← Back to TIL
Second Gear / July 31, 2026

Slow Vehicles Make Practical Robots

Aggregation

World

1 critical / 0 happenings

Aggregation

Tech

1 critical / 1 happenings

Ideation

Ideation

Idea 1 AI Code Acceptance Lab

Sell enterprises an independent acceptance package for AI-generated code before it merges into regulated or contract-sensitive systems. The primitive is not another reviewer bot; it is a signed evidence file that says which agent changed what, what spec it was supposed to satisfy, which checks were independent of the generating agent, and which risks a buyer, auditor, insurer, or acquirer can rely on.

Source Signals

Why Now

Coding agents are cheap enough that code volume can rise faster than senior review capacity. At the same time, enterprises, acquirers, cyber insurers, and procurement teams are starting to care about how AI-written code was governed, not just whether the demo works.

First Wedge

A GitHub/GitLab check for fintech, healthcare, and government-vendor repos that produces an AI-change acceptance dossier on every agent-authored pull request: provenance, spec trace, independent tests, dependency diff, security triage, reviewer separation, and a human signoff trail.

Commercial Model

Security, platform engineering, or compliance teams pay per repo or per accepted agent-authored change. The first budget comes from release risk, SOC2/ISO controls, vendor security questionnaires, and M&A technical diligence where undocumented AI code becomes a deal risk.

Defensibility

The moat is the growing corpus of accepted and rejected AI code changes mapped to real production incidents, audit outcomes, and regulator/customer objections. Incumbent code scanners can add AI language, but they do not own the neutral acceptance record across agents, repos, insurers, and buyers.

Technical Risk

The hard part is proving useful independence. The system has to detect when the reviewing model shares failure modes with the generating model, decide which claims need deterministic tests or formal specs, and produce evidence that auditors will trust without slowing every merge to a crawl.

Market Expansion

Start with AI-authored code in regulated SaaS repos, then expand to vendor-security evidence, software warranties, cyber-insurance underwriting, open-source maintainer gates, and acquisition diligence for AI-heavy engineering teams.

Self-Critique

This could collapse into a nicer CI dashboard if buyers do not yet demand independent evidence. It also has to avoid competing head-on with every AppSec platform; the wedge must stay on acceptance and liability around agent-authored changes.

Next Experiment

In two weeks, run 30 recent AI-authored pull requests from three willing teams through a manual acceptance dossier. Measure which evidence security reviewers actually use, what blocks merge, and whether a compliance lead would pay for the signed packet.

Idea 2 Low-Speed Robot Operations Bureau

Build the operations and compliance layer for neighborhood robots that look like vehicles before they look like humanoids. The first product is a control room and evidence ledger for low-speed EVs, golf carts, campus shuttles, delivery carts, and mobile service robots that use remote operators for parking, dispatch, exception handling, and early autonomy.

Source Signals

Why Now

Physical AI funding is crowded around robot brains and humanoids. The contrarian wedge is that the first mass-market robots may be slow, wheeled, geofenced, and boring enough for remote operators and local permits, but still need professional-grade operations before insurers and municipalities tolerate them.

First Wedge

A managed teleoperation and incident-compliance stack for private communities, resorts, campuses, airports, and LSV fleets: operator staffing, live takeover, route geofences, local rule packs, video/event retention, near-miss logs, and insurer-ready safety reports.

Commercial Model

Fleet owners, robot startups, campuses, and property operators pay a monthly platform fee plus per-vehicle remote-ops minutes. Budget exists because one serious incident, denied insurance policy, or local ban can kill a deployment.

Defensibility

Each deployment produces an operating-domain library: risky intersections, radio dead zones, handoff patterns, local rules, operator interventions, and insurer objections. That dataset compounds across vehicle types, while a single robot OEM is biased toward its own hardware and cannot easily become the neutral bureau for cities and insurers.

Technical Risk

Latency is not the only hard problem. The product has to classify when autonomy, teleoperation, or human dispatch is legally and physically safe; preserve useful evidence without drowning operators; and keep handoffs reliable in messy local connectivity.

Market Expansion

Start with low-speed vehicles in controlled private or semi-private domains, then expand to sidewalk delivery, mobile security patrols, warehouse-yard robots, outdoor maintenance machines, and municipal pilot certification.

Self-Critique

The market could be too early if fleets remain tiny through 2027, and OEMs may try to run remote ops themselves. The company only works if it owns the cross-OEM trust layer: insurers, local authorities, trained operators, and incident evidence.

Next Experiment

Interview 10 LSV or campus-mobility operators, 5 robot OEMs, and 3 specialty insurers. Offer a manual safety packet for one pilot route: geofence, connectivity map, takeover protocol, incident form, and monthly risk report. Sell it before building the full console.