0.0 Context Setting

Tuesday, 21 July 2026 in Portland, Oregon where it’s been hot. There’s been a Fifa Men’s World Cup Football Final which in my humble opinion went to the right team.

0.1 How People Work, Live! ... with Pavel Samsonov

How People Work, Live! is a live version of my workshop, How People Work. I get to invite an interesting guest on and we talk about the hard things to do with the soft skills you need to not just get jobs done, but to do them well. You know, the soft skills like “apparently the quality of my relationship with coworkers appears to matter” and “what do you mean I need to be able to influence people”, along with “no I don’t have time to make this meeting an email” and its sibling “no I didn’t realize I should’ve made this email a meeting”.

Whether you’re a designer, developer, product manager, strategist, researcher, or a mythical 100x unicorn individual contributor, you’ll always need to work with other people. (Even if you’re solo. Other people are going to be using your stuff.)

You can register now for How People Work, Live! ... with Pavel Samsonov, on Friday 24 July at 9am Pacific on Zoom.

1.0 Some Things That Caught My Attention

I am going to try and do some small, short Things this time.

1.1 For All Humanity?

Bruno Dias on Bluesky:

Either LLM model weights aren't copyrightable, in which case distilling an LLM cannot be copyright infringement; or they are the product of mass copyright infringement themselves, in which case all the frontier model trainers own rightsholders their entire valuation.¹

A few thoughts, with the disclaimer that a long time ago I was a baby lawyer qualified in England and Wales:

First! Much as the U.S. Supreme Court can just decide that they actually meant something else, similarly copyright and database rights can just, like, be changed if we want to. Because they are not physical things.

Second! Say you follow through the model and what OpenAI did was a bit like stealing an MP3 and they should be made to pay fines for ripping CDs/novels/webpages, then... how does the fine work?

Now the deal is that OpenAI et al will say that a) they put a lot of effort and money into ripping all those CDs, like, it cost a lot just buying all those CD-ROM drives; b) they should be totally rewarded for all that work, and c) anyway everyone is getting to benefit from it because there’s a free tier to ChatGPT etc.

But see, there’s another way we can deal with “how do we distribute the benefit of a societal good” and that’s also called “taxing them” and then once there’s the tax revenue then the distribution method is your regular political problem which is/should be good because we have whole systems of dealing with that and not leaving the money just in the hand of Mr. Altman or whomever.

The point of this I guess is that we should be able to say “yeah, you getting to have a platform and app and offering it for free to people is not a substitute for making a resource available for everyone to benefit from”.

1.2 Gordon Moore Folds Your Clothes

So Xiaomi, a Chinese company that makes a lot of things, has also made a robot. Or, rather, it has used machine learning to make a model that can be used to drive a robot. So far so AI Company Does Thing.

If you watch the video it’s, like, a fairly robot-does-state-of-the-art stuff video? There’s a robot with manipulators and it does things like fold clothes. It’s interesting how many of these demo videos involve household tasks like folding clothes, picking up pillows, tidying up and all of that stuff.

There’s something where AI boosters do an underpants gnome where they’re like 1. Do AI stuff, 2. Something Happens, 3. AI saves everyone and is magic and a god and everyone is happy.

Invariably one of those 2.5 steps is “AI gets to the point where it just keeps improving itself and we get out of the way”. That’s also things like “humans don’t need to farm food anymore” and “humans don’t build factories anymore”. There is a lot of handwaving there!

But. But, but, but. I made a quip about the Moore’s Law of Folding Laundry. The original Moore’s Law was the not-a-law observation that the number of transistors you could fit in a given area would double every 18 months, which other people took to be shorthand for “every 18 months, computers will double in speed”, which is totally not what it meant anyway. Also money was free then.

The Moore’s Law of Folding Laundry is like this: when you watch that Xiaomi demo video of the robot folding clothes (which should also come with the car ad disclaimer along the lines of “closed circuit, professional robot, do not attempt” and all that jazz), you’ll see in the corner that sometimes there’s a disclaimer that the activity is happening at 1x speed or 3x speed or whatever. Right now, the video shows autonomous shirt folding (new sentence!) at 3x speed.

The boosters are not necessarily wrong. A few years ago, robots could not really fold clothes! They might be able to do so under very constrained circumstances, but, like that’s doing laundry under laboratory conditions. This folding is under less lab-like decisions and you get into issues like extrapolating from “not being able to fold a shirt” to “able to fold a single shirt in 5 minutes” (or whatever) is two datapoints and not one you should be allowed to extrapolate from anywhere because otherwise you end up with rubbish like the AI 2040 plan².

I think it would be reasonable to assume that the robots will get faster at folding shirts. If you look at the slowdown, or the reason why the videos are sped up 3x is because I guess they are thinking very hard (sorry, expending tokens) in figuring out how to move next in the world. That doesn’t say anything as to affordability (what sort of processing power is involved in inferring the next steps for folding arbitrary shirts?)

What’s difficult to navigate here (or not, I suppose) is the continuum of “robots can fold shirts now” to “we now live a life of leisure where we are tended to by machines of loving grace”. What tasks out there are adjacent to “folding an arbitrary shirt”, or in the map of “tasks you can think of” how far are they away from “fold this arbitrary shirt”?

Moravec’s paradox³ is the one that’s all about it being (relatively) easy to do build computing machiens that can do “hard” things like “play chess” but stupendously difficult to do “easy” things like “count the number of letter Ls in the phrase ‘in-house philosopher and ethicist’” or other “easy” things like “why the fuck isn’t the door working in the videogame I’m designing”⁴ (the door problem is brilliant, I love it).

So on the one hand I am railing at the juvenile optimism of the AI 2040 lot, but also trying to take a tempered look at, I don’t know: you might be able to buy a clothes folding robot in 5 years? It might be clamped to a flat surface you’ve got? You might not even be able to afford it in this economic hellscape anyway? It’s just that everything is so jagged.

Short one, that.

How are you doing? I’m doing, well, sort of gesturing at everything really.

Best,

Dan