Winter Garden

February thoughts

I wonder, were there pre­vious sea­sons in the his­tory of com­puting that felt like this one? Reading Fire in the Valley, it seems clear that the birth of the per­sonal com­puter in the late 1970s and early 1980s offered a com­pa­rable rush. Did the web boom of the late 1990s feel so urgent? I don’t know … if you were there, I’d love to hear about it.

My own “core tech industry era” of the late 2000s felt exciting, but less urgent and more … fizzy. It was very Millennial, I suppose: affable, easygoing. Too comfortable, probably!

I was pleased to see this big, meaty Nature comment endorsing my claim that AGI is here 😋

Sev­eral smart friends responded to my first edition saying some ver­sion of this: “It’s not the G for gen­eral that bothers me … it’s the I for intelligence.” Which is totally reasonable, but/and, at this point: that’s simply the term.

I sin­cerely think it’s helpful to under­stand the word “intelligence” in AI or AGI as an abstract token that means “doing what these things do”.

A thing to notice about Claude’s constitution is that it is VERY long — and not just long, but verbose. Floppy.

But of course this is a doc­u­ment for a strange new audience. It’s not pri­marily describing Claude’s aspi­ra­tional char­acter to us; it is describing it to Claude.

The way lan­guage models “read” is very dif­ferent from the way humans read. Basically, they read every­thing at once — every token in the con­text window, slammed into place in parallel. Imagine a lens, focusing that whole field of view down into one decision: which token shall I pro­duce next?

A very fluent human reader can “get” a whole sen­tence in one glance, maybe even a short paragraph, but cer­tainly not an entire 20,000+ word doc­u­ment. That is exactly — and I mean literally, mechanistically, precisely — what lan­guage models do.

This new kind of reading sug­gests new forms of writing, new stan­dards for style and structure. The constitution’s pri­mary author is on the record saying that most people’s prompts are not long enough: you nearly always ben­efit by giving a model more to work with.

I’ve always believed brevity was the soul of basi­cally every­thing, so this makes me feel pretty itchy!

The Gemini API docs discussing the model’s thinking process are fully sci-fi. If you were reading these five years ago, you’d go: wuu­u­u­uuut

Hannu Rajaniemi is both a world-class, bleeding-edge sci-fi writer and a world-class, bleeding-edge entrepreneur. His latest novel Darkome and his new ven­ture Red Queen Bio are two sides of the same intel­lec­tual coin; it’s fas­ci­nating to con­sider them together.

I’ve been rereading William Gibson, also reading a few for the first time; so far, I’ve ticked through the sequence: Count Zero, Mona Lisa Overdrive, Vir­tual Light, Idoru, All Tomorrow’s Parties. (No need to reread Neuromancer, which is already engraved on the inside of my skull.)

They are such weird books … the writing is, more than anything, hypnotic, in a good way. I will never, ever be able to recount the plot of a Gibson novel, but the vibes are immac­u­late and indelible. AND, now is the time to return to these, because his prescience — about AI, the dig­ital occult, the grain of 21st-century life, every­thing — is so profound.

Here’s a 21st-century experience: receiving an email (unsolicited? did you sub­scribe to this? who can recall) with a gnomic link to a cursed PDF with the instructions:

Drop it into an LLM and say “hat on”.

Here’s Jason Willems on the deep puzzle of copy­right in the age of AI.

Here is Steve Krouse calling for powerful tools, not blath­ering agents.

Here is Dave Friedman theorizing about Apple’s on-device infer­ence strategy.

Man … on-device infer­ence makes SO much sense … and it so clearly MUST be where this all ends up … yet I under­stand that, presently, nothing com­pares to the leviathan models run­ning on super­beefy chips in dis­tant data centers. How long does this remote regency last? Five years? Ten??

I’m with Nathan Witkin: the METR graph is misleading.

It’s instruc­tive to actu­ally poke at the tasks in some of these model evaluations. OpenAI’s GDPval is a good example. Go ahead — read a couple of random tasks. You’ll notice, they are “realistic” and also … not. The tasks are weirdly her­met­i­cally sealed: each is served up along­side every­thing you need to com­plete it, like a problem on a stan­dard­ized test. None requires any inter­ac­tions with other workers or organizations. They are fully defined in a way that is basi­cally alien. All of this makes sense — it’s what makes an efficient, repeat­able eval — but it also pro­vides some rea­sons to dis­count the more breath­less reports here.

I think any/every AI eval is worth actu­ally inspecting, if/when you have the time. The ARC-AGI minigames are fun! This one took me a few min­utes to solve … 

P.S. “Make up your own eval then” is a fine retort, and let me tell you, I have ideas for some GOOD ones … 

I don’t want crit­ical engage­ment to crowd out plain recog­ni­tion of the gen­uinely lib­er­a­tory poten­tial of these tools. Because I do rec­og­nize it!

One of my cur­rent favorite internet hang­outs is a cozy space crafted by the writer Craig Mod with the help of Claude Code. Craig recently wrote a bit about the project, and I can add, from my per­spec­tive as a user, that it’s won­derful to use soft­ware so per­fectly con­textual, so opinionated.

I’ve always liked the idea that everyone has inside of them one (1) book. Maybe everyone also has one (1) piece of soft­ware, and now, with AI coding agents, we will get them out … and be done.

Arcee’s Trinity True­Base model is cool — the unmasked shoggoth:

If you’re a researcher who wants to study what high-quality pre­training pro­duces at this scale — before any RLHF, before any chat formatting — this is one of the few check­points where you can do that. We think there’s value in having a real base­line to probe, ablate, or just observe. What did the model learn from the data alone? True­Base is where you answer that question.

Most of my aggre­gate lan­guage model time, 2016-2026, has been spent with base models. We can think of these as pure capa­bility without usefulness: wild­fire rather than hearth, light­ning rather than battery. But, the capa­bility IS all there within them; I want to argue that ALL of it comes from the insanely demanding next-token pre­dic­tion task.

Every­thing that follows — finetuning, rein­force­ment learning, etc. — is essen­tially “EQing” that blinding poten­tial into some­thing people can actu­ally use and/or enjoy.

It’s like hearing a blast of static, yet knowing there’s a Bea­tles song hidden inside, if only you can carve it out. Maybe the AI com­pa­nies ought to start stocking this book in their libraries … 

Spencer Chang:

im bet­ting on a future aes­thetic rooted in proof of longterm exis­tence - a bronze statue worn away where people have touched it, the patina of old plastic doors, the grooves left in wood from con­sis­tent use

This is so canny, so contemporary, so obvi­ously and totally correct. Gibson-level insight.

The great per­sonal super­power of the decade ahead will be: remem­bering what you wanted to do in the first place.

New capa­bil­i­ties emerge, new manias unfurl, and the drum beats loud: you should be trying this … you better not miss that … and: sure, maybe! But maybe you should also remember what you wanted to do in the first place.

What’s actu­ally inter­esting to you? What has been inter­esting for the past ten years? What will still be inter­esting ten years from now? How do new capa­bil­i­ties sup­port that interest, or not?

Sharpen that keel! You’re gonna need it.

I’ll add one more thing. I think we are in a period so inter­esting that basi­cally every­body ought to be writing about it: reporting, reflecting, resisting, re-every­thing — in a blog, a newsletter, an email to friends, whatever.

I said this in my first edition: AI lan­guage models are a particularly, maybe even uniquely, human technology. They are para­dox­ical and poetic, like some­thing out of myth: the uncanny double, the magic mirror. Psy­cho­log­ical intu­itions have proven as useful as tech­nical insights.

You have a stake in this, simply by virtue of being a talking animal. Might as well jot a few notes for posterity, and for the rest of us, here and now.

From the lab,

Robin