Field Notes | Northtek

Field notes

Positions we are willing to be wrong about in public.

Most AI consulting content restates the same four paragraphs about efficiency. These are arguments, drawn from research we read directly and deployments we actually ran, with the primary sources cited so you can check them and disagree.

Written by Kristian Baer, Northtek · Anchorage, Alaska · Updated 2026-08-21

Applied cognitive architecture

Why AI pilots fail

The failure has a name in the research literature. Most pilots give the model one of the three memories a business actually runs on.

3 sources cited →

Reading the frontier

Small models, local data

NVIDIA Research argues most agent work does not need a frontier model. For an Alaska business, that changes what is possible on your own hardware.

3 sources cited →

Brand in the AI era

Founder brand

Buyers decide before they contact you, and AI engines resolve people more readily than logos. Why the founder now carries the company.

3 sources cited →

Original research · in the field now

The Alaska AI Operations Index

Every figure on AI adoption in this state is a national study with Alaska pasted on top. We are running the real count - what got deployed, what got abandoned, and what stopped it - by industry, with the methodology and raw data published. Three minutes to add your organization.

Take the survey →

What we built, in the open

Arguments are cheap. Here is the work underneath them.

Anyone can have a position on AI. What makes a position worth reading is whether the person holding it has had to make the thing work. Each of these is public, with its commit history.

  • GENOME

    A memory server implementing persistent long-term storage for agents, benchmarked against a commercial competitor on public datasets. This is what the memory argument looks like as running code.

  • Kryos

    A programming language and compiler built so an agent’s actions are auditable line by line. Writing one is how you learn which parts of the agent literature survive contact with production.

  • FACTGATE

    A verification gate for model output, with its false-accept rate published including the cases it misses.

We would rather be argued with about these than agreed with about generalities.

Disagree with one of these.

The fastest way to find out whether we know what we are talking about is to argue with us about it for an hour. No charge, and no pitch unless you ask for one.

Book a Working Session Read our source code first

Anchorage, Alaska · info@northtek.io · (907) 903-4353