Weekly Links: Weight Space Programming, SaaS Moats, and US Data centres

A rather late post this week: apologies.

This week in AI: OpenAI claimed it hit 20% of Google's search volume, Airbnb CEO and others say personal agents are not a meaningful share of traffic (like the Internet had no impact on commerce in 1997), and smart rings with joysticks.

On to the main stories:

• Anthropic rewrote its Messages API in Rust. This is a head-scratcher... on the one hand, Anthropic is clearly showing off Claude's capabilities (and perhaps trying to establish a pattern others can follow). On the other hand, this feeds two very negative narratives: developers with specialist skills are no longer needed, and Anthropic may be careless about its implementations. We will clearly reach a point where much of the world's critical infrastructure is written in languages humans no longer understand. It doesn't make much sense to me to hasten that moment within your own company.
• LLMs can natively read & write neural network weights. I guess this was always obviously true, but now, it seems, they can do it successfully. The key quote from the summary is "Just by looking at the raw weights, GPT-6 Astra can tell what a network was trained to do. And without any environment feedback, it has no trouble writing, from scratch, small game-playing networks that outplay networks trained with PPO for 100 million steps." I'm not sure how practically useful this will be yet, but intuitively it is a little like gene-editing, or event-synthesising genetic makeup from scratch to get certain effects. If it worked, it would be way faster than training a network. I just wonder if, at some level, it isn't also another form of distillation.
• Google says there’s no way to use Search without using artificial intelligence. In essence, Google is reacting to a New York City-wide ban on AI in schools, but saying search also won't be available since it is inextricably interwoven with AI. It's unclear whether they are arguing that AI is used to deliver search results (it is, but this has existed for years and clearly isn't the intent of the ban), or whether the new Google AI experience within search is at issue. It's hard to imagine any engineer believing it would be hard to hide the Google search experience (a Chrome plugin could probably do it). It's hard to see this as anything other than a bad-faith shill for AI where it has not been accepted.
• The Only SaaS Moats Left in 2026 [Video]. Rob Walling's MicroConf talk from a few weeks ago covers some exciting ground and offers a nice overview of how to think about SaaS value going forward. I agree with the view that not all SaaS is dead and also with his suggestions on where to find the anchors which ensure it remains valuable. The extension I would make is that these exact same challenges and fixes also apply to new generations of AI-powered products. AI-powered SaaS is still SaaS.
• Datacentres were on the ballot. US voters rejected them. I'm not sure this is the read I would put on the result. Both Republicans and Democrats campaigned against data centres, and many anti-data-centre motions passed. I'd say, though, that AI was on the ballot, not data centres (they are just the manifestation), as well as perhaps the aggressive way in which large tech companies have sought subsidies and tax breaks to bring their data centres to different states. This approach clearly has to flip, and data centres will likely need to promise significant incentives to local communities. Still, data centres in general aren't bad neighbours (try a lumber yard or a truck logistics depot); it seems to me the real backlash is against the job-destroying fear data centres induce. This delay may certainly give other parts of the world a head start in building out compute. Ultimately, I guess it will be solved with promises of more financial payments to local communities.

Wishing you a great Sunday evening.