Journal / research arguments

What would change the view?

Four distinct arguments about what an early-stage AI business needs to prove.

Memory & context · 4 min read

When does memory become a standalone business?

Memory earns an independent budget when it keeps changing knowledge usable across applications and measurably improves work beyond simpler alternatives. Persistence alone is an insufficient investment thesis.

Read article ↗

Published 2026-11-04

Reliable deployment · 4 min read

The cost of a successfully completed AI task.

The useful unit is the accepted business outcome. Retries, external tools, review and failed work belong in its cost, even when the model call looks cheap.

Read article ↗

Published 2026-11-04

Business workflows · 4 min read

What a paid AI pilot does and does not prove.

Payment is evidence of a bounded commercial decision. Repeat work, renewal terms, transferable implementation and delivery costs determine how far that evidence travels.

Read article ↗

Published 2026-11-04

Business workflows · 4 min read

Pre-seed judgment before the metrics arrive.

Before operating metrics exist, investment judgment can still be disciplined: separate founder insight, feasibility and buying behavior, then choose tests that could change the view.

Read article ↗

Published 2026-11-04

Next on the research desk.

Planned fortnightly notes. These dates are a forward editorial calendar.

Planned dateResearch questionStatus
2026-11-18Deleting memory without leaving stale summariesPlanned
2026-12-02When an evaluation score fails to predict a production incidentPlanned
2026-12-16Who owns the integration budget in a vertical AI sale?Planned
2026-12-30How to compare regional distribution without inventing a local moatPlanned
2027-01-13When an agent remembers a fact it can no longer accessPlanned
2027-01-27Switching models without losing the acceptance standardPlanned
2027-02-10Where implementation services hide in AI marginsPlanned
2027-02-24A monitoring alert is not a deployment decisionPlanned
2027-03-10The second application test for shared contextPlanned
2027-03-24What a customer reference can actually establishPlanned
2027-04-07Open-source adoption and the route to a buyerPlanned
2027-04-21How to document a Pass without confusing missing evidence with failurePlanned
Historical DeepTech notes →