Hello friends,
Whilst Hannah was at Øredev, Charles had the extraordinary privilege of representing her at the OpenUK awards in The House of Commons—a full black tie number. Hannah was nominated for AI for the rest of us, alongside David Agbolade, David Sorria Parra for Model Context Protocol, and Peter Steinberger for OpenClaw. OpenClaw quite rightly won in our category, but we were extremely proud to be nominated.
Although the FOMO about the OpenUK awards was intense Hannah had a wonderful time at Øredev in Malmö. A truly unique tech conference. Never before has a speaker’s dinner involved getting naked and jumping in the sea, but we did, and it was awesome!
The call for proposals for our 2-day conference in February has been extended by 2 more weeks. Let’s be honest, with Hannah travelling to Sweden this week and then the US next week there wasn’t going to be time to review the submissions anyway! So if you missed the original deadline, never fear, you can submit your proposal now!
Have a wonderful week,
Hannah and Charles
|
|
Early Bird Tickets
Don't miss out on the £150 super early bird tickets for our two day extravaganza next February!
|
What’s Charles reading this week?
This week’s buzzy but very technical blog shows Frontier LLMs (in this case “GPT-6 Astra”) can directly read and write neural network weights (i.e. the huge lists of numbers that determine how an AI behaves). The authors, four PhD students from MIT’s Computer Science & Artificial Intelligence Lab, call this Weight-Space Programming, and say it “may extend beyond what we’d expect from humans or conventional algorithms”.
The blog describes a series of eight experiments. Amongst them the researchers showed GPT-6 Astra a list of raw weights from a small game-playing AI and asked what game it had been trained on. The AI guessed correctly on all four games, including one (“Freeway”) that was pretty obscure. Those results aren’t necessarily restricted to frontier models either. Simon Willison has managed to recreate this experiment using the much smaller open weight Qwen3.8-27B model on his laptop. It got three of the four games right (it couldn’t identify Freeway but then, to be fair, I couldn’t either and I could watch the gameplay clip).
GPT-6 Astra also wrote the numbers for new game-playing AIs from scratch, and some beat AIs trained the normal way. But the headline grabbing result is when the researchers asked GPT-6 Astra to edit Qwen3.5-0.8B weights to make it better at maths. Astra changed only 0.016 percent of the model's weights, reducing its reasoning loops and improving its math performance considerably: by as much as sixfold on the MATH-500 benchmark.
This does sound remarkable, although there has been some scepticism. Matthew Honnibal, for example, thinks the blog is misleading, quite rightly pointing out that it mixes together very different experiments. In some, the AI really did just look at raw numbers. In others, including the headline maths result, it was allowed to run code and test the model. “Human ML engineers write snippets of code that analyse or even modify network weights ‘directly’ all the time,” he says. Similarly, Sean Goedecke argues that this is all stuff a smart human ML researcher could do. I’m not totally sure what I make of it all just yet: watch this space I guess.
Away from research, The New York Times has reported (also covered by The Guardian without a paywall) that early next year Google will turn off artificial intelligence tools in its popular software for K-12 students (ie Kindergarten all the way to age 17-18), unless school administrators explicitly choose to make such tools available. But it will also disable search. “Search and A.I. are deeply integrated,” Chris Phillips, vice president for education told The New York Times. If a school has banned A.I., he added, students “will not have access to search, and they’ll see a message from their school that indicates it’s not available”.
New York City’s mayor, Zohran Mamdani, who recently imposed a one-year moratorium on generative AI tools in city schools up to and including eighth grade (typically for students 13 to 14 years old), said Google was exploiting its near-total control of online search.
“Google is trying to use its monopoly power to force AI into our public schools,” Mamdani
posted on the dreaded X. “No corporation should have the power to dictate how we teach our children. Our public schools exist to serve students, not to grow the profit margins of Big Tech.”
Also in the US, the midterm results seem likely to slow down the data centre build out. The Guardian’s Johana Bhuiyan reports that voters approved nearly every data centre ban or moratorium on the ballot (dozens across at least 10 states, including 14+ in Ohio), with Harwood, North Dakota as a notable exception. Newly elected governors in Texas, Georgia and Ohio have promised construction moratoria, and analyst Dan Ives expects the AI build out to slow, with Texas the biggest worry for tech investors. I wouldn't be surprised to see more protests in Europe too.
Reddit has announced that it will limit most of its use to logged-in accounts. This is likely at least in part to limit access to AI scrapers and might also create a new business opportunity since the data could be valuable.
“Moving to primarily logged-in will ensure our users have the best Reddit experience possible, and can find the content and communities most relevant to them. It is better for our business because we will have more signals for ads targeting, and being logged-in leads to higher engagement generally. It also allows us to protect access to Reddit’s extremely valuable corpus of data, and helps us better adapt to emerging regulations around the world,” the company says.
And finally from me the Dutch government says that it plans to cut 25,000 jobs with AI automation, saving €3.5bn. For context, the number of central government officials hit an all-time high of 160,000 by the close of 2025, a 41% rise since 2018, according to Bloomberg.
What's Hannah reading this week?
Sometimes I feel like I know more about US politics than I do about politics here in the UK. The results of the midterms in the US this week certainly feel like a good news story for many of us outside of the US, despite the result triggering a panicked sell-off of US tech stocks, with ripple effects throughout the global stock market.
The BBC described it as “fear of a US AI slow down” which feels rather contradictory just a few weeks after Anthropic CEO Dario Amodei called for a “pacing of the frontier” and an intentional slow down. Do we want to slow down or not?
Something tells me the AI giants will continue just as before, regardless of the elected leader in the US.
In Sweden this week the conversation about digital sovereignty and AI sovereignty felt a little less urgent following the midterm results, given the likely change in leadership across the Atlantic. That doesn’t mean it’s not important, Trump’s shenanigans with Anthropic showed the world how dependent we have become on US tech giants and even without a chaotic leader that dependence remains a risk.
I was reminded of an awesome talk by Michael Shanks, CEO and Co-Founder of Budibase at State of Open Con in Belfast back in September. Although the talk “The Cost of Sovereignty” wasn’t recorded I have permission to share the slides. They are worth a quick flick through if you’re new to the topic. The key message was that we have different levels of dependence at each layer of the technology stack, and that the cost and practicality of independence varies dramatically. It’s not a simple yes / no, it’s about risk, reward and cost.
Amanda Brock, CEO of OpenUK commented on the impact the Democrat’s win in the US will have on UK AI policy and the continued investment into open source, not proprietary, AI.
The UK has, to date, taken a non-isolationist view of sovereignty, very different from the EU’s. It recognises its long-standing “special relationship” with the US and its role between the US, China and the EU…
Trump’s protectionist approach towards US closed AI, demonstrated by the recent DeepSeek ban in US federal agencies, and its implications for open weights and models, will be back on the agenda.
For the Democrats this presents an opportunity to push for greater competition, consumer protection and diminishing the concentration of power among a handful of technology companies.
Whether this translates into meaningful regulation, rather than political opposition, remains to be seen.
For open source and AI it’s an opportunity!
At Øredev this week I delivered an updated version of my talk “The Reinvention of the Dev Team” where I discuss the new challenges teams are experiencing with AI-accelerated software development. Even Gartner is now talking about tiny teams. But, in the hallway track afterwards the resounding feedback wasn’t about tiny teams, it was about big organisations and how this shift is playing out for them.
Cross-org coordination is the pain point that keeps on hurting, no matter how fast you can code. Coordination and communication still happen at a human pace and I can’t come up with a novel way that AI can help with that. We need to eliminate that work as much as possible through our org design and architectural decisions. Read Team Topologies remains the best advice I have on that topic.
Just before I went on stage, Clive Thompson did a wonderful job of articulating the changing work of software developers. The satisfaction and the “buzz” of the struggle and triumph is largely gone. The stereotype of the antisocial developer who doesn’t like going to meetings might be gone with it too.
I agree with Clive that many developers are mourning the loss of a craft they enjoyed, but perhaps, like pilots who are required to fly manually for a minimum number of hours, we need to make sure we can still fly our plane without the autopilot, and writing code manually might remain an important and necessary part of the job.
The topic of our next meetup in London will be “Software Craftsman For The AI Age” and I’m looking forward to a lively debate about where the industry is going. Some people want the dark factory, others worry we’ll lose our ability to comprehend our own systems. These aren’t mutually exclusive goals and right now a lot of teams are figuring out how to achieve both, and do it well. We're excited to have Meri Williams - The Geek Manager - opening the discussion with their thoughts on the subject.
Another common thread that ran through many of the talks this week was about decision fatigue and context switching. I loved this blog post about the challenge of context switching: Context switching in AI development, or the twenty minutes between knowing and having by Fernando Alvarez.
"I have nothing to do until the agent catches up. That's a context switch waiting to happen, and how I handle it decides whether AI makes my day faster or just more fragmented."
Fernando is has captured some really awesome, good advice, that most of us know but probably don’t follow. He proposes that one deep thread of work is enough, and any additional threads you decide to parallelise should be limited to smaller, low risk work items. The context switch costs less when the work is well understood and requires less engagement and review.
He also shared that feeling we get when we try to juggle too much.
“A day spent bouncing between three agents and a chat window doesn't feel slower. It feels frantic, and the frantic feeling is the bill.”
I’ve always liked Kanban more than Scrum but this feels like a strong argument for minimising work in progress. Frantic isn’t sustainable. Exhaustion isn’t sustainable.
Finally, I wanted to shout out to the team at Stacklok who, instead of tokenmaxxing, have managed to reduce their token consumption by 42%. If you want to learn how they’ve shared their learning on their blog. Thanks for being role models!
Updates
|
|
Early Bird Tickets
Our next 2 day conference is set for 25-26 February 2027. Grab your early bird ticket for £150.
|
|
|
Nov Meetup
Our November Meetup is scheduled for Tuesday 24th. We'll be talking about software craftsmanship.
|
|
|
Between Jobs?
Interested in a more human approach to career change and AI? Join the Surviving The Gap community for practical workshops, useful resources and early access to Career Core.
|
|
|
Follow us on LinkedIn
Stay up to date on the latest news from the community!
|
|
|
Follow us on BlueSky
Stay up to date on the latest news from the community!
|
|
|
Agent Craft Talks
Agent Craft talks are now live!
|
