AI for the rest of us

Welcome to the AI community for everyone.

Hello friends,

Thank you to everyone who joined us at Agent Craft this week. It’s amazing what happens when you put smart people in a room together and just let them talk about what they care about. A few people said it was their highlight of London Tech Week, but the greatest compliment was one attendee who said “I’ve met people here I genuinely want to see again” and that’s the whole point of community.

We all have so much to learn and the AI space can feel especially fast and overwhelming. AI for the rest of us is a community that’s navigating these changes together.

As an unconference we built the agenda on the day based on what attendees wanted to talk about. This is the agenda we came up with. Hannah particularly enjoyed the sessions “Quick start advice for building agents” and “Measuring business outcomes and ROI”.

Yes, we had an anxiety party. This is a tool that Hannah has used in a work context to create space for her team to voice their most irrational or catastrophic fears. It was cathartic to hear the AI enthusiasts in the room share their worries in an honest and open way. You might want to try running an AI Anxiety Party with your team at work!

While working on a case study for a client Charles invented a new job role - the DevOops engineer - and thinks that should really be a thing.

Charles and his family were at the theatre on Wednesday afternoon to see “Thespians” — subtitled “Greece the musical (not that one)”. It’s by Mischief, the company behind the “goes wrong” series of plays, and purports to tell the story of the invention of acting in ancient Greece, finding little has changed. The proto-thespians play Zip, Zap, Boing — “How does this help?” “I don’t know!” — over-dwell on their motivations, and grapple with one star reviews.

The production heralds shoestring production values (this Greek chorus numbers, er, two) with lots of fourth wall breaking, groan-guaranteed puns and daft wordplay. Writer Sayer and composer-orchestrator Zanders check off every musical theatre staple, from the “I want” song to the act two opener, a big villain tune and an 11 o’clock number. It was excellent. Clever, funny, camp, and very, very silly, with moments that are sweetly affecting as acting teaches our motley crew the value of empathy. When two of the male thespians finally kissed there were audible “aaahs” from a group of schoolchildren from many rows away, and afterwards Charles overheard one of them excitedly tell another, “It was the best thing I’ve ever seen”. Theatre is wonderful.

Have a wonderful week!

Hannah & Charles

What’s Hannah reading this week?

It was London Tech Week so I rented an apartment on Brick Lane and embraced the chaos; attending 3 conferences, 2 meetups and an over 30’s day disco (the latter was not part of London Tech Week but it was wonderfully silly!)

The Minister for AI and Online Safety Kanishka Narayan MP spoke at The AI Summit about the importance of open-source AI, declaring that there will be investment, support and access to the compute we need.

“Britain needs to become the global home of Open-Source AI… We will back you with the compute you need to keep going…Here’s the compute, here’s the support, let’s go and build”

I also appreciated the message about responsible AI. Put simply “Yes, but safely”. We must not wait for “perfect certainty” if we want to be leaders in AI.

FYI: perfect certainty isn’t coming. Not with LLMs anyway, that’s not how the technology works. If you want certainty, build a deterministic system. I feel like one day we’ll look back at the lunacy of this particular moment in tech history and laugh about all the things we tried to make LLMs do for us.

We have many international readers of this newsletter so I won’t inflict too much UK politics on you but there is a growing sentiment in the UK and throughout the EU that dependence on US based tech giants is a risk, and we should be investing in the infrastructure we need to provide more options for local companies and institutions. You can’t go to a tech event without the words “Sovereign AI” being muttered.

As an ambassador for OpenUK I was happy to see open-source at the heart of Kanishka’s speech on Wednesday, however Amanda Brock (CEO, OpenUK) thinks we need to go further. If anyone knows how to make open-source succeed it’s Amanda and she shared her views with Computing:

“We need the actual infrastructure that sits behind the open-source ecosystem in place and proper funding of open-source. The words “open-source” are not magic words. We need to follow the tried and tested recipe for open-source success if the goal is to lead”

With a US government directive causing Anthropic to withdraw access to both Fable 5 and Mythos 5 to any foreign national (even Anthropic's own team), I am once again left worrying about the inequality we may be creating when one single government can decide who gets access to this technology, and who does not. Open models are part of the solution for sure. They must be truly open and accessible to all for this to work.

Multiple tech unicorns seized the moment at London Tech Week to promote their hiring initiatives in London, including Cursor announcing 200 new jobs. It’s bloody brilliant to see jobs being created, it really is, but I live in a village in Yorkshire and I’m committed to country life. Remote work is fundamentally more inclusive than office based hiring so I hope there will be at least a few remote UK hires in this glut. I don’t need a shiny office, I need to be near my family. That’s it.

This week I also enjoyed Laurie Voss’ explainer “How do you make an LLM anyway?”, Josh Bresser’s tear down of current security practices “We have to change the rules of security” and the hilariously stupid idea that Klarnas CMO can let an AI Agent listen to his employees complain instead of actually listening to his employees complain “Why Klarna's CMO built an AI replica of himself for employees to vent at”.

What's Charles reading this week?

Our trip to the theatre got me thinking about the value of art, and why it is that I can’t get excited by attempts at AI generated art, whereas I found earlier attempts at procedural music — Mozart’s dice game, or systems music — fascinating. So, at the risk of sounding desperately pompous, I'm going to risk sharing those thoughts with you. If you'd prefer that I hadn't, might I suggest you skip the next four paragraphs.

I think that art needs to do one or more of four things. The first is to inspire wonder. That sense when you see a great play or film or hear a piece of music of “How on earth did you make that?” It doesn't have to be high art either; a good blockbuster film or a well crafted pop lyric can do it just as well.

The second thing I think is curiosity. Perhaps it makes you wonder whether something is true. Or perhaps it makes you curious to try to understand how the thing was made. The best art even survives that latter process. I spent many happy hours in the Barbican library trying to figure out what made the second movement of Sibelius’ third symphony tick, because the first time I heard it it made me cry. Even now I know the mechanics and can see the machinery, when I hear it performed all of that melts away and it remains just as moving.

The third and fourth are to do with empathy and self reflection. I think a lot of great art helps us to reflect on our own actions and think about occasions when we've maybe not been as good as we want to be. As a side effect I think it also helps us to empathise, to, as they put it in Thespians, “stand in another person's sandals”. Musk, of course, thinks that empathy is a weakness, and I think he is an idiot, albeit a very rich one.

AI generated music has none of that: there's no wonder as to how it was created; it doesn't extend empathy; it doesn't prompt self reflection. It puts me in mind of the Auditors of Reality in Terry Pratchett's "Thief of Time”, who try to understand art by grinding it up, and are baffled as to why that doesn’t work: "In fact, they had spent some time in the Royal Art Museum only that morning. They had dismantled several priceless landscapes into their component atoms in an attempt to find out where the beauty was kept.” I suspect in time great artists will find ways to use AI in interesting ways, but a piece of art that is made solely by an AI just bores me, and I think it is because there isn’t any work going on.

Whilst Thespians is wild and brave, Anthropic’s publicly released Claude Fable 5, its first “Mythos-class” model and one it says surpasses its previous frontier Opus models in overall capabilities, risks being overly cautious. The model comes with safeguards designed to prevent it from answering queries on topics like cybersecurity, biology, and chemistry, where the company has publicly worried about its potential impact to “uplift” malicious actors. If the safety classifiers detect a query regarding biology, chemistry, or cybersecurity, the system automatically redirects the query to Claude Opus 4.8 (an older, less capable model) to answer it instead.

I don’t have an Anthropic subscription so have been unable to test it myself but Ars Technica’s Kyle Orland and others noted that the filters are so strict that Fable 5 is refusing to answer basic, completely benign educational and medical questions, such as general queries about cancer, the mitochondria, or standard high school biology. Anthropic states that these filters trigger in fewer than 5% of user sessions, but the friction has caused frustration among researchers and medical professionals who find themselves blocked from standard scientific work. In Anthropic's defence AI safety is complex, and I’d rather they were overly cautious than not cautious enough, but I’d be even happier if they spent more time to get it right.

[Since writing access to Fable 5 has been withdrawn due to a US government directive citing national security concerns. As the BBC’s Harry Sekulich notes “Anthropic has found itself in the crosshairs of the Trump administration recently.”]

Apple held its big WWDC developer conference this week. During the keynote they finally introduced the long-delayed “Apple Intelligence” update for the Siri voice assistant. Senior Vice President of Software Engineering Craig Federighi opened that segment by saying that “AI is an incredibly powerful technology with the potential to shape society in profound ways, and, with proper care, unlock meaningful benefits for people everywhere. Still, some appear to be racing forward, seemingly pursuing AI for the sake of AI, without clear regard for the people, all of us, that it’s ultimately meant to serve.” Subtext: “but not us.”

The new “Siri AI” — now being promised for OS updates rolling out “this fall” — will come alongside a new Google-powered update to Apple’s on-device Foundation Models, as well as tighter integration of all these AI capabilities across Apple’s many operating systems. The company demoed it in a series of scripted conversations with Siri AI, complete with seemingly unedited, multi-second pauses between each spoken prompt and Siri’s response. It looks impressive, even if it's doing things we’ve seen elsewhere before. There’s also some welcome options to tweak the pace and expressiveness of Siri’s voice. When it ships it will only be available in English, and some features will have daily usage limits requiring a subscription for more use.

Apple is building this whilst spending a lot less than its competitors. In 2026, the four largest American cloud and AI infrastructure providers — Amazon, Google, Meta, Microsoft — committed to a combined spend of $670 billion on capex. Apple, in contrast, spent $12.7 billion on capex last fiscal year and projects $14 billion for 2026, i.e 8% of its peers’ average spend. So Apple is making an enormous bet on AI: their bet being that they don’t need to spend hundreds of billions per year on AI infrastructure to reap the benefits.

Apple also spent a lot of time emphasising that Siri is “privacy first”. Following the keynote, Federighi and other executives got on a smaller stage to explain to the press and other media how it plans to preserve user privacy while still getting the kind of compute capacity it needs from Google, for whom privacy is rather less of a strong suit. My invite obviously got lost in the post(!), but Ars Technica’s Andrew Cunningham has a good write-up of how it works. In brief, Apple is for the first time expanding its Private Cloud Compute (PCC) framework outside of its own data centers to run complex, heavy AI workloads on Google Cloud infrastructure using NVIDIA GPU. By partnering with Google Cloud and NVIDIA, Apple gets the hardware scale it needs while using architectural locks to ensure that Google functions strictly as a blind host for the data, keeping user requests completely private.

Apple has also confirmed that due, it claims, to the Digital Markets Act (DMA), it is not be able to ship Siri AI in the European Union with the release of iOS 27 and iPadOS 27, and that “there is currently no timeline for Siri AI’s availability in the EU on iOS and iPadOS.” Siri AI will though be available to EU users on macOS 27 and visionOS 27.

Looking at the AI segment as a whole there are some odd contradictions. Apple emphasised that it had “deep respect for the art of photography,” for instance, but appeared to have rather less for the art of writing or other art forms.

As Elton John famously noted, Sorry Seems to Be the Hardest Word. Apple didn’t manage to apologise for the horror of last year's MacOS Tahoe interface, but it is working to fix some of its worst decisions. Tahoe runs on a client laptop I have to use and annoys me every… single… day. Golden Gate, this year’s MacOS, has improvements to readability, including a new slider to let you change the level of transparency of “liquid glass”, a proper toolbar design, and the removal of some of the more egregious interface distractions. As a dyslexic writer and Mac user I really appreciate this. My own Mac is still running last year’s Sequoia but might get to run Golden Gate.

Finally from me in this very long update, one of the joys and absurd privileges of what I do for a living is that occasionally people send me previews of books they've written and, assuming I like the book, I get to interview them about it afterwards. I even get paid for it. Admittedly not much or very often, but it all helps. On the GOTO book club this week, I sat down with Elisabeth Hendrickson and Joel Tulsi to talk about their systems thinking book “Signals and Levers” which is coming out on IT Revolution in the autumn. It's a fantastic book and they were both wonderful podcast guests. It is a bit off topic but I would really encourage you to give it a listen and read the book when it comes out.