Robin Sloan
the lab
March 2023

Phase change

An extremely close-up photograph of a snowflake, looking almost architectural.
Snowflake, Wilson Bentley, ca. 1910

Earlier this week, in my main newsletter, I praised a new project from Matt Webb. Here, I want to come at it from a different angle.

Briefly: Matt has built the Braggoscope, a fun and useful appli­ca­tion for exploring the archives of the beloved BBC radio show In Our Time, hosted by the inim­itable Melvyn Bragg.

In Our Time only provides HTML pages for each episode — there’s no struc­tured data, no sense of “episode X is connected to episode Y because of shared feature Z”.

As Matt explains in his write-up, he fed the plain-language content of each episode page into the GPT-3 API, cleverly prompting it to extract basic metadata, along with a few subtler properties — including a Dewey Decimal number!?

(Explaining how and why a person might prompt a language model is beyond the scope of this newsletter; you can read up about it here.)

Here’s a bit of Matt’s prompt:

Extract the description and a list of guests from the supplied episode notes from a podcast.

Also provide a Dewey Decimal Classification code and label for the description

Return valid JSON conforming to the following Typescript type definition:

  "description": string,
  "guests": {"name": string, "affiliation": string | null}[]
  "dewey_decimal": {"code": string, "label": string},

Episode synopsis (Markdown):


Valid JSON:

Important to say: it doesn’t work perfectly. Matt reports that GPT-3 doesn’t always return valid JSON, and if you browse the Braggoscope, you’ll find plenty of ques­tion­able filing choices.

And yet! What a technique. (Matt credits Noah Brier for the insight.)

It fits into a pattern I’ve noticed: while the buzzy appli­ca­tion of the GPT-alikes is chat, the real workhorse might be text trans­for­ma­tion.

As Matt writes:

Sure Google is all-in on AI in products, announcing chatbots to compete with ChatGPT, and synthe­sised text in the search engine. BUT.

Using GPT-3 as a function call.

Using GPT-3 as a universal coupling.

It brings a lot within reach.

I think the magnitude of this shift … I would say it’s on the order of the web from the mid 90s? There was a radical simpli­fi­ca­tion and democ­ra­ti­sa­tion of software (architecture, development, deployment, use) that took decades to really unfold.

For me, 2022 and 2023 have presented two thick strands of inquiry: the web and AI, AI and the web. This is evidenced by the structure of these lab newsletters, which have tended towards birfucation.

Matt’s thinking is inter­esting to me because it brings the strands together.

One of the pleasures of HTTP (the original version) is that it’s almost plain language, though a very simple kind. You can execute an HTTP request “by hand”: telnet 80 followed by GET /.

Language models as universal couplers begin to suggest protocols that really are plain language. What if the protocol of the GPT-alikes is just a bare TCP socket carrying free-form requests and instructions? What if the RSS feed of the future is simply my language model replying to yours when it asks, “What’s up with Robin lately?”

I like this because I hate it; because it’s weird, and makes me feel uncomfortable.

I think it’s really chal­lenging to find the appro­priate stance towards this stuff.

On one hand, I find critical deflation, of the kind you’ll hear from Ted Chiang, Simon Willison, and Claire Leibowicz in this recent episode of KQED Forum, appro­priate and useful. The hype is so powerful that any correc­tive is welcome.

However! On the critical side, the eval­u­a­tion of what’s before us isn’t sufficient; not even close. If we demand humility from AI engineers, then we ought to match it with imagination.

An important fact about these language models — one that sets them apart from, say, the personal computer, or the iPhone — is that their capa­bil­i­ties have been surprising, often confounding, even to their creators.

AI at this moment feels like a mash-up of programming and biology. The program­ming part is obvious; the biology part becomes apparent when you see AI engineers probing their own creations the way you might probe a mouse in a lab.

The simple fact is: even at the highest levels of theory and practice, no one knows how these language models are doing what they’re doing.

Over the past few years, in the evolution from GPT-2-alikes to GPT-3-alikes and beyond, it’s become clear that the “returns to scale”—both in terms of (1) a model’s size and (2) the scope of its training data — are expo­nen­tial and nonlinear. Simply adding more works better, and works weirder, than it should.

The nonlin­earity is, to me, the most inter­esting part. As these models have grown, they have undergone widely observed “phase changes” in capability, just as sudden and surprising as water frozen or cream whipped.

At the moment, my deepest engage­ment with a language model is in a channel on a Discord server, where our gallant host has set up a ChatGPT-powered bot and laced a simple person­ality into its prompt. The socia­bility has been a revelation — multiplayer ChatGPT is much, MUCH more fun than single player — and, of course, the conver­sa­tion tends towards goading the bot, testing its boundaries, luring it into absurdities.

The bot writes poems, sure, and song lyrics, and movie scenes.

The bot also produces ASCII art, and SVG code, and PICO-8 programs, though they don’t always run.

I find myself deeply ambivalent, in the original sense of: thinking many things at once. I’m very aware of the bot’s limitations, but/and I find myself stunned by its fluency, its range.

Listen: you can be a skeptic. In some ways, I am! But these phase changes have happened, and that probably means they will keep happening, and no one knows (the AI engineers least of all) what might suddenly become possible.

As ever, Jack Clark is my guide. He’s a jour­nalist turned AI practioner, involved in policy and planning at the highest levels, first at OpenAI, now at Anthropic. And if he’s no longer a disin­ter­ested observer, he remains deeply grounded and moral, which makes me trust him when he says, with confidence: this is the biggest thing going, and we had all better brace for weird times ahead.

What does that mean, to brace for it?

I’ve found it helpful, these past few years, to frame my anxieties and dissat­is­fac­tions as questions. For example, fed up with the state of social media, I asked: what do I want from the internet, anyway?

It turns out I had an answer to that question.

Where the GPT-alikes are concerned, a question that’s emerging for me is:

What could I do with a universal function — a tool for turning just about any X into just about any Y with plain language instructions?

I don’t pose that question with any sense of wide-eyed expectation; a reason­able answer might be, nothing much. Not every­thing in the world depends on the trans­for­ma­tion of symbols. But I think that IS the question, and I think it takes some legit­i­mate work, some strenuous imagination, to push yourself to believe it really will be “just about any X” into “just about any Y”.

I help operate a small olive oil company, and I have spent a bit of time lately consid­ering this question in the context of our business. What might a GPT-alike do for us? What might an even more capable system do?

My answer, so far, is indeed: nothing much! It’s a physical business, after all, mainly concerned with moving and trans­forming matter. The “obvious” appli­ca­tion is customer support, which I handle myself, and which I am unwilling to cede to a computer or, indeed, anyone who isn’t me. The specific quality and character of our support is important.

(As an aside: every customer support request I receive is a miniature puzzle, usually requiring deduction across several different systems. Many of these puzzles are challenging even to the general intelligence that is me; if it comes to pass that a GPT-alike can handle them without breaking a sweat, I will be very, very impressed.)

(Of course, it’s not going to happen like that, is it? Long before GPT-alikes can solve the same problems Robin can, using the tools Robin has, the problems themselves will change to meet the GPT-alikes halfway. The systems will all learn to “speak GPT”, in some sense.)

The simple act of asking and answering the question was clar­i­fying and calming. It plucked AI out of the realm of abstract dread and plunked it down on the workbench.

Jack Clark includes, in all of his AI newsletters, a piece of original micro-fiction. One of them, sent in December, has stayed with me. I’ll reproduce it here in full:

Reality Authentication

[The internet, 2034]

“To login, spit into the bio-API”

I took a sip of water and swirled it around my mouth a bit, then hawked some spit into the little cup on my desk, put its lid on, then flipped over the recep­tacle and plugged it into the bio-API system.

“Authenticating … authentication successful, human-user identified. Enjoy your time on the application!”

I spent a couple of hours logged-on, doing a mixture of work and pleasure. I was part of an all-human gaming league called the No-Centaurs; we came second in a mini tournament. I also talked to my therapist sans his augment, and I sent a few emails over the BioNet protocol.

When I logged out, I went back to the regular internet. Since the AI models had got mini­tu­ar­ized and prolif­er­ated a decade ago, the internet had radically changed. For one thing, it was so much faster now. It was also dangerous in ways it hadn’t been before - Attention Harvesters were every­where and the only reason I was confident in my browsing was I’d paid for a few protection programs.

I think “brace for it” might mean imagining human-only spaces, online and off. We might be headed, paradoxically, for a golden age of “get that robot out of my face”.

In the extreme case, if AI doesn’t wreck the world, language models could certainly wreck the internet, like Jack’s Attention Harvesters above. Maybe we’ll look back at the Web Parenthesis, 1990-2030. It was weird and fun, though no one in the future will quite under­stand the appeal.

We are living and thinking together in an inter­esting time. My recom­men­da­tion is to avoid chasing the ball of AI around the field, always a step behind. Instead, set your stance a little wider and form a question that actually matters to you.

It might be as simple as: is this kind of capability, extrap­o­lated forward, useful to me and my work? If so, how?

It might be as wacky as: what kind of protocol could I build around plain language, the totally sci-fi vision of computers just TALKING to each other?

It might even be my original question, or a version of it: what do I want from the internet, anyway?

From Oakland,


March 2023, Oakland