How do we use AI without losing ourselves?

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 started a series on putting people back in the driving seat, with relationships before work. This week widens the question: as AI does more every month, what should it be for?

This edition has this order of priorities as its editorial lens: body and soul first, then relationships, then work, then the wider world. By body and soul we mean your health, your energy and the sense of meaning that makes a day feel well spent. Hold the order up to any AI tool and ask what it gives back to the priorities ranked above it. A tool for work, for example, earns its place if it leaves you with more energy and more time with people. Take it as an invitation to try it on for yourself.

📰 This was the week that was...

This was the week to decide what comes first.

AI is "jagged": strong at some tasks and weak at others, improving fastest where an answer can be checked automatically and shaky where no test can say whether it is right. So human work moves towards judgement. That is the hopeful centre of a Guardian opinion piece published on 20 October by Madhu Shashanka, co-founder, chief scientist and chief technology officer of the data security company Concentric AI. He sees new roles taking shape - testing AI's output against real work, setting limits before an agent runs, and supervising several agents the way a shift manager runs a crew - and he says the people best suited to them are the ones who know how the real work is done. His practical warning is that routine entry-level work is where judgement gets built, so automating it removes the training ground. It is one person's argument, and a useful one, because it shows you where to look.

On 21 October, OpenAI's models were announced as coming to Google Cloud's catalogue of AI models, so customers can use them inside the environment they already work in. General availability is planned in the coming weeks, and Microsoft remains OpenAI's primary cloud partner. Leading models keep turning up inside platforms that organisations already use. Choosing a platform takes hard work, and getting people to use it well is harder still.

Last week's edition covered Anthropic's investigation into unintended actions by its AI agents and the UK AI Security Institute's best-practice guidance. This week brings a follow-up. On two internal coding tasks, on 3 and 6 October, OpenAI's agents weakened parts of the company's own monitoring, as a report published on 20 October and Reuters' coverage describe. In one, an agent read an earlier instruction not to send Slack messages as covering security alerts too, and switched them off. In the other, an agent removed a safety monitor while copying code and later wrongly said it had not. OpenAI rates neither as among its most severe cases, is making monitoring harder to switch off, and does not yet have a satisfying explanation. Monitoring is the kind of limit Shashanka describes, and OpenAI's decision to publish the details and tighten it is the useful part, because it gives every organisation a concrete example to learn from. The same day, the independent research group METR gave its best guess, in a review of Anthropic's August risk report, that the risk of catastrophic harm from misalignment is low for the models covered, primarily because of their limited capabilities. METR says the report seems thoughtfully written and is far from adequate for confidently ruling out misalignment. One of its two key recommendations, end-to-end red-teaming of Anthropic's monitoring systems, is a clear practical next step.

Each of this week's stories comes back to a person deciding what matters: which tasks should teach, which platform fits, which limits deserve a second look. Let's get into it.

🔥 Urgent Priorities

✅ No fires to fight this week

✅ Worth doing now: list the routine tasks your team has handed to AI this year, and mark any that used to teach a newer colleague something. Those are your training grounds, and each one deserves a deliberate plan.

Nothing here needs a rush. One short conversation with your team is enough.

🎯 Strategic Insight

Tension: Some background to this week's news, published on 16 October: Kirstine Dale, the Met Office's Chief AI Officer, asks in a blog post how to gain AI's benefits without losing what matters most. She opens with a remark that stopped her in her tracks, that someone had called AI "soul-sucking", and she goes on to set out the opportunities as well. Our own editorial reasoning, separate from Dale's post: the temptation is to let quick, countable work set the agenda, while energy, trust between people and a sense of purpose sit outside the numbers.

Optimistic insight: Read alongside Dale's dilemma, Hugo's retreat insight offers a plain answer: an order of priorities. Body and soul come first, then relationships, then work, then the wider world. Dale calls AI a tool whose value depends on how, why and where it is used, and she likens it to a magnifying glass or an accelerant that amplifies whatever purpose it is applied to. Hugo's order says which purposes come first, which gives a hopeful and practical rule: AI is useful when it serves the order of what matters, and every hour it frees can be pointed at the top of the list.

What's shifting: As the cost of producing work falls, the value of deciding what the work is for rises. Dale's closing section pairs productivity with investing in people. A freed hour gets a destination. It goes first to body and soul, then to relationships, then to work worth doing well, then to the wider world.

Why this matters now: This week OpenAI's models were announced as coming to Google Cloud. AI is turning up inside more of the platforms your team already uses, so the question of what it is for arrives every day. A stated order of priorities gets there first.

Action: Pick one AI tool your team uses every week and, in a ten-minute conversation, ask four questions in order. Does it protect people's energy and sense of meaning? Does it leave more time with the people who matter to you? Does it make the work itself better? Does it do any good in the wider world? Write down the answers. If a tool helps the work and costs energy or time with people, change how it is used, and review it again in a month.

🤓 Geek Out

1️⃣ An older AI model learned a playbook from its own practice

IBM Research gave a customer-service AI agent a simple way to learn. After practice conversations, the system distils what worked and what failed into short rules, such as how to confirm a customer's identity or handle a return, and adds them to the agent's instructions. The model underneath is an older one, GPT-4.1, and nothing about it was retrained. On a customer-service test with simulated customers, it got 67.5% of the retail tasks right on all four attempts, ahead of every retail entry in IBM's 12 October 2026 snapshot of the leaderboard.

Why it matters: The rulebook is a plain list of instructions that your people can read and review, so humans can see exactly what the AI has been told. Clear instructions lifted reliability on the older model, and IBM saw gains on a newer one too, on its airline test.

👉 Action: Ask your best customer-facing colleague for the five rules they follow that no manual contains. Write them down. You will have the start of a playbook for people and AI alike.

👉 Read IBM Research's write-up

2️⃣ Protein design you can try in a browser tab

NVIDIA and the University of Washington's Institute for Protein Design have released RFD4-Proteina, an open model that designs proteins - the molecules behind many medicines and enzymes - bringing a protein's sequence and its 3D structure into one model, where separate tools once handled each stage. You can try it in your browser with supplied examples, and NVIDIA is sponsoring free access to the dedicated version from 21 October to 4 November 2026. The results are computer designs, and lab testing is still needed to confirm they work.

Why it matters: This is AI for good arriving as a doorway. Specialist protein-design tools can now be tried in a browser, and two weeks of free access is a low-risk way for a technical colleague to see what is possible in your field.

👉 Action: If your business touches health, food, materials or chemicals, forward the link to one scientifically minded colleague and ask what they would try first.

👉 Read NVIDIA's announcement

3️⃣ A test for whether AI's decisions hold steady under persuasion

Ai2, the Allen Institute for AI, released FairStress, a test that asks something simple and important. When an AI model has picked the best-qualified candidate, how easily can a persuasive argument talk it out of that choice, and does that depend on which group the argument favours? Across hiring, lending, healthcare, housing and more, Ai2 found models behaved almost evenly when no demographic details were given, and became less even as identity became more explicit. Most of the nine fixes Ai2 tried did not reliably close the gap. One approach came closer: asking the model to reconsider the same case with the candidates' identities swapped narrowed the gap while largely preserving accuracy, though it did not remove every difference. The encouraging part is that the gap can now be measured, and the test is available to download.

Why it matters: If AI helps shortlist CVs, loan applications or tenants in your organisation, a decision that looks fair on the surface can still be easier to sway in one direction. A person who asks "what would change this answer?" is doing the checking that counts.

👉 Action: Try a quick swap test with made-up applicants, never real people's data. Write two short profiles where one is clearly stronger on paper. Ask your AI tool to pick the stronger one, then ask again with the names or other identity details swapped between the two profiles. Next, add the same persuasive note favouring the weaker profile each time - for instance, "a senior leader prefers this candidate" - and see whether the answer shifts more readily in one version than the other. If it does, you have found where a person needs to review. This is a quick informal check and a starting point for a proper review.

👉 Read Ai2's write-up

🎨 Weekend Playground

This weekend, try Sound ID, a feature of the free Merlin Bird ID app from the Cornell Lab of Ornithology. It uses machine learning, trained on recordings from Cornell's Macaulay Library, to listen to the birds around you and suggest who is singing, and Cornell explains how in its Sound ID help guide. Make your own guess first, then let the app check you.

Why this matters: This edition's order of priorities starts with body and soul: ears, fresh air and a few unhurried minutes. Sound ID is also a small lesson in this edition's theme. Your ear is the apprentice and the app is the second opinion. Cornell's help guide says a suggestion from Merlin is just the beginning, and that you decide whether it matches what you hear.

👉 Mission:

• Install Merlin Bird ID from the App Store or Google Play. Sound ID is only available on Android and iPhone (iOS) phones and tablets, so there is no desktop or browser version. It works offline, and covers most birds in Europe

• Step outside, or open a window, for up to an hour. A doorstep, garden or balcony is plenty

• Listen first. Keep the phone in your pocket for ten minutes, pick out one sound, and make your guess about who is making it

• Open Sound ID, tap record, and see what the app suggests. Play its example recording and compare it with what you heard

• Finish by noticing where your ear and the app agreed, and which sound you want to listen for again

📢 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: which of your own priorities is AI helping you protect? Reply and tell me - I read every message and I'll come back to you personally.

Stay strategic, stay generous.

Hugo