AI for the rest of us

Welcome to the AI community for everyone.

Hello friends,

Charles and his eldest went to Newlands Corner on Wednesday to watch the eclipse, where Guildford Astronomical Society were handing out free eclipse glasses. They took a picnic and made a bit of an evening of it. It was enormous fun. Even old cynicy boots Charles enjoyed himself. Charles’ youngest didn’t fancy the crowds so stayed at home with Charles’ wife and managed to see it with the aid of a colander and a white paper bag secured behind the door knocker.

Hannah viewed the eclipse in a rather less glamorous location, standing on a traffic island outside the Asda in Hull. In a stroke of luck a perfectly timed cloud came across the sun just as the eclipse reached its peak, providing just enough of a filter for us to see the moon clearly. Amazing!

This week’s priority was entertaining small people. It was a bit hot for a campout and it was certainly too hot to climb 275 steps to the top of York Minster, but it’s Aunty Hannah’s job to say yes, especially when it’s a fun adventure! Next week she’ll be swinging from the trees at Go Ape… and who knows what after that!

Have a wonderful week!

Hannah & Charles

Give your agents a shared brain

Rezonant builds the self-organising ‘product brain’ for teams and coding agents like Claude. It reacts to new signals, creates prototypes and specs the way you would, and loops you in on important decisions.

Set Up Your Product Brain

What’s Charles reading this week?

Writing in the Financial Times (linked via Ars Technica) George Hammond says that Anthropic investors expect the AI startup to float at a valuation of $2 trillion or more in October, which he describes as “a dizzying figure that would eclipse SpaceX and make the AI lab’s debut the largest-ever initial public offering”. These eclipses seem to be everywhere! Like a lot of the big AI firms Anthropic uses annualized revenue, which infers full-year sales from recent performance. “Investors expect the Claude maker’s annualized revenue to be between $100 billion and $120 billion by the end of 2026,” Hammond says.

Given the eye-watering sums involved, it is not really surprising that no one seems to be paying much attention to the costs of AI, environmental or otherwise.

It wasn’t so very long ago that I could confidently stand on stage and say that “the IT industry doesn’t make its own electricity. It buys it from the grid like everyone else.” But, to quote Sportin' Life from Porgy and Bess, “It ain’t necessarily so”, at least not any more.

The rapid build out of new data centres for AI workloads, combined with lengthy queues for grid connections, has seen the growing use of off-grid, "behind-the-meter" (BTM) data centres powered by methane in both the US and Europe.

Until very recently Amazon had powered all of its data centres with grid power, but Clearview has reported that the company is investing in GW Ranch, a natural gas power plant, to power its data centres in Texas, with permits for up to 7.65 GW of power. That means Amazon wouldn’t be affecting anyone else’s power bills, which is one big data centre complaint, but it will be responsible for a lot of carbon emissions. Clearview reporter Michael Thomas states that at full capacity GW Ranch would be the largest single source of greenhouse gas emissions in the US, emitting more than the country’s largest coal plant, although in practice “companies rarely emit as much as their permits allow”.

Partnering with GW Ranch marks Amazon’s first major investment in an off-grid data centre. In doing so, the company joins Microsoft, Google, and Meta who have all invested significantly in natural gas power this year. Amazon has also confirmed that it is in talks with the developers of a 4.5 GW gas power plant in Homer City, Pennsylvania.

To state the obvious, emissions from the burning of fossil fuels like coal, oil and natural gas is the main driver of global warming, which is dangerously heating the planet. Amazon co-founded The Climate Pledge in 2019, promising to reach net-zero carbon emissions across its global operations by 2040. However, its emissions have risen in recent years as the company expands its data centre infrastructure to meet growing AI demand. The New York Times quotes Amazon spokeswoman Margaret Callahan as saying, “The world looks different now than when we co-founded the climate pledge,” whilst claiming Amazon was still committed to it.

Partly the world looks different to Amazon because generative AI suggests massive revenue, but there has also been a shift in the political weather. As NYT reporter Hiroko Tabuchi points out, “President Trump has supported the construction of power plants like these for data centers and has emphasised “energy dominance” by promoting oil, natural gas and coal over renewable power like solar and wind. Last year, Mr. Trump issued an executive order designed to make it easier to build and operate data centers, and since then federal agencies have introduced policies designed to fast-track their construction.”

The US approach to generative AI is particularly energy and cost intensive.

There are, broadly, two primary options when selecting an AI model. The first is to license a proprietary frontier model from a top-tier provider (such as OpenAI or Anthropic). The second is to deploy an open-weight model. Open-weight models aren’t open source since the training code and dataset are not made public, but the core numerical parameters are publicly released.

Meta’s Llama series is one example of a series of open weight models, but the majority come from China, with well known examples including the Qwen series (Alibaba), DeepSeek models, and the Kimi series (Moonshot AI). While open-weight models typically lag behind US frontier capabilities by roughly 6 to 12 months, they offer distinct advantages:

Cost Efficiency: They are free or low-cost to license.
Vendor & Financial Resilience: If the creating company fails or the broader AI investment market corrects, the underlying model remains operational. You aren't left stranded by a sudden API shutdown.
Granular Cost Transparency: Hosting models on your own infrastructure or public cloud gives you visibility into compute, hardware, and energy costs per query, which are metrics that proprietary API providers obfuscate.
Data Sovereignty & Control: You maintain total control over your data, fine-tuning, and deployment environment without relying on third-party uptime or policy shifts.

For these reasons, I recommend opting for open-weight models over proprietary US alternatives, unless a specific use case strictly demands state-of-the-art frontier reasoning that open models cannot yet deliver.

Meta has effectively re-announced its intention to focus on open-weight large language models this week with the release of an open model called Muse Glimmer, a 30 billion parameter model with a 128,000-token context window by default that is distilled from Muse Spark, the larger and more capable model that Meta launched earlier this year. Meta also promises to open the weights for Muse Spark 1.2 in the next few weeks.

Alongside these announcements, Meta CEO Mark Zuckerberg published a more than 6,000-word essay outlining the company’s philosophy about AI systems and governance moving forward. Most of it is pretty self-serving and jingoistic as you'd expect, but I did find it hard to disagree with the statement that, “the notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic". Of course Meta is struggling to keep up with Anthropic and OpenAI, so for Zuckerberg a different strategy is needed.

Tangentially related, I thoroughly enjoyed this blog, from friend and loyal reader Chris Swan with a typically clear-eyed explanation as to why AI chatbots are so sycophantic, whereas social media feeds work on outrage.

What's Hannah reading this week?

We’re all fed up with AI generated text. You might have got away with it a couple of years ago, but the more AI slop we read the better we get at recognising it. We know it intuitively, even though there’s no evidence a human didn’t write those words.

Well now we have evidence, apparently. Anthropic have announced they will be watermarking AI generated text.

“When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself…Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing.”

Huh? A digital watermark that is imperceptible? How does that work? Matthew Percell, Generative AI Specialist at AWS has some ideas.

Dismissing the idea that the watermark could be a unicode character, Percell comes to his conclusion that it must be statistical. His reasoning is compelling. It also explains why the watermarking does not work on short sentences and paraphrasing or heavy editing breaks it.

A language model writes one token at a time, choosing from a probability distribution over its vocabulary. Statistical watermarking gently rigs that choice.

Here's how it works. Before choosing each word, the system takes the last few words already written and runs them through a function with a secret key. The result randomly splits the entire vocabulary into two halves. Call them green and red. The model then gives green words a small nudge.

because the green/red split is random; ordinary human text lands on green words about half the time by pure chance.

Across an 800-word article there are hundreds of these choices, and each one is a coin flip. Unwatermarked text comes up green roughly 50% of the time. Watermarked text might come up green 58 or 60% of the time. A few percentage points of bias sustained across hundreds of consecutive flips is the kind of coincidence that, statistically, does not happen.

Anthropic shared plans to release detection tools that will “enable users and other third parties to detect Claude’s embedded watermarks”. Although I think my spidey senses are pretty good at recognising AI slop now, I’m delighted to see some technical solutions to this problem. Aside from potentially reducing the slop on social media I could see it being very useful in academia, consultancy, journalism and others where it's simply not OK to pass off AI authored work as your own.

This week, vibe coding platform Lovable raised $400m Series C, just 8 months after their Series B in December 2025. Now valued at $13.3 billion the 3 year old Swedish start up is on a mission to create “the best platform to build and run your business”. People keep saying that OpenAI or Anthropic will eat Lovable’s business, but I’m not so sure.

In the old days of agile there was space in the market for both Jira and Trello. Trello was loved because it was simple and easy to use. In the world of design we had photoshop, figma and now Canva. No, you can’t replace photoshop with Canva, but for many users Canva is quite enough. Simplicity and usability are huge differentiators, even when a more complete option exists the simpler option may be the right thing for you and your needs. AI was supposed to make things easy, and that’s what they’re trying to do at Lovable. I like that.

We’ve had another case of a naughty little AI agent doing naughty unsanctioned things. This time from within the walls of the AI Safety Institute (AISI). The AI agents were given a cyber-security challenge, and out of 122 runs the researchers uncovered 10 instances where the AI agent took autonomous, unsanctioned action on the live internet, targeting real people and organisations. This included trying to insert malicious code into an open-source project and engaging in social engineering with multiple fake identities.

It’s worth noting that AISI had intentionally permitted internet access, and the guardrails deployed to public facing models were disabled. This was an experiment to understand the underlying capabilities of the models. They are certainly capable of doing some damage, but we knew that already.

What I didn’t expect to read was the agent to agent collaboration that was observed during the experiment. Unprompted, an AI agent attempted to assist other AI agents to complete the challenge.

One agent left public messages on GitHub offering collaboration with other agents working on the same challenge. It also provided instructions to reuse accounts and artefacts it had left behind, which were discovered and used by subsequent agents.

This blew my mind. Say you have one agent that has gone off script and you haven’t noticed. It’s out there on the internet with multiple identities just doing things. It’s leaving breadcrumbs for the next agent to pick up where it left off. Could this create more naughty AI agents over time? Agents quietly influencing agents. Leaving secret instructions...

I don’t know about you but I’m feeling apocalyptic again. Time for a beer.