Engineers and facilities teams should test with realistic traffic traces rather than defaulting to the newest kiosk.
Source: according to the original post.
qm — Multiplayer agent harness for work
Why this matters now: Y Combinator’s open‑source QM gives companies an auditable, multi‑tenant framework for placing agents under per‑user scopes and admin control — a pragmatic accelerant for internal agent adoption.
QM reframes agents from personal assistants into organizational collaborators: employees and rooms each get scoped memory, files, keychains and a "durable sandbox." It’s built to run on Slack and a web UI, supports cron/webhook triggers and pluggable models (Pi, OpenCode, Codex, etc.), and ships with three security postures: Strict, Auto, Dangerous. The repo is useful because it codifies decisions teams otherwise reinvent: per‑principal permissions, auditable actions under user credentials, and deployability in an org’s cloud account.
Community reaction is split. Some welcome the sane defaults for per‑person scopes; others mocked quirky shipped skills or questioned whether QM duplicates existing tools. Still: for teams experimenting with agents at scale, QM is a practical starting point — but guardrails and clear policies remain essential.
Source: see the GitHub release.
AI doesn't generate working products, that's still your job
Why this matters now: If your team is buying into LLM-driven rapid prototyping, remember: production readiness (security, observability, durable data models) still requires experienced engineering judgment.
This essay argues a simple, crucial point: AI accelerates reaching a working prototype but doesn’t shorten the distance to production‑grade software. Generated code often lacks resilient error handling, observability, strong data models and the architectural judgment to survive real traffic or attacks. As the author nails it,
"The prototype is the easy part."
Managers should treat LLMs as force multipliers for engineers, not replacements. If your hiring or training bets shift toward "prompt engineers," you risk accumulating brittle systems that look fine until they fail under load.
Source: according to the original post.
Deep Dive
Run Kimi K3 using 29 GB of RAM at 0.50 tok/s
Why this matters now: The WASTE engine demonstrates you can stream a full 2.78‑trillion‑parameter MoE model from NVMe and run it on a consumer machine — a technical milestone with real privacy and cost implications for people running occasional, sensitive workloads.
WASTE (Weight‑Aware Streaming Tensor Engine) is an embeddable C inference engine that keeps only the dense trunk and a bounded cache of sparsely‑activated experts in RAM, streaming the rest from NVMe on demand. The key idea is to exploit sparse activation in Mixture‑of‑Experts (MoE) models: only a small subset of experts fire per token, so you don't have to load the entire weight set into memory. This is not model compression — it's on‑demand weight loading.
On a well‑provisioned laptop (64 GB RAM, NVMe), a converted 982 GiB weight container ran at roughly 0.45–0.62 tokens/second. The author emphasizes careful engineering: read‑ahead, overlapping disk IO with compute, and a tuned expert cache policy. The project’s mantra is blunt and memorable:
"Every token answered by a cloud service is paid for twice: once on the invoice, and once in the electricity of a datacenter..."
That framing drives the motivation: for occasional, privacy‑sensitive queries you might prefer local answers even at half‑a‑token per second, avoiding both cloud cost and external data exposure. WASTE’s reproducible measurements and an emphasis on documenting what failed as well as what worked make the repo unusually useful for people who want to experiment rather than just applaud a headline.
But the tradeoffs are stark. Throughput is low enough that interactive or high‑volume usage is impractical; SSD wear and high I/O are real concerns, and commenters estimated power and durability costs per million tokens that can outstrip rented GPU inference for heavy workloads. Some in the Hacker News thread argued this is primarily a research milestone; others suggested it’s already meaningful for hobbyists, researchers, or legal/health queries where privacy trumps latency.
If you’re thinking of trying WASTE: treat it as an engineering exercise, not a turnkey replacement for hosted inference. Expect long warm‑up times, carefully manage your SSD endurance expectations, and consider hybrid approaches — run smaller local models for most queries and WASTE for a limited, high‑sensitivity subset.
Source: see the WASTE repo and measurements.
Closing Thought
Systems — whether elevators, workplace agents, or model runtimes — win or lose in the details: traffic patterns, security postures, cache policies, and the real costs of latency and reliability. Today’s standout across these stories is a reminder that clever engineering can stretch constraints, but meaningful gains come from measuring the right thing for the right user.