LodeHQ
AI news for people who shipFree, by email

Product · Saturday, 22 August 2026 · Issue 82 · 14 stories

AI makes building cheap, so decisions now bottleneck

Execution is nearly free. That means the hardest part of your job just became choosing what to build, and the default way most teams decide is the least efficient one.

Product Management

AI Made Building Cheap, So Your Decision Process Is Now the Bottleneck

AI agents are making execution nearly free, which means choosing what to build is now the real bottleneck. David Pereira argues the default response, more meetings and wider consensus, is the least efficient one: talk breeds confidence, not evidence. The better pattern is decision-empowerment, handing product people an outcome to figure out, not a pre-scoped solution to approve. Match the process to the stakes, not to the crowd.

dpereira.substack.com ↗9 min read at source

Faster Code Isn't Faster Revenue: The Missing Ingredients Between Ship and Sell

AI coding is making software cheaper to build, which is convincing founders and CEOs that faster development means faster revenue. Mironov's shopping list says otherwise: code is the cheapest ingredient, while pricing, support, compliance, and go-to-market are what paying customers actually buy. He splits the product into three tiers, code, commercially deployable software, and revenue-ready product, and itemizes what's missing at each stage. Good ammunition before you promise the board a number.

English becomes the interface: AI built a dress its creator never could

Yana Welinder asked AI to operate CAD software she'd never learned, and it turned her sculpture-inspired vision into a dress that was "previously unmanufacturable." Tom Tunguz reads this as the end of software's expertise tax: English is becoming the universal API, translating intent into the grammar of accounts, orders, and splines. The consequence is an inversion of product design. PMs stop designing every screen and start building the harness that lets agents act safely, and documentation becomes the product.

tomtunguz.com ↗2 min read at source

Design & UX

AI Images Pass the Trust Test: No Penalty vs. Stock Photos

NN/g ran a controlled study with 77 users rating identical fictional company pages that swapped only the hero image. Unlabeled AI-generated photos scored no worse than real stock on trust and professionalism, and the tiny authenticity edge for real photos (0.4 points on a 7-point scale) was marginal. The implication: what an image depicts matters more than how it was made, so teams hesitant about AI imagery can stop worrying about a perception penalty.

nngroup.com ↗8 min read at source

Strategy & Growth

Tools & Launches

Vercel's Zero: the graph-first programming language built for AI agents

Vercel Labs just shipped Zero, an experimental, open-source language that inverts the agentic coding loop. Instead of agents writing text files and hoping the build works, they query and patch a semantic program graph; the compiler validates every change before it sticks. Humans only ask for outcomes and review readable .0 projections when useful. It's a bet that the compiler, not source text, should be the source of truth for AI-written code.

zerolang.ai ↗2 min read at source

Reverse-engineered HID protocol brings the Attack Shark X6 gaming mouse to Linux

A Go + React configurator decodes the Attack Shark X6's proprietary HID protocol, captured from the Windows app and validated against the dongle, unlocking 8 DPI stages (50-26,000), battery heartbeats, and per-device profiles on Linux. Writes 56-byte SET_REPORT configs over hidraw, gated by udev seat permissions instead of root. Beta, MIT licensed, build from source. A Windows-only peripheral now has a first-class Linux story.

github.com ↗2 min read at source

Reverse-engineered Bambu's 3MF color format so GLB keeps every color zone

Every other 3MF-to-mesh converter throws away the slicer's per-face paint data, since Bambu Studio's color encoding is undocumented. This browser-only, MIT-licensed tool is billed as the first to carry those color zones into GLB, then lets you recolor them live, preview true-size in AR, and export GLB/USDZ/OBJ or a re-tinted 3MF. Works with Bambu Studio, OrcaSlicer, and Snapmaker U1 files, 100% client-side with no uploads.

github.com ↗19 min read at source

A math course built a compiler that compiles itself

not-abc is a teaching compiler written entirely in itself: just 64-bit values, no static type checking, and a handful of built-ins, yet the whole toolchain is complete. It comes out of Michael Lehn's undergraduate HPC course, a live demo that a surprisingly small language can host a full compiler. Reading how the bootstrap works is a fast way to un-mystify compilers for anyone who skipped the CS courses.

github.com ↗10 min read at source

Antigravity agents now live inside VS Code, JetBrains, and Zed

Google's bet was that coding happens in an agent-first surface, but it just conceded you still live in your editor: Antigravity is now a lightweight extension for VS Code, Visual Studio, JetBrains, and Zed. All four share your agent conversations, customizations, and multi-agent context with the standalone app and CLI, so you can review diffs and debug without switching tools. It is the rare launch that removes a migration cost instead of adding one.

antigravity.google ↗3 min read at source

Static security gate for Postgres extensions, before they ever touch a database

Postgres extensions cross a brutal trust boundary: install scripts can run with elevated privileges, and native modules execute inside the server process. PgExtAssure scans an extension's full source tree before that decision, flags security-relevant patterns, and can fail CI at a chosen severity. It never builds, installs, or runs what it inspects, so the evidence is reproducible without touching a live database. No hosted account or telemetry, and public runs from independently controlled forks back the claim.

github.com ↗15 min read at source

MiniMax H3, a 33B video model, runs on 16GB Macs days after release

Days after MiniMax opened the weights, vpipe has the 33B H3 FL2VA/REF2VA models generating video and its soundtrack on a 16GB MacBook Air. It is a native-Metal pipeline with no third-party tensor runtime, and benchmarks 11m25s per 5-second clip, beating the reference h3.c implementation on identical M5 hardware. Weight streaming plus 4-bit prep squeeze a model class that previously required a workstation into a fanless laptop.

github.com ↗20 min read at source

A 3-day subagent refactor experiment birthed a durable AI-team runtime

Ad-hoc subagent delegation is fragile: a closed terminal or interrupted provider session can erase the only record of what happened. After running a three-day, subagent-driven codebase refactor, this developer built Oh My Subagents, a local runtime that commits child assignments, persists parent waits, and supervises every return so managers never poll in a loop. Delegation becomes recursive and checkpointed, with completed work preserved even after crashes. The repo ships docs and starter teams for Codex and Claude.

github.com ↗7 min read at source

Swat this browser fly: its brain is the real FlyWire connectome

daedalOS, the open source web desktop, now ships pet fruit flies driven by a live spiking simulation of 1,275 real FlyWire connectome neurons. Each fly has its own brain and senses: window edges are walkable terrain, the cursor looms, clicks swat, and a takeoff startles nearby flies. That turns a dataset usually browsed one neuron at a time into recognizable behavior you can poke, groom, and send to sleep in a browser tab.

github.com ↗4 min read at source

You’re caught up.

That was all 14 stories in the 22 August issue.