Hi Reader!

Last week was a whopper week, spare a thought for Anthropic's lawyers who've probably had the worst weekend.

Let me take you through it.

And if all the dystopian angst is too much for you, skip to the happy bit right at the bottom.
​
You won't want to miss that.

Wed am: Dario's essay

On Wednesday 10 June, Anthropic CEO Dario Amodei published 'Policy on the AI Exponential', a sweeping, serious piece of writing, a far cry from his 'Machines of Loving Grace' that painted a picture of an abundant possible future. In the latest essay, he sets out all the possible things that can go wrong with AI and he calls for regulation of frontier models similar to what we have for aviation or pharma, where there is global testing and safety standards and less of a need for speed. He called for government powers to block deployments where necessary. On the face of it, a responsible thing to do.

And yet.

My read is that Dario is carefully building a record. Documenting his warnings, marking his position, so that when this comes to a head he can point back and say: I told you so. However, it's all very well calling out the bad things, while also running the company that is leading the way on building 'the thing'.

I was left with a feeling of despair thinking about who is actually going to save us?
You would expect our governments to look after us and step in. But governments are structurally, almost constitutionally, too slow. And the tools they have at their disposal like regulation, export controls, policy, all seem to be fit for a world that has gone.

Wed pm: Fable dropped

Wednesday evening, Anthropic released Claude Fable 5 to the public. Their most capable model, the one previously known as Mythos, was suddenly open to anyone on a paid plan.

I used it. You probably did too.

I gave it a multi-step task for something I was working on. And it did everything, start to finish, no edits required from me. It came up with the answer, wrote it, designed it, and for the first time since I started working in this space, I had a genuine moment of disorientation. Not excitement, not fear. Just a glimmer of 'what about me?' What will I do now?

That feeling, more than any of the benchmarks, was the honest signal for me that this model has entered into a whole new level of capability.

Within days, Fable was widely regarded as the most capable model in the world.

​
Friday 5.21pm

Friday evening the US government issued an export control directive ordering Anthropic to suspend all access to Fable 5 and Mythos 5 for any foreign national. That included Anthropic's own foreign national employees and you and me. To comply, Anthropic had to disable Fable (and Mythos) for every customer, everywhere. Active sessions ended. All new sessions fell back to Opus 4.8.

The stated reason: a jailbreak technique believed to bypass Fable's cybersecurity safeguards, reported by Amazon (bah!). Whether that's true or not, or it's just Trump's crew getting at Anthropic because of the whole Pentagon x Anthropic 'you're not using our work to drop bombs on innocent people' battle of several weeks ago. Timing-wise, embarrassing.

Spare a thought for Anthropic's legal team. That was not a pleasant weekend for them!

​
​

​
The sovereignty problem

This all raises the issue that we need to think seriously about AI sovereignty. Sharon Richardson of Hoare Lee called out that the overriding theme of London Tech Week was about the importance of data and AI sovereignty. It's a hot topic and playing out live on our screens.

So what happens if the model you're using disappears on a Friday night?

Where AI Institute stands

At the start of this year I set a goal: one AI agent per business pillar by the end of 2026. We are not halfway through the year. We have four.

I am sharing that not to show off how great we are, but because it a demonstration of how fast the world is moving around us. Even the stretchy goals we set in December now read like they were written by a different version of ourselves.

The other goal I set has now gone from 'ambitious end-of-year target' to 'this needs to happen soon': set up a local model for day-to-day tasks.

For some time now, I have felt genuinely uncomfortable using frontier-class compute for basic work. Summarising a document. Drafting a short email. Finding a recipe to use up the rest of the marscapone cheese. None of that needs Fable. None of it needs Opus. And after this week, none of it needs a model that a foreign government can turn off without warning.

Time to elevate the goal of having our own local models in place.

​
What a local model actually is

The Great Christopher Penn has a number of videos that take you through how to set up your own local model. He is a Genius. Listen to anything he says. A local model is one that runs entirely on your own hardware. No cloud, no API call, no third party. You download the model weights, they sit on your machine, and every query stays there.

This used to mean serious hardware and mediocre results. But not anymore. Models like Qwen 3.6, Minimax M3, and DeepSeek can run on a well-specced MacBook Pro and are every bit as good as what we called frontier, just 3 months ago. You don't have to have frontier for most day-to-day tasks.

The Fable weekend fiasco has made the case in my mind for introducing open models sooner than end of year.
I like the fact that no government directive can switch off a model that already lives on my hardware.

That is what AI sovereignty looks like in practice.
More from me on that later, no doubt.

Now to the good news : Claude in Chrome

So while all the big-picture, macro stuff is going on, let's zoom right in to the micro, and a happy story.

Last week on LinkedIn I ran a poll asking who is using the Claude extension in Chrome. The majority of people who responded are not. This has to be put right, because Claude in Chrome tickles me with such delight on a near-daily basis that I feel like I have to share this.
​
It does the things I genuinely do not want to work out how to do myself.

• Like beating the finest UX minds in the business, who hide unsubscribe buttons.
• Or setting it loose on the likes of Kit, and watching it whizz around doing the operational tasks I would rather not spend my brain juice on.
• Beau, our Copilot lead, suggests using to to point at Teams on Chrome, give it a knowledge base and have it respond to any questions that get sent to you.

Here is how it works. Claude sits in a side panel in your browser (see below). It can see what you see. You tell it what you want to do, for example I asked it to check that this email will pass anti-spam filters.

First it makes a plan which you need to approve. Then it just gets on with it - does the tasks, navigates pages, clicks buttons, fills forms, and completes multi-step tasks across tabs all while you watch, or while you get on with something else. It really is a case of 'look Ma, no hands!'

I like to use it for filling in forms and working out how to use saas software I couldn't be bothered learning how to use. In the case above, I got it to perform a check I probably wouldn't do if I was being my normal lazy human self. So not only does it free you up to do other things, but it also improves the quality of what you're working on.

How to get started with Claude in Chrome:

1. Go to the Chrome Web Store and search 'Claude', or go straight to claude.com/claude-for-chrome. Click 'Add to Chrome'.
2. Sign in with your Claude account. It needs a paid plan (Pro, Max, Team or Enterprise).
3. Open any new tab, click the Claude icon in your toolbar, and it appears as a side panel. Enable it for a conversation via the Connectors dropdown in Claude.ai and then you're all set.

Start small. Open a page you need to summarise and ask Claude to pull out the key points. Then try something with a few steps. Once you stop expecting it to be instant (browser AI is deliberate, not fast), the quality of what it produces is genuinely surprising.

​Tell me what you think. What did you use it for? I'd love to know.

Where's the podcast? ChattingGPT

We'll be back with a new series of the podcast soon. Some great conversations being recorded with interesting humans who are implementing AI. If you know someone who would be good to have on the show, please make a suggestion. Or if there's someone you'd love to listen to, just drop a name and I'll see what I can do.

Thanks for reading to the end.
​

As the old Indian proverb goes:
​

It'll all be alright in the end.

and it's not alright...

It's not the end.

See you next time.

AI Institute Team