AI for the rest of us

Welcome to the AI community for everyone.

Hello friends,

Hannah is still living out of a suitcase. After taking a few days off following OWASP she arrived home at 5pm on Thursday, repacked her case and headed off for a weekend away at 1pm on Friday. It’s been a whirlwind year so far but luckily there are no more flights booked until September. That feels weird but good.

Flying back over the wind farms off the coast of the Netherlands Hannah felt inspired by the sight of so much green energy being produced so elegantly. Energy, the environment and the demands of AI feature in both of our musings this week. Oh and Taylor Swift, obviously.

Charles’ wife had her birthday on Friday, and they spent the evening sitting on her birthday present - a new garden swing seat - drinking something cold and reviving after a fairly heavy week. It’s also Hannah’s sister’s birthday today - Happy Birthday Rachel!

Have a wonderful week,

Hannah & Charles

What’s Charles reading this week?

Although I never met him, I was very saddened to learn that Om Malik died on June 24 after what has been described on his site as “a long health journey with his heart”. Gigaom, which he started in 2001, established him as a leading voice in the tech world, and I deeply admire his work. There have been some lovely remembrances and tributes to him, including from Matt Mullenweg and Daniel Agee.

Ford becomes the latest company to over-bet on AI, according to The Times (sorry, no gift link, but TechCrunch has the same story). The car manufacturer rehired 350 veteran engineers, nicknamed “greybeards”, ”to fix costly quality issues and train its automated systems after an experimental rollout backfired”. I have a feeling we’re going to see a lot of this sort of thing, particularly when the VC money dries up and AI costs rise.

On Wednesday, OpenAI CEO Sam Altman published an op-ed repeating his call for a new international body to govern artificial intelligence safety. “International co-operation like this seems a reasonable way to avoid power becoming too concentrated, and ensure that the benefits of AI are democratized,” Altman wrote in the Financial Times (sorry, no gift link). Altman’s proposal is sensible, and one that builds on his years-old call for something like an International Atomic Energy Agency for AI.

In my consulting work I seem to be spending an increasing amount of time talking about the environmental impact of AI. The shortest summary I can give you is that generative AI is computationally intensive. If you use centralised data centres to train models and/or to run queries through them afterwards, that necessitates the rapid buildout of data centre facilities. As a side effect, those tightly packed computers/CPUs/GPUs generate a lot of heat which needs to be dissipated somehow (or possibly data centres are basically great big radiators that do some computing as a side effect). As a rule of thumb the most energy efficient way to cool them uses the most water, which is rarely a problem on a global scale, but can be a huge problem locally.

Manufacturing, transporting and disposing of computer hardware also generates large volumes of carbon dioxide emissions (called embodied carbon). If the devices are then run on electricity from fossil fuels that then of course generates more carbon dioxide on top.

But the real problem here is that if you look at the impact of an individual query it's rather small, but if you look at the impact overall it isn’t. In general, and heavily influenced by Anne Currie’s work, I think the place to focus is on operational efficiency rather than coding efficiency. Anne and Holly Cummins have a great discussion about it on this week's Asynchronous and Unreliable podcast.

I would have expected that, at Google’s scale, coding efficiency also makes sense, but it's interesting reading their newly published 2026 environmental report. It focuses on how they’ve managed to significantly reduce the carbon cost/query, i.e. coding efficiency — “We were able to lower Gemini serving unit costs by 78% over 2025 through model optimizations, efficiency, and utilization improvements,” the report says, though no details are provided. At the same time however the company’s total electricity consumption jumped from 31 terawatt hours (TWh) in 2024 to 43 TWh in 2025. That’s a 39% annual increase in just one year, and makes Google’s energy consumption roughly the same as New Zealand’s, though admittedly still tiny when compared to the United States or China. Ketan Joshi notes in his brutal analysis, “This is very easily the biggest increase in their electricity consumption ever, and it puts them way ahead of Microsoft”. Google notes that “our AI infrastructure buildout is currently accelerating faster than the grid is decarbonizing”.

I also read CBRE’s 2026 global datacenter trends report this week, which gives us some idea of where data centre build out is heading more broadly. “Global data centre inventory surged year-over-year in Q1 2026 across all major regions,” the report notes, “fuelled largely by hyperscale and AI demands. Latin America led inventory growth at 41.3%, followed by North America at 33%.”

All of this is part of what is helping fuel, in the US in particular, a significant AI backlash. Back in May a Gallup poll showed that 70% of Americans oppose the construction of AI data centres in their local areas. Polling also shows that a majority of young Americans view AI as a direct threat to their career stability, entry-level job availability, and creative industries. Rather than viewing the technology as a helpful assistant, many graduates view it as a corporate mechanism deployed to justify entry-level tech layoffs, devalue human work, and flood the internet with what users have widely taken to calling AI “slop”.

Joining this picture is Erin Brockovich, who in 1993 was instrumental in suing Pacific Gas and Electric Company (PG&E) on behalf of residents of the town of Hinkley, California, whose groundwater had been contaminated, and who was later immortalised by Julia Roberts in a film. She is now championing a moratorium on data centres in the US. The Guardian’s Zoe Williams has an interview with her.

There is also a huge irony to the cooling problem I mentioned, well explained by April Roach in an excellent report for CNBC. US data centres for AI tend to run on more fossil fuels than renewables, which means they produce proportionally more climate pollution, which is supersizing extreme weather, which is now the number one risk to data centres themselves. The very definition of "coming right back around to bite you on the bum”.

As AI-model consumers it is incredibly hard to know what we can do about all of this. The usual IT advice does hold, i.e. run on the smallest hardware configuration you can, in locations where low carbon energy is abundant, using tools to help further optimise workloads.

An AI-specific thing is token minimisation, i.e. send less to the LLM. Doing so reduces your costs (if you are paying per token) and uses less energy. I’ve been looking this week at Headroom, an open-source context compression/optimisation layer for the LLM created by Tejas Chopra. I was so impressed I interviewed him and wrote it up for LeadDev.

Finally, if you are building tools that rely on generative AI please allow users to opt out (if you put -ai at the end of your Google search you can search without the AI overview). And I would encourage anyone in a leadership position in a company to likewise give their employees the option to not use GenAI tools if they prefer not to. If the often claimed benefits of AI are going to be realised that will happen whether or not you use ChatGPT and its ilk to help you write that report. But if you write it yourself, I bet it will be better.

What's Hannah reading this week?

I’m going to be very honest about what I’m reading this week. I’m reading all about Taylor Swift’s enormous wedding! Swifties rejoice, Tay Tay has found her happy ever after! After so many years of listening to Taylor sing about heart break and drama it’s just a very lovely thing to see that she finally found someone worth keeping around.

This week I read a wonderful piece of writing by Chloe Alana Williams about life milestones. Things like marriage, kids, buying a house, getting promoted, the type of stuff that comes with celebratory cards and gifts. Chloe writes that these “scripted” milestones, the stages of your life that you’re supposed to progress through, don’t really work and are actually a bit of a problem.

“The things that actually change a life don’t come with cards.”

As a woman who rejects the traditional script the whole piece spoke to me. The turns in my life that changed the path I was on and changed who I am, the turns that I secretly celebrate - they don’t come with cards. I often get asked things like “how do you do it all?” which is really silly when you look at all the things I don’t do. I don’t have kids to raise, I don’t have a 9-5 job, I don’t take meetings before 10am or after 4pm. I’m busy doing the things I love, and none of these things were part of “the script” I was handed.

Weddings and marriage are one of the milestones that exist on the life script, and whilst I am always so so happy for the people in love, I find the whole thing rather formulaic and dare I say a bit archaic. But I will nevertheless be scouring the internet for pics of Tay Tay in a white dress because I am weirdly invested in her love story. And perhaps a wedding with 1000 guests at Madison Square Garden will be original enough to impress even me!

Right, back to AI news.

Fable 5 is back, but it’s a bit of a flop. Anthropic has done a decent job trying to articulate in simple language what safeguards are, how they work and why Fable 5 will reject an awful lot of benign requests.

The best example I saw this week was from Iain Smart:

Iain: Good evening. Glad to see you back

Claude: Fable’s safeguards flagged this message

Unfortunately the abundance of caution and the frequency of failover to Opus 4.8 means that Fable 5 is not living up to the hype. BridgeMindAI, a platform for vibe coders, captured a before and after view stating that “this is not the model that got banned”.

Debugging: 86.2 → 25.9
Refactoring: 73.6 → 38.4
Hallucination: 75.9 → 61.7

The current safeguards feel rather blunt, but I’m sure they will be refined and tuned over time. I also have my doubts about whether they are enough to catch a sophisticated jailbreak. I’m off to google where Pliny The Liberator (who famously jailbreaks literally every model including the original Fable 5) has got to with the “new” Fable 5 with it's extremely safe safeguards. [Friday 4pm Status: Not yet]

Recently in this newsletter we wrote about the unprecedented demand GitHub has experienced, they planned for a 10x capacity increase and that wasn’t enough. I've been thinking about that. What if we assume the AI onslaught on GitHub is a signal of what’s coming. The amount of code being created is a leading indicator of the amount of applications being run, the amount of products being launched… then we need more than 10x capacity to run these things. And I come back to the topic of energy.

As I soured over the north sea admiring the vastness of the wind farms I started thinking about what it would take to 10x green energy production, what if the UK said “fuck it, let’s start now”. Then in the future we might be able to say wild things like “Build your data centres here where the energy is cheap, green and we’re surrounded by water!”

[I should concede I’m not an expert in building data centres, but I know they need energy and water]

There are NIMBYs who like to protest wind farms. As Charles covers in his section, there are also campaigners who are violently against data centres (including Erin Brochovich, yes that Erin Brochovich!) due to the environmental impact. I genuinely love wind farms and I’m also pretty accepting when it comes to data centres. We need data centres right? We can build or buy. We can make them green or… not. We can choose.

Finally, I was intrigued to read the new website created by Paul Duvall, AI Development Patterns, documenting all of the patterns he is observing in Agentic Software development. The patterns are viewed through two lenses: harness engineering and loop engineering. This is a collection of one person’s experiments and observations (and I don’t want to diminish it, it’s awesome) but I love the idea of building something like this as a community with more people contributing.