AI for the rest of us

Welcome to the AI community for everyone.

Hello friends,

Join us on Wednesday 5th August at 12:00 BST for our August virtual event. Hannah will be joined by Emma Burrows, founder of Rezonant, who will share how to build a self-driving product. Hannah has been quoting Emma at conferences all year, and we finally met in person for the first time at the AI for the rest of us summer social a couple of weeks ago. If you're curious about how to use AI in product management or you’re working in software development you’ll love this talk!

The virtual meetup will be hosted on LinkedIn Live and streamed on YouTube. See you there!

Here in Yorkshire we've kicked off the school holidays in chaotic Foxwell style. Hannah and her niece opened an art gallery in the garden, the whole family was treated to a “gig” with original songs composed by the seven year old and a wooden spoon drumming performance from the ten year old that had the neighbours grumbling. On friday we took to the water at Allerthorpe Lakeland Park where we rented a canoe, only to travel around in circles for an hour because we’re not exactly accomplished rowers. Hannah returned home on Friday covered in lake water and sweat with no voice left. Summer holidays are THE BEST!

Charles and his wife have been watching the BBC’s new show Evolution, which takes one animal an episode and delves into a particular aspect of it and the evolutionary journey it represents. It is presented by Chris Packham in a style reminiscent of a younger David Attenborough, and the programme has lots of proper science but with Packham happy to express a sense of wonder at the miraculous-seeming nature of it all. It's excellent.

Have a wonderful week,

Hannah & Charles

Teach your agents when to say no

Enforceable boundaries around every AI agent. The autonomy you can't fully predict becomes autonomy you can trust — because it knows where to stop. Trust infrastructure for AI

What’s Charles reading this week?

I didn’t cover this story last week partly because it is actually quite an old story, and partly because I wanted some time to reflect on it because it really bothered me. I have a bit of a thing about the unnecessary destruction of objects, which I think I’ve always had. When I was very young (I don’t remember what age) I remember watching an episode of Top of the Pops with my sister where the act destroyed a piano on stage with, I think, axes. I remember having a weird, visceral reaction to it and finding the whole thing unsettling. I had a similar reaction when I watched British artist Michael Landy, who destroyed all 7,227 of his possessions in a 2001 performance art project titled Break Down at a former C&A department store on Oxford Street, London. I also found myself thinking, “If you don’t want the stuff, why not donate it to people who do”.

So…if you read last week’s newsletter you may remember I talked about the Anthropic copyright case and how training on books was fair use, but that Anthropic's piracy of those books (rather than buying them) likely wasn’t. What does that mean in practice?

Well, with much of the content available online now exhausted, and increasingly polluted with poor quality AI generated writing, the AI labs are turning to uncorrupted texts published pre-2002. Anthropic and other AI companies have been buying up lots of rare books, and using them to train their AI models. Which might be good, although perhaps only briefly, for antiquarian book sellers who, in my experience, tend to be lovely people. The downer is that the books don’t survive the process.

Why don’t the books survive? Court documents detailed how the AI lab used a “hydraulic powered cutting machine” to “neatly cut” millions of books before scanning the pages “on high speed, high quality, production level scanners”. The remnants of the books are then pulped. Anthropic has thus far destroyed about 2 million books in total. They described the project, called Project Panama, as “our effort to disruptively scan all the world’s books” in an internal memo. It also said, “We do not want it to be known that we are pursuing this project”.

The story is back in the news because 404 Media published a new/updated version of it but it's actually been known for a while. One small book seller said that in April he suddenly went from selling no more than 20 books a week to hundreds, and he’s almost certain that the customers are AI labs, noting the random selection of the books and how they all have ISBNs. He added that his inventory is full of rare and out of print books, meaning that an AI company could be destroying some of the few remaining copies that can be found.

“I personally have mixed feelings about all of this,” the bookseller told 404. “It benefits me financially as well as by clearing out old inventory that is otherwise unlikely to sell. I’ve been well-suited for these sales with inventory from overseas and foreign language books. On the other hand, I don’t like the end-use, and I don’t like that uncommon books are being pulped.”

One company, ISBNdb, which boasts that it has the “world’s largest book database”, now offers bulk book buying for AI labs, “up to one million titles per order”, including “older, rare and specialist volumes”.

“The world’s best AI training data is sitting on a shelf,” it states on its website. In an article, ISBNdb stated, “Anthropic kept Project Panama confidential, not because it was illegal, but because it looked bad. Destroying millions of books evokes images of burning libraries, even if the law was on their side. When the project became public through discovery documents in early 2026, the narrative was already set against them.” Of course there is another option; scan the books the hard way and preserve them in a library.

Universities are worried about AI-enabled cheating, but some are backing off from ‘AI-detectors’ because they’re basically snake-oil. The FT’s Ima Jackson-Obot has an excellent write-up on this.

There’s an interesting proposal from the British energy regulator Ofgem, that data centre developers in the UK pay a deposit between £237,500 to £712,500 per megawatt — meaning data centres seeking 1 gigawatt (GW) of power would have to pay hundreds of millions up front, paid back if the project was completed. The plans are aimed at tackling "speculative projects from securing scarce network capacity without any firm intention to connect", the regulator said. Writing for the BBC Laura Cress notes that, “The amount of electricity capacity being requested by projects seeking to connect to the grid had risen from 41 GW to 125 GW in the past year, reflecting a sharp increase in demand. This is significantly more than double 2025's peak electricity demand in Britain of around 46 GW.”

HAWK, a quantum-resistant cryptography algorithm that was under consideration as an official US standard has been taken out of the running after an Anthropic security model helped find a flaw that rendered it broken. Prior to this HAWK had already survived two rounds of expert human scrutiny over two years, a sign it was considered reasonably solid.

Claude Mythos found a mathematical shortcut that nobody had spotted before: a hidden pattern (called a "symmetry" or "automorphism") in the underlying maths HAWK relies on. Exploiting this pattern lets an attacker crack HAWK roughly twice as fast as previously thought — in cryptography terms, it's like discovering the lock has effectively half as many possible combinations as advertised. Concretely, for one version of HAWK (HAWK-256), the estimated cost of breaking it dropped from an astronomically large number (2^64 attempts) to a merely very large number (2^38 attempts) — still not practical to crack today, but a big jump, and one that matters more for HAWK's bigger key sizes, which were supposed to offer much stronger, longer-lasting protection.

HAWK isn't actually deployed anywhere yet, so nothing you use today is affected. But because of this finding, HAWK's designers would need to roughly double their key sizes to restore the intended level of security. That extra size removes part of what made HAWK attractive as a candidate in the first place (smaller, faster signatures than competitors). Anthropic disclosed the finding responsibly, sharing it with HAWK's own designers in advance and coordinating public release through NIST's official channels. The AI vendor sees the whole episode as the standardisation process working as intended, by catching weaknesses before a scheme gets locked in and widely deployed, rather than after.

What's Hannah reading this week?

I’m starting with some good news this week. Not all billionaires are evil.

Cliff Obrecht and Melanie Perkins, co-founders of Canva have pledged 30% of an estimated $42bn to their non-profit foundation, The Canva Foundation, and they are asking the world to tell them how this money should be put to work. The Goals Platform already has two pilot programmes up and running in Australia and the US but is asking the world to share their goals. Go ahead and submit your goals, add your voice and influence how these billions are invested!

What's one goal you'd like to see in your lifetime for the world, your country or your local community?

On LinkedIn Lavinia Kanagandran writes “AI isn’t the problem, capitalism is”. Lavinia’s post echoes many of the concerns I share, and concerns I’ve heard expressed from others in the community. The impact of job losses felt by employees at the bottom, the surplus flowing into a concentration of a few companies at the top.

We are living through what might be the fastest redistribution of economic power in modern history, and the public conversation about it is almost entirely about technology.

In October this year Deeply Human Innovation will host their first summit in Paris. I spoke to Founder and CEO Iliana Grosse-Buening about the event and was impressed by how direct Iliana and her team are being about action. This is not an event to sit and listen, it’s an event to organise and act. I’ll share a discount code with this community as soon as it’s available but in the meantime save the date for October 2nd.

Another piece of good news this week is that LinkedIn has introduced the “Seems like AI Slop” option to report a post. I can’t tell you how happy this has made me! I will be using the button! I hope other people join me in some gleeful reporting of AI Slop so that the posters eventually stop. If you want to share your thoughts online then just share your bloody thoughts online! If you don’t have thoughts to share online then simply don’t post… it’s not that hard.

As a Maths nerd (first class degree don’t you know!) I was interested to read this story in The Wall Street Journal about Jacob Tsimerman, who just won a prestigious Maths prize, left academia and took a job at OpenAI.

“This is someone who recently wrote a paper categorizing the ways AI might kill everyone, applying a mathematician’s instinct for order to the apocalypse… It turns out he’s so worried about the dangers of AI that he’s now pivoting to work on AI safety.”

We have seen versions of this story before. People who want to have an impact are drawn towards the places and organisations where they feel they can have the most impact. I hope Tsimerman can have that impact at OpenAI.

Another interesting move was Tom Blomfield, the serial entrepreneur who founded GoCardless and Monzo. Tom has been working as a partner at Y Combinator but has left to work full time at Anthropic. In a recent talk Tom shares his vision of what an agentic self-improving company might look like.

This is a great talk and well worth 13 minutes of your time. Part of Tom’s vision involves recording everything that happens at a company and distilling that into context to be used later.

That might be necessary to give the AI agents what they need but does it give the humans what they need? Do we want to be recorded while we work? Does it create a safe space to disagree? To share unfiltered opinions or ideas? What does a future of constant recording and inspection do to psychological safety?

The answer has to fall somewhere in between. We definitely need better systems and processes to capture the context of our “work” in a way that’s low friction and practically invisible. But we can’t do it in a way that impacts the performance of our people and people need psychological safety.

Finally, I wanted to share the reports that the conflict between US and Iran may have moved to cyber-warfare, with CBS reporting that the U.S. is investigating if Iran was behind cyberattack on water systems in 7 states, including Minnesota. I share this because I constantly bang on about cyber-security in this newsletter and it can often seem hypothetical. We need to protect ourselves in case there is an attack vs. we need to protect ourselves for when there is an attack.

I’ve been focussed on user research for BIMP over the past couple of weeks so I’ve been interviewing teams about some of their cyber-security practices. I realise how hard this is, so there is no shade intended, but OMG the real world is worse than I thought. I had a rant about it on LinkedIn because it’s genuinely terrifying.

I urge everyone to do a pre-mortem this week. Pretend you got hacked. How did they get in? You might find this activity alone is enough to drive a few critical improvements.