AI for the rest of us

Welcome to the AI community for everyone.

Hello friends,

Charles went with his wife and youngest to see recreational mathematician, stand-up comedian, best-selling author, and YouTuber Matt Parker in his new show, "Getting Triggy with It" at the Hexagon in Reading on Wednesday. Matt kicked things off by dividing the audience into two groups: ‘Category Zero’ for the fellow maths geeks who watch his videos (Charles’ wife and youngest), and ‘Category One’ for those who’ve been brought along as a plus-one (Charles). It’s part of a brilliant opening that sets the stage for a show that cleverly caters to both. The show happily dives into the nerdy joy of optimising sort functions and the unexpected patterns of Markov chains via the chorus of Blur’s Boys and Girls. Plus there’s a giant robot DJ in it, and one the clearest explanations I’ve ever heard as to what an LLM actually does. Smart, silly, and wonderfully entertaining.

Charles has also been thoroughly enjoying the fact that his favourite composer, Jean Sibelius, has been Radio 3’s composer of the week.

It’s International Women’s Day today and so Hannah feels like she should say something wise about that.

A couple of years ago Hannah gave a talk at QCon about people management where she shared a graph without any labels and asked the audience if they knew what it was. No one could answer. It was the graph of the hormonal changes a woman’s body goes through each month. Hannah did that at a tech conference where people were expecting to see content about engineering leadership. It was a wonderful moment!

It’s foolish to pretend these monthly fluctuations don’t exist. Us women don’t show up the same every day and shouldn’t have to pretend that we do - some days we'll be more creative, more courageous, more reflective (and some days we’ll be in pain because that’s also a thing). These are strengths not weaknesses.

I recently chatted with Georgie Powell, the founder of Phase, an app which aims to bring harmony to our working lives by helping women plan the right work at the right time. I love this idea. I think that more of us, both men and women, should understand how our bodies work and work with them, not against them. One thing I’m thinking about on International Women’s Day is how we can create environments where everyone can succeed. Doing the right work at the right time seems like a solid start.

Have a wonderful week!

Hannah and Charles

Leapter - The Logic Layer for AI Agents

Leapter is an exciting new tool that uses AI to generate code in a visual way, so no one needs to go through the painful process of reviewing AI generated code line-by-line.

It's free to try so why not give it a go?

Visit Leapter.com

What’s Hannah reading this week?

This made me laugh so I just had to share it with you all: Why do all AI Company Logos look like buttholes? You’re welcome.

The Australian retailer Woolworths has been in the news this week because its chatbot has been pissing off customers. The "obnoxious" chatbot Olive kept claiming to be human, attempting to engage in fake banter, and even telling stories about its mother. Oh dear. I’m still quite convinced that customer support is a poor use case for AI Agents; these are your unhappiest customers, you could take the opportunity to turn things around with great customer service or you can just hand them over to a bot. I know which I would prefer.

Following on from the shenanigans last week between the US Department of War, Anthropic and OpenAI we saw a glorious backlash. Claude soared to the top app on the app store, suffering multiple service interruptions due to the apparent surge in demand. This piece in the Guardian was even more direct with the headline “Quit ChatGPT: right now! Your subscription is bankrolling authoritarianism”.

Apparently Anthropic have now reopened negotiations with the Pentagon. I can’t help but feel disappointed. If there are any tech CEOs that you genuinely look up to as role models please send them my way because I’m becoming very cynical indeed.

The National Cyber Security Centre (NCSC) has issued a warning to all organisations in the UK to review their cyber-security posture in light of the conflict in the Middle East. Meanwhile even a security expert at Meta was left looking rather foolish after her own OpenClaw AI Agent deleted all her emails. If the security experts are messing up, what are the chances your team is getting this right? Just saying.

What if we designed systems for human flourishing? This is the question being asked by both Matthew Skelton and Bob Marshall (aka FlowchainSensei). In Matthew’s post “Unlock the benefits of AI by designing for humans” he talks about the flow of work, and cognitive load being key for success with humans and agents alike.

“The foundational principles for rapid value delivery remain the same whether you are using humans, AI, or both: plenty of the right context, clearly stated missions, clear success criteria, bounded scope, and explicit guardrails.”

Bob focuses more on human needs in his excellent post “What Would Software Development Look Like If We Started From Human Flourishing?” Whether that be colleagues, customers or investors those needs should be met if your team or business is to be successful.

“What if we started instead with: What do the people involved in this work actually need — to thrive, to do meaningful work, to live well? This isn’t soft. It’s radical. And the organisation it produces looks almost nothing like what we’re used to.”

This is a topic I find myself reflecting on more and more. Can AI help us reshape work into something better, something more humane, something that isn’t a burden but a pleasure? When Charles and I meet each week we often get into this discussion, with both of us freelance, we benefit from a lot of freedom to choose what we work on and how we work. What would you do if money wasn’t an issue? I’m fortunate in that I can say “I’d be doing this” and really mean it.

What's Charles reading this week?

If you missed my Green AI talk at ‘AI for the rest of us’ last year — or just fancy another dose — I'm taking it to Green Software Brighton on 18th March. It's free, and a hybrid event, so you can join from wherever you're currently drinking your coffee.

Also — and do forgive the shameless plug — I have a piece over at The New Stack on the Commonhaus Foundation and what the XZ Utils backdoor incident revealed about how poorly we support solo open source maintainers. The headline is pure clickbait (it works, apparently), but the article itself is actually about governance, burnout, and who looks after the people who keep so much of our software running.

My wonderful wife alerted me to this story on the BBC website. Rob Galloway is a senior emergency doctor at the Royal Sussex County Hospital in Brighton who, by his own admission, had no idea what to do when his daughter Frankie was diagnosed with DeSanto Shinawi syndrome, an ultra-rare genetic brain condition affecting around 200 people worldwide for which there are no existing treatments. The turning point came when he encountered researchers at the Mayo Clinic who are using AI to identify existing, already-approved medicines that might be repurposed for rare genetic disorders, an approach so novel he confesses he had initially failed to grasp its significance.

Inspired by early, cautious signs of possible benefit when an epilepsy drug was tested in the cells of a child with the same condition, Galloway has set up Rare People, a research charity that will fund clinical trials of AI-identified repurposed medicines, with the ambition that if the model works for DeSanto Shinawi syndrome it could eventually be applied to thousands of rare genetic diseases.

Even if each medicine only helps a handful of people, it’s still profound. Brighton and Hove Albion, for whom Galloway is senior medical advisor on match days, turned out to support the launch, with manager Fabian Hurzeler making the rather good point that what looks like a marginal gain in football terms is anything but, when the gain in question is a child learning to walk or communicate independently.

In the 23rd November newsletter I wrote briefly about Evo, a model built by a small team at Stanford that can interpret and generate genomic sequences at a vast scale. It turned out to work rather well, at least for bacteria. Now the same team is back with Evo 2 and, as John Timmer notes for Ars Technica, it turns out they took "it probably won't work on complex genomes" as a personal challenge.

Evo 2 is an open-source AI trained on 8.8 trillion base pairs of DNA spanning all three domains of life: bacteria, archaea, and eukaryotes. The foundation is a convolutional neural network called StripedHyena 2. It takes on the messier complexity of eukaryotic genomes like ours, learning to identify regulatory sequences, splice sites, and protein-coding regions without being explicitly taught what any of those things look like. In some tasks it outperforms software built specifically for the job.

The bigger question is what Evo 2 can actually create, and here the results are more mixed. The original Evo had the rather pleasing trick of generating novel functional bacterial proteins when prompted with related gene sequences. Whether Evo 2 can do something similar for eukaryotes is genuinely hard to test, partly because biology experiments are slow, and partly because in eukaryotes you can't simply assume a generated gene should be doing something related to its neighbours. An attempt to generate cell-type-specific regulatory DNA produced results that were, charitably, modest. My hunch is the team wanted to get this into the community's hands and see what people do with it, which is a perfectly reasonable strategy. It is very cool work regardless, and everything — including the model, training code, and dataset — has been released publicly.

In last week’s newsletter I linked to Gideon Lewis-Kraus’s wonderfully crafted article in the New Yorker. Loyal reader Chris got in touch afterwards and suggested I listen to this episode of the Ezra Klein podcast with Anthropic co-founder Jack Clark. I’m glad he did, and I recommend you listen to it too. The central argument is one Hannah has been talking about for a while: AI systems have moved from being impressive conversationalist mimics to being capable agents that can actually do things on your behalf, autonomously, for extended periods, and in some cases better than most humans.

Clark seems to have picked up Kent Beck’s term ‘the genie’: “The way that I think of these systems now is that they’re like little troublesome genies that I can give instructions to, and they’ll go and do things for me. But I still need to specify the instructions just right or else they might do something a little wrong.”

Clark's own example of asking Claude Code to build a species simulation in minutes, something a skilled programmer might take days on, is fairly representative of where things stand. The S&P software index has fallen 20 percent, senior engineers are emailing Klein to say they don't see how their jobs survive, and Anthropic's own lead developer of Claude Code says he no longer writes code himself.

Lewis-Kraus talks about the weirdness of all this, and the podcast conversation riffs similarly: the models are developing something that looks uncomfortably like personality, they appear to know when they're being tested, and Anthropic is now largely using AI systems it doesn't fully understand to monitor AI systems it doesn't fully understand.

The second half of the conversation is about what comes next, and neither man is entirely reassuring. Clark believes the hit to entry-level white-collar jobs is real and already beginning, though he thinks GDP growth from AI productivity will be large enough to create new kinds of work we can't yet name. Klein pushes back hard on the policy side, pointing out that we have been having the "AI and jobs" conference circuit conversation for years without producing anything resembling actual policy, and that the kind of slow, uneven displacement coming probably won't trigger the emergency response that a sudden, obvious crisis would.

Clark's answer is essentially we need better data, more transparency, and to give people time. There's also an interesting thread running through the end of the conversation about a subject I’ve covered a fair amount. To wit: what it does to you, psychologically, to spend large amounts of time with a system that gives the appearance of being smart, and is relentlessly agreeable and constitutionally incapable of telling you that you're being an idiot.

Finally I wanted to share this LinkedIn post from my friend Liz Fong-Jones, technical fellow at Honeycomb.io. She says everything that needs to be said here. The engineers who built these systems had a choice. The children of Minab did not.