01 / AI operating systems
The company brain
Designing the context and tool layer that makes institutional knowledge usable by people and agents.
Read the thinkingAI, in practice
AI moves fast. Useful work→ still takes judgement.
The models will keep changing. The harder question is what to trust, where to use it and how to keep people in control.
A compressed history of AI
Seventy years
in one ascent.
Progress was never a straight line. The durable pattern is a climb from hand-written rules toward systems that learn, reason and act.
- 1950
The test
Alan Turing replaces a philosophical question with a practical one: can a machine converse indistinguishably from a person?
Turing, Mind ↗ - 1956
The field gets a name
The Dartmouth workshop coins artificial intelligence and sets out an audacious research programme.
Dartmouth proposal ↗ - 2012
Deep learning lands
AlexNet changes computer vision and proves what large neural networks, data and GPUs can do together.
NeurIPS paper ↗ - 2017
Attention scales
The transformer architecture makes it possible to train general language models at unprecedented scale.
Attention Is All You Need ↗ - 2022
AI becomes a product
ChatGPT puts a capable general interface in front of the public and changes expectations overnight.
OpenAI ↗ - NOW
Models begin to act
Reasoning, tools and agent loops move AI from answering questions toward completing work.
The current frontier ↗
Four levels of applied AI
Capability rises
with the harness.
Most organisations are using Level 1 and expecting Level 4 outcomes. The model matters. What surrounds it determines the result.
- 01The chat window
Prompt engineering
Where everyone starts: AI that answers. Simple, safe and built for the masses. How you ask changes what you get back.
- 02The terminal
~/work $ claude
Welcome back.
› read the contract, flag risky clauses
● reading contract.pdf … 3 clauses flagged
Context engineering
AI that acts under your direction: reading files, running commands and doing instead of describing. What the model sees decides what it can do.
- 03The Harness
The whole car is the harness.
Harness engineering
The model is the engine: raw power that goes nowhere on its own. Context, memory, tools, guardrails and observability turn it into a car you can drive.
- 04Effective automation
Frontier evolutions
The frontier is effective automation: systems that select methods, plan, act, validate and recover around an objective while human judgement stays at the consequential gates.
Executive Operations Audit
A focused, independent look at one critical operation.
Bring the operation that feels harder than it should.
- 01
Map
A clear picture of how the work actually moves.
- 02
Evidence
The facts behind what is slowing it down.
- 03
Recommendation
A practical next move, including when not to use AI.
- 04
Readout
A concise decision pack and private walkthrough.
The Audit is independent of any build. Sometimes the right answer is to leave AI out.
Selected focus
Work at the edge.
Built for the centre.
02 / Harness engineering
AI that can do the work
Moving past prompts into reliable workflows, tooling and agentic systems.
See the field note03 / Organisational enablement
Champions, not spectators
Equipping the people who turn isolated experimentation into an organisational reflex.
View the approachLatest writing
Read the
latest analysis.
The three most recent essays - dated, answer-first readings of the current model cycle and the regulation deadlines that decide what ships in Europe.
01 / AI security≈ 6 min
An AI found a WordPress zero-day for $25. Read the fine print.
A researcher used GPT-5.6 to build a working WordPress Core exploit for about $25 - a bug brokers value in the hundreds of thousands. What is verified, what is his own account, and what actually changed for defenders.
Read the essay02 / AI security≈ 6 min
The closing cyber gap: open-weight AI is four to seven months behind the frontier
The UK's AI Security Institute has, for the first time, measured how far downloadable open-weight models trail the closed frontier on offensive-cyber tasks: four to seven months, down from six to ten a year ago. Why a measurable, shrinking gap changes what defenders should assume.
Read the essay03 / Governance≈ 6 min
WAICO: China launches a global AI governance body - what it can and can't do
Twenty-nine states founded a China-led AI cooperation body in Shanghai on 16 July - without the US, EU, UK, Japan or South Korea. What a governance bloc with no binding power actually changes.
Read the essayAbout the author
The operator
behind the ideas.
Dominic J. Arrow builds AI systems from Amsterdam, informed by a career inside the messy reality of delivery, stakeholders and operational change.
14+ yearsinside delivery, systems and operational change
- 08 ●
Sep 2023 – Present
Manager, Systems
CSMART Almere · Maritime
- 07 ↗
Mar 2021 – Jan 2023
Technical Program Manager
Xero · FinTech
- 06 ↗
Sep 2019 – Mar 2021
Senior Business Analyst
Toyota NZ · Automotive
- 05 ↗
Feb 2019 – Sep 2019
Business Analyst
ezyVet · Software Development
- 04 ↗
Jan 2018 – Jan 2019
Strategic Program Manager
UCG NZ · Infrastructure & Build Construction
- 03 ↗
Sep 2013 – Jan 2018
Program Delivery Manager
Chorus NZ · Telecommunications
- 02 ↗
Jan 2012 – Sep 2013
Project Manager
Vocus Group · Telecommunications
- 01 ↗
2010 – 2011
Computer Science
Left after first year, self-taught since
Executive Operations Audit