Back to first principles: what leaders need before they ask AI anything

Friends,

Your weekly AI briefing is here - designed to help you respond to AI, not react to the noise. No curveballs. No chaos. Just clarity.

Last week's edition offered an order of priorities for deciding what AI is for. This week asks what a leader needs to understand about their own organisation before asking AI to help with any of it.

📰 This was the week that was...

This was the week to know your own organisation first.

On 27 October, Anthropic published an update on its work with six drug-development partners. The result most useful to any leader with a pile of paperwork comes from one of them, Enveda, whose manufacturing chemists spent up to 75% of their time on documentation. In its own report, published the same day, Enveda describes a system built on Claude that checks a completed batch record against the company's own approved instructions. Human review with the system's comments in front of the reviewers was about nine times quicker. The comments are advice: Enveda's reviewers decide which ones need action. These are early, company-reported results from one clinical campaign, and they say nothing yet about how any medicine performs in patients. Our reading: a check like this is easier where an organisation has already set down in writing what "right" looks like.

On 27 October, the US General Services Administration announced an expanded agreement with Anthropic under which government buyers now pay 40% less per token of Claude usage, with no minimum orders. Anthropic's own post estimates the discounts will save government $1 billion. It is one negotiated agreement for one set of buyers, so it settles nothing about prices in general. If terms like these become more common, more organisations will be able to try capable AI, and clarity about what to ask it for will matter even more.

Let's get into it.

🔥 Urgent Priorities

✅ No fires to fight this week

✅ Worth doing now: before your next AI tool decision, ask whoever is proposing it which question it will answer for your organisation, and write the answer down.

Nothing here is pressing. A calm half hour on this is plenty.

🎯 Strategic Insight

Tension: This week's GSA agreement makes capable AI cheaper to try for one set of buyers. Whatever happens to prices elsewhere, every leader can answer one question now: what is this organisation asking AI to help with?

Optimistic insight: Hugo's reflection since his retreat is that the useful move for leaders is a return to first principles, and in practice that means good management. AI can spotlight problems and evaluate business cases, but it only does it well for a leader who understands what is happening in their organisation before asking AI.

A recent episode of The Mayvin Podcast, released on 20 October, Earthy Intelligence: the Odyssey in the Age of AI, describes The Odyssey as "a poem about practical wisdom, appetite, limits, and how hard it is to come home". Our own reading: practical wisdom is knowing how your organisation really works, appetite and limits are choosing what to ask AI for, and coming home is returning to first principles before any tool decision.

What's shifting: If capable AI becomes more widely available, the advantage may move to organisations that can say clearly what they want from it. Two early examples fit that idea: The first is this week's Enveda story, where the system checks records against instructions the company had already approved. The second is a preprint posted on 27 October, in which Saiful Khan and Afrah Farea report that an AI agent given the process steps behind a company's software services completed 38.7% of 81 multi-step tasks, against 10.3% for a baseline that searched text chunks. It is an early, author-reported result. Their research results remind us that first principles are key when making effective use of AI. When a question is unclear, AI can return confusion quickly and at volume. A leader who has done the thinking gets that same speed working for them.

Why this matters now: Rolling out AI tools asks a lot of leaders, first in choosing well and then in helping people make the tools part of daily work. A clear picture of how your organisation works supports both steps, and it costs only time.

Action: Write a note covering four things: what your organisation is for, how work really flows from request to result, your top three priorities, and what good looks like. Leave out anything confidential. Then ask an AI tool to point out gaps and contradictions in the note, and to test one live business case against it. Pay attention to the places where you had to stop and think. Those are the places to look at the organisation directly - in the real world - not in the AI world.

🤓 Geek Out

1️⃣ A world you can steer with plain sentences

Google DeepMind has announced Genie 4, a "world model": an AI that simulates what happens next in a scene, so you can move around inside it in real time. According to Google DeepMind, it generates sound and spoken dialogue alongside the picture, and you can change the scene by typing instructions, such as changing the weather or adding a person or object. Google suggests uses from training robots and AI agents to new interactive experiences. It is a research preview for trusted testers, with a waitlist for wider access, and the capability claims are Google DeepMind's own.

Why it matters: Safe, adjustable practice worlds could one day let a team rehearse a difficult site visit or a rare emergency before it happens for real. That is our extrapolation from Google's description, and it is early days.

👉 Action: Nothing to try yet. Note down one situation where your people would benefit from rehearsing in a safe copy of the real world, and keep the note for when access opens.

👉 Read Google DeepMind's announcement

2️⃣ A satellite-image AI model, shrunk by more than 90%

Researchers have shown that an AI model that sorts satellite images by land use can be cut down by more than 90% of its parameters (the internal settings that make up a model) and still keep strong accuracy. Their method lets a "teacher" model pass its knowledge to a smaller "student", then trims the student step by step. On two public collections of satellite images, accuracy stayed strong beyond that 90% cut. Lower power use on small devices is the authors' aim, and the results they report are about accuracy. It is a preprint, also published in the proceedings of a geoscience conference, IGARSS 2026.

Why it matters: Smaller models that keep strong accuracy may help analysis run closer to where the images are taken. For anyone who relies on satellite or aerial imagery, where the analysis runs is a question worth asking.

👉 Action: If you buy satellite or aerial image analysis, in farming, property, insurance, land management or logistics, ask your supplier whether the analysis must run in the cloud or could run on smaller equipment.

👉 Read the preprint

3️⃣ An AI answer that shows which evidence mattered most

A research team has proposed "Filtered Decision Logs": each AI answer arrives with a short reason, plus the supporting evidence ranked by how much it shaped the answer, without exposing the AI's raw step-by-step thinking. In a small study with 50 participants, 92% rated the ranked evidence as helpful for understanding and trusting the recommendation. When the researchers removed the top-ranked source, the answer changed more than when they removed a low-ranked one, which they take as a sign that the ranking tracks what the answer depended on. These are the authors' reported results in a preprint, and the study is small.

Why it matters: A decision you can trace is a decision you can question. Seeing which evidence counted most helps a leader accept an AI recommendation, correct it or ask again.

👉 Action: Next time an AI tool gives you a recommendation, ask it for a one-line reason and the three pieces of evidence that mattered most, then check the top one against your own records. This is a rough version of the idea, since a chatbot's own account of its reasoning works differently from the paper's method.

👉 Read the preprint

🎨 Weekend Playground

This weekend, try Stellarium, an open-source planetarium that shows the sky above you as it really looks. The clocks went back on Sunday 25 October, so evenings now get dark early.

Why this matters: Looking up is a free way to reset your sense of scale before Monday. Stellarium includes constellations from many cultures. The phone finds the stars, then you do the looking up.

👉 Mission:

• On any phone or computer, open Stellarium Web in your browser and allow it to use your location. On Android, you can also install Stellarium Mobile from Google Play

• Step outside, or go to a window, for up to an hour once it is dark

• Find one constellation you know and one that is new to you. In the mobile app, you can point your phone at the sky to identify a star

• In the Stellarium Mobile app, switch on another culture's constellations in the sky-culture settings and read the story behind one of them. If you are using the browser version, look for a similar setting in its menus

• Put the phone away for ten minutes and just look up, then tell someone the story you found

📢 Share the Optimism

If The AI Optimist helps you think more clearly, forward it to someone else handling the shift.

And here's the question I'm curious about this week: what is the one thing you would want to be clear about before asking AI anything for your organisation? Reply and tell me - I read every message and I'll come back to you personally.

Stay strategic, stay generous.

Hugo