AI for the rest of us

Welcome to the AI community for everyone.

Hello friends,

Hannah is writing this week’s newsletter from Florence where she’s been absolutely delighted to discover their famous / infamous Wine Windows! Tiny little hatches scattered throughout the city through which you can purchase a glass of wine to take away. They date back to 1559 when they were originally installed as a way for families to sell their wine without opening a shop and incurring taxes, but made a comeback during the pandemic in 2020.

Picture the scene: you open the tiny door, ring the tiny bell… “Due vino rosso della casa per favore”... a hand with a card machine emerges somewhat breaking the medieval vibe… then two lovely glasses of local Chianti emerge from the hole in the wall. She drank hers on one of Florence's many bridges as the sun was setting. Bellissimo!

Hannah is making her way through Italy towards Milan where she will be MC for the Business Track at Platmosphere, a conference for Platform Engineers and Engineering Leaders. The theme for this year is “Mastering the Vibe” and there is sure to be a lot of discussion about AI. You can expect to hear all about it in next week’s newsletter!

Charles has had a cultural week, attending a lecture with his eldest at The British Library on Monday on “Self-Help from the Middle Ages” from historian Peter Jones, and two lectures from Pint of Science in Guildford on Tuesday. If you are not familiar with it, Pint of Science is an annual festival that brings science talks into pubs and bars, making research accessible and informal. It’s awesome. His general observation from all three lectures is that there is something wonderful about hearing people who are both deeply knowledgeable and have a profound love for their subject getting up on a stage and talking about it.

Don’t forget that we’re hosting a seriously awesome one day event during London Tech Week. Agent Craft is an unconference for Agent Builders and Agentic Engineering. We’ll be going beyond the basics of AI Agents to talk about the real world failure modes and risks that come with this incredible technology. Tickets start at £96 and we expect to sell out.

Have a great week!

Hannah & Charles

The AI Summit London

We're excited to share that we're now an official partner and fringe event of The AI Summit London 10-11 June, during London Tech Week. The team at The AI Summit have shared this exclusive 20% discount code for our community: ARU-20DEL

What’s Charles reading this week?

Andrej Karpathy has joined Anthropic. He co-founded OpenAI in 2015, was director of AI at Tesla (reporting directly to Elon Musk) from 2017–2022, went back to OpenAI in 2023, and then left again in 2024 to start an AI education company named Eureka Labs. He also coined the term “vibe coding” in February last year, but we probably shouldn’t hold that against him.

Anthropic and OpenAI are both rumoured to be getting ready to IPO. Anthropic may also be about to turn profitable, according to The Information (sorry, no gift link). Sri Muppidi writes that the firm “expects to generate $559 million in operating profit at the end of the June quarter, becoming profitable on an operating basis for the first time, according to a person with knowledge of the financials”. So now it can share its profits with all the creators whose IP it stole. We eagerly await the world's most overdue royalty statement.

The Guardian’s Aisha Down has a story that made me pause: more than 100 new UK datacentres are planning to burn gas to generate their own electricity. Some of them permanently.

This isn't a scandal, as such. It's more of a policy collision. The Government wants Britain to be an AI superpower. It also wants to hit clean power targets by 2030, with unabated gas (i.e. no attempt to capture, store or otherwise abate emissions) supplying less than 5% of electricity. These two ambitions are now visibly in tension, and nobody seems entirely sure what to do about it. The UK has an enormous queue of datacentre projects largely driven by AI demand and all waiting to connect to the National Grid. The queue is so long, and moves so slowly, that developers are giving up waiting and reaching for gas instead. Over 15 terawatt hours of gas demand annually, if all those requests materialise. Enough to power London for roughly four and a half months. The grid can't be upgraded by wishful thinking, and "we'll sort the emissions out later" is doing a lot of heavy lifting in current policy.

I’m slightly unsure on this, but I suspect the climate picture is worse than the headline figures suggest. Natural gas is mostly methane, and methane that escapes unburned — through leaks in pipelines or imperfect combustion in generators — is around 80 times more potent as a greenhouse gas than CO2 over a 20-year period. Emissions figures based purely on what gets burned will, I think, tend to understate the true impact.

For context, as long-term subscribers to this newsletter already know, in the US this is already well advanced. Datacentres built for companies including Microsoft, Meta and OpenAI are projected to collectively emit more carbon than the entire country of Morocco. Elon Musk's xAI has been running dozens of methane-powered generators without proper air permits, within a few kilometres of six schools. Datacentres already consume around 6% of electricity supply in both the UK and the US. That number is heading in one direction.

Back in April I shared this terrific essay and Decoder podcast from Nilay Patel, chief editor at The Verge, in which he explores the disconnect between the software industry’s view of AI and the vociferous public opposition to it. We can now see this in real life. Earlier this month, Gloria Caulfield, a vice president at real estate company Tavistock Development, took the stage at the University of Central Florida to deliver an uplifting message to soon-to-be grads. Then she said, “The rise of artificial intelligence is the next Industrial Revolution,” and the students, many of whom had studied journalism, writing and filmmaking, began to boo. A few days later, when former Google CEO Eric Schmidt took the commencement stage at the University of Arizona, a similar thing happened.

Loyal reader Sophie emailed me to ask what I made of this and I replied, perhaps a bit glibbly, that it turns out that “we stole everyone’s work and burned a small planet to build a glitchy plagiarism engine in order to kill your job prospects and line the pockets of a few tech billionaires” didn’t land well with the public. Whoever would have guessed? Joanna Sterne however called Houda Eletr, a recent UCF journalism grad and aspiring poet, to find out exactly what she was booing. Being too old for TikTok I wish the video was longer, but it is worth a watch.

In my latest article for LeadDev, I take a look at how AI is increasing asymmetry in open source. If we don't start valuing stewardship over sheer output, we're going to drown in our own automated technical debt.

Meta Platforms had said last month it would lay off about 8,000 people on Wednesday, roughly 10% of its 78,000 employees, as it embarks on its latest effort to cut costs and offset the heavy investment in artificial intelligence. It has now begun notifying thousands of employees that they are being laid off, The Information’s Jyotii Mann reports.

The Musk vs OpenAI trial ended after a nine-person jury unanimously decided on Monday that Elon Musk took too long to file his lawsuit that accused OpenAI of stealing a charity. Musk filed his suit against the $730 billion artificial intelligence start-up in the summer of 2024, but the jury found that he was aware of the behaviour discussed in his complaint as far back as 2021. Musk plans to appeal. I think the point of this lawsuit is just to make Altman’s life as unpleasant as possible and distract the team at OpenAI as they get ready for an IPO, rather than actually trying to win. This update quoting Judge Yvonne Gonzalez Rogers’s “poetic” jury instructions is just lovely:

“A jury reflects the attitudes and mores of the community from which it is drawn,” she said, paraphrasing another judge. “It lives only for the day and does justice according to its limits. The group of jurors who are drawn to hear a case make a decision and then melt away. It is not present the next day to be criticised. It is the one governmental agency that has no ambition. It is as human as the people who make it up.”

What's Hannah reading this week?

At BIMP we’ve been using Cloudflare as part of our hosted solution and so I’ve recently taken more of an interest in their writing. This week Grant Bourzikas, Chief Security Officer from Cloudflare shared their learning as participants of Project Glasswing, gaining early access to Anthropic’s Mythos model.

I appreciated Grant’s explanation of how Mythos works differently to other frontier models in the way it identifies, tests, exploits and chains together small bugs into a much more severe exploit. This is why everyone is so worried about Mythos. How many security teams currently spend their time triaging and prioritising vulnerabilities? That low severity issue may not be as harmless as you think.

One of the most interesting take-aways from his post is how difficult it is to leverage these frontier models for security research on a large, complex code base. You need a harness that is designed for this type of very focused explorative work. Simply pointing a model, even Mythos, at your code and asking it to find issues won’t do the job.

[Since writing this newsletter Anthropic have also published an update on Project Glasswing. The headline seems to be that Mythos has found over 10,000 critical and severe vulnerabilities already and we can expect that to continue]

Cloudflare was in the news last week when their share price took a hit after Anthropic announced a $1.8bn contract with Akamai, their main competitor. That’s BILLION not million. That’s a crazy number and Akamai’s share price is soaring as a result. If you have shares in Akamai, well done. Sadly I do not.

We need to talk about burnout. Some recent research by Obsolete.com found that 92% of AI time savings are converted back into more work. AI does not reduce work, it intensifies it. There’s also a new phenomenon called “AI Brain Fry” whereby people experience acute cognitive overload from monitoring, evaluating and correcting AI output. I’ve requested the full report and I’ll share it here when I receive it but you can review the summary on LinkedIn.

Bryan Ross, Field CTO at GitLab writes about the unsettling feeling many people are experiencing when the “doing” of our jobs is being “done” by AI. In his post “The Identity Crisis AI Didn't Warn You About” he talks about how exhausting it is to work in IT right now.

Jamie Dobson shared the same question this week on LinkedIn asking, “How burnt out are we all?” Is the speed that AI brings actually setting unrealistic expectations about individual output? Are all the layoffs taking their toll on the people left behind? Are we feeling the pressure to do even more because it’s so quick and easy to slop up a first draft? Are managers and leaders even interested in measuring real productivity, or just volume?

“Manager” can be a bit of a dirty word. They are the baddies, right? They always want you to do more with less, those bastards. I actually LOVE being a manager and one of the things I like to do in my teams is provide focus. Focus is finite, we only have so much head space and we must use it on the most important work. Managers who don’t provide focus can create an environment where everyone is busy and nothing is getting done. AI mandates and workslop can have a negative impact if they distract us from what’s actually important and I wonder how many teams are trapped in that place right now, producing a lot of work and not getting anything done?

One of the sessions I’ll be hosting as part of Agent Craft on 12th June will be about the role of Humans in the Agentic Org. Where can we safely offload work to agents and where do we think humans will always have a role to play. What skills do we need for this new world and how do we make sure the work that’s left is fulfilling and not exhausting?

Finally, if you're a regular at our meetups you may have met Paul Markham who is one of the most active members of the community. Paul is hosting a webinar on June 2nd that explores real time language translation and is looking for multi-lingual folks to try it out. Connect with Paul on LinkedIn if you're interested in playing with this exciting new AI tech!