At home in high-dimensional space

In this edi­tion of my tech­nical newsletter, I want to share a few thoughts and links that have stacked up in my notes recently.

First, though, I want to be sure you know about Moonbound, my new novel coming in June.

This novel is more ger­mane to the inter­ests of this newsletter than my pre­vious work, for a few reasons:

• Where my pre­vious two novels are notion­ally realist fiction, though clearly influ­enced and ener­gized by the vibes of sci­ence fiction, this novel is: straight­for­wardly sci-fi!

• My exper­i­ments over the years with AI and writing (e.g. this one in 2016, this one in 2022) resulted not in a story written with AI tools — in the end, that approach just was not inter­esting to me — but rather a story informed by the “grain” of that material, as I’ve come to under­stand it. It makes a rich, weird vein through Moonbound.

• The novel also boasts an in-depth treat­ment of the math­e­matics and aes­thetics of high-dimensional space. This sub­ject emerges organ­i­cally in the story; there is, at one point, an urgent need for my pro­tag­o­nist to get good, fast, at high-dimensional thinking! But, I’ll confess, I really love “didactic stories” (like the inim­itable anime Dr. Stone, which I raved about here) so I was happy to take the oppor­tu­nity to tell one.

I’ll now encourage you to pre­order Moonbound, which you can do any­where books are sold, in any format you like, print or dig­ital or audio. Barnes & Noble is a great option; Amazon is, of course, very convenient.

I know you under­stand very well the power of the algorithm: the way atten­tion compounds. What you might not under­stand is the rel­a­tively modest scale of book pub­lishing success. It only requires sales in the single-digit thou­sands to pop a book onto the best­seller lists, which can become gate­ways to fur­ther success. The point of the pre­order, then, is to focus a dif­fuse field of interest into the hot week of a book’s release.

Delivering the mail in Val Town

I’ve pre­viously expressed enthu­siasm about Val Town. Now, I’ve actu­ally used it for some­thing, and I can report that my enthu­siasm has only grown.

Val Town offers a light­weight web editor for Type­Script func­tions that can run in a variety of ways. For me, the killer app is the built-in email han­dler, which works like this:

1. You create a “val” and des­ig­nate it an email han­dler. That val is auto­mat­i­cally con­nected to an email address that looks some­thing like this: yourUserName.coolEmailHandler@valtown.email.

2. You write a func­tion that accepts an Email object. This func­tion will be called auto­mat­i­cally when new emails arrive, and it can do anything! Maybe you want to save the mes­sage to a database. Maybe you want to parse the body and exe­cute some command. Maybe you want to send a reply — Val Town will happily do that.

3. You start sending emails to your val!

I think a par­tic­u­larly pow­erful option here is to pass the body over to a lan­guage model, along with a prompt explaining what kind of infor­ma­tion you’d like to extract. In this way, an email han­dler val can become a bridge between the chaotic, unstruc­tured world of email and what­ever more formal, schematic require­ments you might have.

Indeed, this is part of what’s hap­pening behind the scenes with my pre­order registration, described above. I wanted to avoid even the min­imal fric­tion of “please fill out this Google Form”, and since an online order is almost always rep­re­sented by a con­fir­ma­tion email, I wondered, “What if people could just for­ward those emails along? What if I … didn’t have to read them all myself?”

Get­ting that flow up and run­ning as a val was easy and fun.

This is not a huge deal; honestly, a Google Form would have been fine. But, I am an incor­ri­gible sender of emails-to-self; it’s how I log and manage most of my notes, including many of the items that you’ll find in this newsletter. So, this expe­ri­ence has got me thinking: how might I enrich that flow, adding some struc­ture along the way? What new fla­vors of emails-to-self might I conceive, with what kinds of useful “side effects”?

This is all to say, Val Town has fired up my imagination — always a good sign. The com­pany and user­base alike are small enough to feel convivial and responsive; it’s a cool plat­form at a cool time.

Robin Rendle feels pow­erfully the romance, the energy, the POTEN­TIAL of modern CSS, and he expresses all that in his new blog and newsletter, The Cascade. I was a devoted reader of Robin’s con­tri­bu­tions to CSS Tricks, so I am all in for this new stream of writing and discovery.

Meta proclaims that its Llama 3 lan­guage model was “pretrained on 15T tokens from pub­licly avail­able sources”.

Fif­teen tril­lion tokens! Call that eleven tril­lion Eng­lish words. If, just for fun, we say that 100,000 words equals a book, that’s equiv­a­lent to a hun­dred mil­lion books — only slightly less than the total number (estimated by Google Books) to have been published, ever, since the inven­tion of the printing press.

I have to confess, it remains strange to me that the AI folks worried, for so many years, about the com­po­si­tion of their training corpora — not only its legal status but also its struc­ture, its ori­gins and omissions … 

 … and then, starting some­time around 2021 or 2022, they simply: didn’t.

I think that’s because beg­gars can’t be choosers, and every engi­neer of a large model is presently, where data is concerned, a des­perate beggar indeed. The require­ment for SO MUCH DATA prob­ably reveals some­thing deeply weird, and far from ideal, about the cur­rent approach to mod­eling and training these systems. Yet the con­nec­tion is clear: they do pow­erful things when you dump more data into them, so, ideal or not: more data must be had!

A recent edi­tion of Jack Clark’s indis­pens­able AI newsletter dis­cussed the pro­duc­tion, to this end, of syn­thetic data — i.e., data not gath­ered from the “real world”, but pro­duced expressly for AI training. Some­times it’s pro­duced using straightfor­ward com­puter code; some­times it’s pro­duced by another AI model 😵‍💫

It’s odd to con­tem­plate these vast new “books”—training corpora — that will never be read (could never be read) by any human. They can be under­stood and eval­u­ated only statistically, through spot-checking or auto­mated analysis.

I’ll never stop saying: it’s tragic that Borges missed all this. He would have loved it.

The dig­ital essay Models All the Way Down by Christo Buschek and Jer Thorp takes a beau­tiful swing at these complexities. I like the bit that Robin Rendle picked out:

Here we find an impor­tant truth about [this dataset]:

It con­tains less about how humans see the world than it does about how search engines see the world. It is a dataset that is pow­erfully shaped by com­mer­cial logics.

A different activity altogether

When you raise ques­tions about AI training data — anything related to copy­right, fair use, attribution, etc. — you’ll often encounter a defense that goes some­thing like this:

What’s the big deal? Robin Sloan “trained” him­self on a ton of copy­righted books, didn’t he? He learned from them, then went on to write books of his own, and nobody calls that copy­right infringement!

This might be a rea­son­able argu­ment if AI models oper­ated at the speed and fidelity of human writers and artists. It’s true, Robin Sloan did read a ton of copy­righted books. However, he did not read all the copy­righted books, and even then, the task took him decades. Furthermore, he gen­er­ates output at the rate of approx­i­mately one book every four years, which works out to approx­i­mately one token per hour 😉

When capa­bility increases so substantially, the activity under dis­cus­sion is not “the same thing, only faster”. It is a dif­ferent activity altogether. We can look to the physics of phase change for clues here.

Basically, I want to immu­nize you against this analogy, and this objection. There’s plenty to debate in this nascent field, but any com­par­ison between AI training and human edu­ca­tion is just laugh­ably wrong.

This is what dig­i­ti­za­tion does, again and again: by removing fric­tion, by col­lapsing time and space, it under­mines our intu­itions about pro­duc­tion and exchange.

No human ever metab­o­lized infor­ma­tion as com­pletely as these lan­guage models. “As our case is new, so we must think anew, and act anew.” You’re gonna need fresh intu­itions.

Black box science

What’s the value of these ever-growing AI models, really? I know sev­eral people working in this domain who believe the goal is one (1) thing of over­riding consequence, which we might call vir­tual Wolf­gang Pauli, or maybe on-demand Albert Ein­stein: an AI model that can actu­ally pro­duce path­breaking sci­en­tific theory.

In this formulation, eco­nomic and social trans­for­ma­tion would be a second-order effect: not of “AI itself”, but of tech­nology derived from sci­ence it pro­duces.

I think this vision is weirdly more plau­sible than AI as “general labor replacement”. I sup­pose you could counter by saying, if they can engi­neer a vir­tual Pauli, they can FOR SURE engi­neer a vir­tual employee; one task is strictly “easier” than the other. But I don’t know if that’s true. Was Albert Ein­stein a good employee?

I think about this sce­nario a lot. For me, it’s more imag­i­na­tively compelling, with fewer ready­made sci-fi precedents, than “AI leisureworld”.

Reading his­to­ries of physics in the early 20th century, it’s thrilling to learn about the intel­lec­tual and social ferment, the rich net­work of “whys” behind every step for­ward. In an imag­i­nary AI-powered annus mirabilis, those “whys” might be absent. No story; no development; just theory. But also, perhaps, testable predictions, and new expla­na­tions for weird phenomena, as con­se­quen­tial as Ein­stein’s for the pre­ces­sion of Mercury.

But sci­ence is a social process; the AI folks under­stand this very well. How would AI-generated “raw theory” be chan­neled into the real world of sci­ence and tech­nology? How would you know when your vir­tual Pauli had a theory worth testing? What if it spat out a mil­lion theories, and you had good reason to believe one of them was correct — a real paradigm-buster — but you didn’t know which?

I come down on the side of skepticism, but/and … it’s chewy stuff! Fun to think about.

My memory of the ter­rific TV series Halt and Catch Fire has grown a bit murky, which means it’s almost time for a rewatch; that would make my third viewing. If you’re a tech­nical or tech­nical-adjacent person, you really ought to make time for this one. Don’t let the melo­dra­matic opening episodes throw you — the series quickly matures into one of the best-ever tele­vi­sual pro­duc­tions about col­lab­o­ra­tion and creation.

I encourage you to take just a moment to con­tem­plate this twenty-sided die, and the one depicted at the top of this edi­tion, too. These were warmed by palms nearly two thou­sand years ago. You could play D&D with them today, if the Met would let you.

Durable delights. Com­puter pro­grams can’t do this — not yet.