A creationist's soul-searching model makers

Weeknotes 367 A shorter newsletter this week, but with some thoughts about models doing soul searching. And a rough overview of the news of last week.

An interpretation by Midjourney

Dear subscriber, dear reader,

This week, the newsletter is a bit different than usual because —bluntly—of a lack of time. As you might know, we have the TH/NGS conference on this Friday. As always, the final weeks bring a lot of work—apparently.

Week 367: A creationists-mode of model makers

This means that I am capping my time for this newsletter to less than usual. Practically, I drafted quickly and leaned on my “helpers”—AI tools—to rewrite my first thoughts. Normally, it takes a couple of iterations, often ending up with me writing it myself, using only Lex's tips (lex.page that is). So this edition is a bit more AI-shaped and a bit less me—maybe nice to see sometimes.

This week’s triggered thought

Thoughts about Anthropic's leaked "Soul" document—a set of narrative-style instructions describing how their model should behave. It's more than a system card. System cards set defaults. The Soul doc goes deeper: it's about upbringing.

Call it "god-mode" learning. You give the model impulses, let it learn, and anchor it in a value reference. You're not just configuring—you're raising. Values are baked in at the moment of creation.

This is a creationist design. Anthropic defines the good, encodes it, and the model inherits it—like a child inherits a worldview from parents who believe they know the right way to live.

But is this the right approach?

The creationist model assumes a lot: that Anthropic knows what's good, that the good is singular, that values can be fixed upstream and still hold downstream in messy, contested contexts. It assumes a single creator and a model that follows.

There's an alternative: evolutionary design. Values emerge through selection pressure—what works in context, what users need, what situations demand. Fitness isn't about matching an ideal; it's about adapting. Values are bottom-up, emergent, and plural.

Notions from last week’s news

Also, a bit tighter in this part. Normally, that takes a few passes through Mail, RSS, and Reader: rough selection, then reflection. This week, that was too much, so I collected items and simply linked them below. I specifically kept my selection to the articles on my only reading list out of curiosity.

I listened to a bit fewer podcasts, but still from Hard Fork to Tech Won’t Save Us to Dithering, the AI Daily Brief, Nate B. Johnson Notebook, Pivot, and Sharp Tech. You can fill a week with all of these…

One returning topic: model competition—“best model,” “code red” at OpenAI—drew a lot of attention. It is a valley-dance. OpenAI feels pressure from Gemini; days later, it claims that Anthropic’s new model is best for everyone. That would make you nervous if your business runs on keeping promises about delivery. Otherwise, you’re “cooked”—it is this week’s meme.

Is “smartest” the goal? If survival of the fittest applies to models, “fittest” = best fit to context—not raw performance. Not the fastest runner, but the most situated. This shift may deflate AGI claims in the coming year—perhaps acceleration follows. The Overton window for AI is shifting; what counts as AGI is a moving target.

Ok, on with some captures from the news, a set of items I’m curious to read more about.

Human-AI partnerships

Robotic performances

Immersive connectedness

Tech societies

Misc

What’s up for the coming week?

ThingsCon will keep me busy this week, and I’m kicking off my Cities of Things research project this week. That will be more intense over the coming weeks and months. More on that in the next newsletter. What I can say is that it centers on what has happened in the context of Cities of Things. Comparing the dilemmas we laid out back in 2018, and what we learned in the activities over the years. What is next?

Interestingly enough, just this afternoon in a session on a new educational program at Rotterdam University of Applied Sciences on Learning Communities Slimme en Sociale Stad, we discussed the themes for the topic of Clever objects & environments. The dilemmas we described in the paper are now coming into the practice of city planning, according the representative of City or Rotterdam.

That’s my week for now. Hope to see you at TH/NGS!

Have a great week!

About me

I'm an independent researcher through co-design, curator, and “critical creative”, working on human-AI-things relationships. You can contact me if you'd like to unravel the impact and opportunities through research, co-design, speculative workshops, curate communities, and more.

Currently working on: Cities of Things, ThingsCon, Civic Protocol Economies.