Month 6 at the AI Lab: Spec Coding, Local Models, and AI in the Businesses Nobody Blogs About
Notes from the June 25th meetup — the end of "vibe coding," why fine-tuning fundamentals matter, and applied AI showing up in fencing and independent restaurants. Plus, Second Robotics moves in.
AI Builders Meetup: June 25th
The June meetup circled around three ideas that reflect where the field actually is right now.
1. Vibe Coding > Spec Coding
Walter Estrada — a major open source contributor (Grails) — rejected the "vibe coding" label outright and offered a replacement: spec coding. Write a tight specification first, then let the agent execute against it. Specs give the agent something to be held accountable to.
He also coined confabulation gates — application-layer guardrails that catch LLM hallucinations before they reach the user. A concrete architectural pattern where most people just wave vaguely at "hallucination prevention." He's building an ACT/SAT study aid inspired by his own son, and the confabulation gate is the reliability layer that would let a student actually trust the tutor.
Not a coincidence: Z.ai's GLM-5, released mid-June, subtitles its release "From Vibe Coding to Agentic Engineering." When a frontier lab prints that on the release notes, the phrase is real. The industry is quietly moving from "watch what happens if I prompt it" toward "specify what it must produce, and verify."
2. Knowing what's under the hood
Mouna brought the technical depth. They referenced GLM 5.2 as an open-weight pick, LoRA fine-tuning as routine, and sub-Chinchilla training — models trained below the ~20-tokens-per-parameter ratio DeepMind's Chinchilla paper set as the compute-optimal baseline. Most deployed models are still under-trained relative to their size.
Mouna had some memorable lines: Regarding good use of dev cycles - "Sure, you could have a room full of engineers to do calculations for you — or you could have them build a calculator?"
Regarding space data centers - "will solar flares cause bit flips?!"
3. AI is showing up in the businesses nobody blogs about
The two strongest applied AI stories in the room came from the least tech-shaped domains.
A local fencing company is running on the Entrepreneurial Operating System— and has now wired AI directly into it, using OpenClaw to hit their CRM and generate customer estimates. A real business, real quotes, an agent with actual system access. Not a proof of concept.
Greg Hard is building what he calls the Restaurant Operating System. Cameras on the floor, QuickBooks on the numbers, and decades of operating wisdom encoded as an AI a working owner can actually talk to. His thesis is that AI can save independent restaurants. Margins are brutal, turnover is brutal, and most owners run on instinct and exhaustion. Greg wants to give them a second brain.
Together with Walter's EdTech app, the pattern is hard to miss: the interesting AI right is the embedding into real businesses like fencing, restaurants and tutoring.
New Startup at the AI Lab: Second Robotics
Second Robotics has moved into the lab. Apollo Timbers — hardware engineer and Renaissance man (see photo) — is building robots that run local AI models for visual detection: plants, rocks, environmental sensors. The application in his sights is a lunar rover: a machine that can navigate and identify terrain without phoning home to a cloud API.
That rover already has a name: ATHENA, a lunar micro-rover built to carry a 1U CubeSat-class payload across the surface while conducting its own science along the way — imagery, light sensing, and radiation measurements. The shop motto sets the bar: "We do not fail the mission."
That's the kind of use case that keeps pushing on-device inference forward. When latency has to be human-fast and connectivity can't be assumed, the model has to fit on the device. Robotics has been one of the strongest forcing functions for smaller, faster, locally-runnable models — and it's now happening on our bench.
Apollo's latest painting — a satellite drifting through a cosmic nebula — is already on the wall. Part maker space, part gallery.
A New Space: The Focus Room
We added a new space to the AI Lab this month — a quiet, enclosed room for focused work or calls. The centerpiece is a solid wood desk, hand-refinished by Dan's mom. (see photo)