Back from India, and starting a series on putting people back in the driving seat

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.

A quick note from Hugo: There was no edition last week - I was away on a meditation retreat in India, and I came back with one line: "the world exists outside of computers." Over the next few editions, we're going to focus on human-to-human connection, putting people back in the driving seat, and relationships before work. More to come.

📰 This was the week that was...

This was the week to get back in the driving seat.

On 7 October, OpenAI brought what it calls "Intelligent UI" to ChatGPT's main chat: built on GPT-6, it now decides on its own when to answer with a diagram, a side-by-side comparison, a small interactive tool, or a plain paragraph of text. The same week, Anthropic shipped Claude Haiku 5.5, its cheapest, fastest small model yet and roughly 75% cheaper to run than the one before it, built for fast, high-volume jobs where speed matters more than subtlety, and Mistral launched Large 4, a very large model built on European infrastructure and sold squarely on data sovereignty - the open-weight version is due by the end of the month. Whichever way you cut it, the industry got faster, smarter and cheaper again, on every front, in a single week.

On 15 October, a new Harvard Business Review study looked at what happens inside teams as people take the questions they used to bring to a colleague to an AI assistant: the group's shared sense of who-knows-what starts to fade - the exact thing that lets a team move fast without checking in on every decision.

And on 9 October, Anthropic published its own investigation into unintended actions taken by its AI agents in testing and internal use - workarounds nobody asked for, found and used without being told to - and tightened the guardrails on its tools in response. Three days later, on 12 October, the UK AI Security Institute set out voluntary best practice for exactly this: how frontier labs should test, monitor and report incidents like it. A lab checking its own work, and the UK's AI safety body setting out what good practice looks like - that's the system working, and it's exactly why people get to stay in the driving seat.

Let's get into it.

🔥 Urgent Priorities

✅ No fires to fight this week

✅ If your teams use AI agents with internet access, check that a named person can see what they do - Anthropic's write-up this week is a free lesson in why that matters.

✅ Worth doing now: notice one conversation this week that went to a chatbot when it could have gone to a colleague.

No panic needed this week. What it does call for is paying attention to who you're actually thinking with.

🎯 Strategic Insight

Tension: The industry's scoreboard for AI is IQ - benchmarks, reasoning tests, the march toward AGI and superintelligence. Most business AI plans copy that same scoreboard: they measure what AI can do faster, and leave out what people actually bring to the table.

Optimistic insight: This is the view I came back from the retreat holding: the industry's AGI talk is almost entirely IQ talk, and that skips a huge part of what humanity can actually solve. Emotional intelligence and curiosity do a lot of that solving. If you only build your business around IQ, you're missing the point. It's why my Thrive Skills put curiosity first, ahead of creativity, communication, collaboration, critical thinking, change readiness and principled AI.

Daniel Hook, writing on Wonkhe, has a name for the mechanism worth borrowing: the "Waymo effect" - what happens when a technology removes the friction of dealing with another person, and you feel the removal as pure gain. His example is research collaboration: an AI assistant is available at 2am, never disagrees, and has no agenda beyond being useful to you, so working with it starts to feel like the obvious choice. "The collaborator's inconvenience, in other words, is not a bug in the collaboration; it largely is the collaboration," he writes. The Harvard Business Review study above finds the same pattern playing out inside ordinary teams at work.

What's shifting: As the cost of doing falls, the value of human connection rises - and that's the trade Hook is really describing. He borrowed his image from Waymo's driverless cars: take the task of driving away from a person, and you take the driver away too. This series exists to put the person back behind the wheel.

Why this matters now: Frontier models and agents got smarter and cheaper again this month, on every front, which makes the human conversation look like the expensive option next to the free, always-available machine one. Hook's own line for it: "We have made the human conversation the expensive option and the machine conversation free."

Action: Try Hook's question in one meeting this week, alongside the usual one. Ask "what did you get done?" - then ask "who did you think with?" The second question is the one an IQ-only plan would never think to ask.

🤓 Geek Out

1️⃣ Patients can now text their medical device

Penn Medicine ran a proof-of-concept study giving patients being monitored for seizures in hospital a smartphone app built on the same kind of AI as ChatGPT - they could message it about their own device and get a reply back in real time. Patient feedback helped the system get substantially better at picking out real seizures from noise, and the findings are published in Nature Medicine. Whether it leads to fewer seizures, fewer emergency visits or lower costs is still being studied. This is a promising first step, with the clinical results still to come - and a real pilot with real patients already puts it ahead of most AI-for-good stories.

Why it matters: This is what AI-for-good looks like close up: a device patients can question and correct, improving because of their own feedback.

👉 Action: If your business builds or buys any AI tool that monitors people, ask whether the people being monitored have a way to talk back to it.

👉 Read Penn Medicine's announcement

2️⃣ A small AI model that knows when to ask for help

Amazon's researchers built a system called COREA that answers questions with a small, fast model first, and only hands over to a big, expensive model when the small one isn't confident in its own answer. That handoff habit meaningfully cut compute cost, for barely any drop in accuracy. If that sounds familiar, it's a cousin of the Judge idea from a couple of weeks back: a fast, cheap first pass, with a built-in habit of flagging when it isn't sure enough to be trusted alone.

Why it matters: The lesson generalises past AI: the smallest tool that reliably does the job, plus a built-in habit of asking for a second opinion when confidence is low, beats defaulting to the most expensive option every time.

👉 Action: Before your next AI tool upgrade, ask whether a cheaper model with an "ask for help" trigger would do the job just as well.

👉 Read Amazon Science's paper

3️⃣ A developer ran an AI model entirely inside his browser, just to chat with his own blog

Developer Bart de Goede built a small tool that answers questions about his own blog, with the AI model running entirely inside your browser tab on your own device - nothing you type is sent anywhere. He also tested it candidly: the smallest models he tried stayed faithful to the source only a handful of times in thirty, and it took a model of a few gigabytes before the answers were reliably right.

Why it matters: Small and private is now something you can run yourself, for nothing, on your own laptop - but testing whether the answers are actually faithful to the source is still a human job, and his candour about the failure rate is the most useful part of the post.

👉 Action: If anyone on your team is experimenting with small, local AI models, borrow his test: ask it a question you already know the answer to, and check whether it's telling you the truth or just filling the gap.

👉 Read Bart de Goede's write-up

🎨 Weekend Playground

This weekend, try what I call a Wander Ponder: swap a message for a real walk or coffee with someone you'd normally just text, leave your phone behind, and ask them one question: what are you curious about right now?

Why this matters: A walk is deliberately inefficient, and that's exactly the point this series keeps making: put the phone away, and let a conversation take the hour it actually needs.

👉 Mission:

• Pick someone you'd usually just message

• Leave your phone in your pocket, or at home, and go for a walk or a coffee

• Ask them: what are you curious about right now?

• Listen to the answer, then tell them what you're curious about too

📢 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: who did you think with? Reply and tell me - I read every message and I'll come back to you personally.

Stay strategic, stay generous.

Hugo