This is Robin Sloan’s lab notebook. It’s about media and technology, creative computing, AI aesthetics, & more. Here's the RSS feed. My email address: robin@robinsloan.com
I encourage you to open Levent Alpöge’s PDF note announcing his proof of the Riemann Hypothesis with the help of AI — because although the work was AI-assisted, this note is so clearly and beautifully human. It’s fabulous.
I have noticed the steady creep of AI-generated text into the publications of the big AI companies themselves, and while it tends not to be on the level of “egregious slop”, it does lend everything a … smoothness. I note also that these publications are very rarely signed; they come from The Blob. Who wrote this?Who wrote this? I know it’s customary for big companies to speak Blob-like, but/and I still find it jarring.
Counterpoint Alpöge: his note is not only better than a smooth, AI-generated summation, it’s also better than an academic paper, because his passion — his appreciation of surprise and beauty — is so clear on the page. It seems almost to rise off the screen, a 3D layer of sparkling energy.
Tell you what, you can use all the AI you want in your project, if you write about it like this, in your real human voice.
Anthropic’s IPO is a buzzy, valedictory moment, particularly here in the San Franscisco Bay Area, but of course I always want to pop things out into temporal perspective. It’s instructive to think about Apple’s IPO in 1980, Amazon’s in 1997, Google’s in 2004 … all of them victory laps in their own way — yet those companies were, in those moments, barely premonitions of what they would become. Apple without the Mac or the iPhone; Amazon without AWS; Google without Chrome, YouTube, or, of course, Gemini.
Thinking about Anthropic, it’s strange to imagine a product as profound and pathbreaking as Claude to be “merely” its Apple II. Yet that’s surely correct. And, I don’t just mean that Claude 11 will make current Claudes look doltish; I mean the interface will change completely — even the deep sense of what the product “is”.
AI is all a rush — economic, existential — yet anyone who imagines that the finish line (superintelligent? utopian? apocalyptic?) is within sight is kidding themselves. This transformation — this sequence of transformations — has barely begun. I believe the RSI risks are real and urgent, but I also believe humanity is going to find a way through these next few years — not elegantly, but, that’s fine — and then … we just keep going! Always in tension, balancing and learning. You’re never late to the party because the party lasts forever. There are algorithms undreamed of, out there in high-dimensional space. Somebody will bust us out of the chat box, at last. The robotics thing is coming, but it will take decades of refinement. Decades!
You know the hoary old Churchill quote — you can almost hear his voice, can’t you? “This is not the end. This is not even the beginning of the end. But it is perhaps the end of the beginning.” Will AI progress slow down? I hope so — and I hope it’s less the feeling of an industry furiously restrained, and more the feeling of a field settling in for the long haul, all the hard work and invention still to come.
First, open this page, the configurator for various flavors of PyTorch — a grid that will be VERY familiar to some of you, as it is to me!
Then, scroll down to the bottom of this page, a presentation of a new AI-ready UI library. Click “copy prompt”, and notice with appreciation the little matrix of options that appears, to tailor the block of text you will imminently paste into a thread with an AI agent.
There you have it: the old software distribution, and the new.
It’s all just eye-wateringly weird, beyond (or off to the side of) any prediction made by ANYONE as recently as five years ago. Yet this fluid, organic capability was always there, latent in language, in Piles of Text. Well — maybe not always … but certainly since the year 2000. We had enough of a Pile by then to get started.
I think it’s great. The installer, the .tar, the .msi, the .dmg, has been superceded by, essentially, the postcard.
(Hey, that sounds like a fun way to sell software, doesn’t it? Pay $10 dollars, and within the week you’ll receive a postcard in the mail containing installation instructions along with a unique key. Take a picture and show it to your AI agent. Then put it in a drawer for easy reinstallation.)
I make this observation with zero snark, only appreciation: there are clearly a handful of rich, robust 3D modeling and gamedev RL environments being peddled in the marketplace right now, being snapped up by the big AI companies, for the specific reason that models with these environments in their post-training mix can produce spectacular, shareable visual demos. “I asked Model X to optimize my COGS” is one thing; “I asked Model X to let me ride a giant crystal dragon through a dramatic mountain landscape”, followed by a swooping, scintillating video, is quite another.
In terms of model capabilities, these demos seem to me more like superimpressive recapitulation than truly dynamic invention. “Sufficiently advanced imitation is indistinguishable from magic”, etc. But! The value of these capabilities lies elsewhere: as marketing juice, grist for the “wow!” mill. And in that respect, they’re totally brilliant. I just hope the price of acquiring the RL environments is being tallied in the “marketing” column 😉
The most helpful thing you can do is probably to tell your agent to submit diaries of its experience building with Exact to the feedback endpoint. Detailed logs of agents’ experience are gold for pinpointing what to fix and improve. Thanks for checking it out.
This makes me think of all those UX researchers over all those years struggling to prise feedback out of users: some sense of “tell me what you’re thinking” or “how you’re struggling” … And now, the agents just record it as a matter of course, because their work IS the transcript of their work.
Frameboy is a very clever and appealing demo from World Labs, purveyor of 3D “world models”. Good use of skeumorphism — it’s basically impossible not to pop at least one “cart” into the little console.
Anytime I wander around one of these on-demand spaces — Google DeepMind has its own offering—I am struck by the dazzling fidelity … and total lack of interest. Within moments, my attention “slides away”, because on some level — possibly animal — I understand there’s nothing important to look at. Every detail is only and exactly “what you’d expect” from the given environmental trope.
It makes me think of a few things. For starters, Max Kreminski’s great paper, The Dearth of the Author:
[ … ] a condition that arises when AI-based creativity support tools for writing allow users to produce large amounts of text without making a commensurate number of creative decisions, resulting in output that is sparse in expressive intent.
Visually rich but “sparse in expressive intent” is exactly how I’d describe the World Labs environments.
It makes me think also of dreams — the way the setting of a dream often just represents the gestalt, the core vibe, of a real place. You “know” you’re back in high school, but if, upon waking, you had to sketch the scene, there wouldn’t really be anything to draw. The most accurate depiction might simply be the words: “HIGH SCHOOL?!”
Of course, AI almost always makes me think of dreams.
I’m certain there will be ways to pack these on-demand worlds with more creative decisions; one imagines the artist of the future wandering a World Labs scene under construction, saying, “No, change this lamp. Delete that wall. Let’s put some evocative items on the fridge.” In his latest novel, Ken Liu imagines a new medium: vivid shared dreams authored and guided by an “oneirofex” of this kind. This feels totally plausible to me, eventually — but we’ve got a long way to go yet.
P.S. Update: we found the art!!! I strongly encourage you to boot up one of the Frameboy carts, choose a direction, and: keep walking. (Remember, you can use Q and E to adjust your elevation, if needed.) The engine will gradually “slow you down”, and it eventually feels like moving through molasses, but you must endure, for the spectacular breakdown that awaits you, as you reach the bounds of the model’s “imagination” … Try the cyberpunk cart: wander around until you find the window with a view of the city, then: walk straight out, and keep going.
It’s been a long time since I purchased anything from a Crowd Supply campaign, but I open their email newsletters eagerly and promptly, because the projects are so unlike anything else in the tech tealm — hyperspecific, fully niche, a bit raw.
Case in point: here’s a modern, drop-in replacement PCB for old iPods — a whole new brain and heart to slip inside that classic body. What a cool idea.
I highly recommend subscribing to the newsletter—it’s honestly a tonic, every project an emblem of some nerdy preoccupation. Plus, it’s hard to argue with their rallying cry: for hardware that is “original, useful, respectful”.
Today’s “ask for what you want”: a rich visual research report on “box logics” in print media, from the beginning of time until today. I’ll put my full prompt in a details element below. The short version is: I am interested in a broad family of graphic design ideas, and I wanted an AI agent to fetch me examples.
No particular sense of “best practices” here — I share the prompt only because it can be interesting to see how other people approach this. As usual, I put mine in its own little file, RESEARCH-IDEA.md, and I spent 5-10 minutes writing it. I find this more effective and frankly tolerable than plunging directly into a back-and-forth with an AI agent.
Here it is.
We’re working on some refreshed design concepts for use across both print and web, for a series of books AND stationery items that are accompanied by a ton of web content.
Here are some threads of inspiration, in no particular order.
One is the book cover system used decades ago by Faber & Faber, with example PNG files in this dir. Note the simple stacked boxes, the way they become the visual anchor placed over rich, full-bleed images.
Another is TV guide grid layouts — the interesting sort of information design that emerges from the need to show all these different channels, streams, of information.
A third is “nerdy data displays” on things like baseball cards, comic book character guides, the backs of toy boxes, collectible trading card games … anything that wants to show you “the stats”, whether real or fictional.
Yet another one is 1980s and 1990s computer GUI designs, especially those with hard corners and edges. NeXTSTEP is an example.
Here’s what appeals to us about these “box logics”:
They go against the grain of modern “round, friendly” digital design. The right angle! The hard corner!
They offer an extensible system for different needs and contexts.
They can easily “float above” richer graphical content – as seen in the Faber & Faber book covers. Our materials WILL be highly graphic, so having a design scheme that’s more sober, minimal, even severe, makes for a nice mix.
I should stipulate, we’ll be using these grid systems essentially in the context of “covers”, both in print and on the web … putting the key data about a content item into these boxes: title, author, publication date, other little fun bits of metadata, taglines, summary lines, etc. AT THE SAME TIME, it should “scale” to handle other forms of information … metadata at the bottom of a web page, the content on a book’s copyright page, etc. The point however is that these materials aren’t going to wall-to-wall grids. This will be a tool used sparingly, but always in the spotlight.
Okay that’s are starting point. We’re interested in doing a few things. The outputs here will be standalone web page reports with a focus on visual examples.
Here’s what we want:
A comprehensive survey of “box logics” across print media, going back as far in time as we can. We are looking for hi-res, useful visual examples of things including baseball scores, stock market reports, TV guide listings (as mentioned) … anything and everything that “puts text and data in a neat little box”. Tabular data, too. Let’s dig up a TON of good/interesting examples. This survey should be global — be aggressive about searching beyond English!
A survey of flexible grid-based graphic design in digital systems specifically — various UIs, GUIs, OSes, tiling window managers, etc. So, this is the “digital / screen-based” analog to the print research run above.
Finally, a “what if?” exercise, one that combines our starting impulses with the results of #1 and #2 above and synthesizes them into some weird/wacky ideas about potential box logics, grid systems, etc. Creative formulations and applications. Notions that take my assignment here and sort of twist it through another dimension. These should be live HTML/CSS as appropriate. Expect the unexpected! You get it.
Okay, that’s a good bundle of work. Please research very widely and take your time. Ask me any questions needed, then proceed with the project.
I recommend setting these kinds of queries to run overnight, so you can enjoy the results over coffee in the morning. (My overnight computer task, years ago, was the language model training run. In the morning, I’d see how low the loss had dropped, generate some sample lines. We’ve come full circle.)
You’ll find it’s like reading a little magazine, just for you. Is it the world’s best magazine? No. Is it weirder and more specific than any magazine could ever be? Yes! I prefer local web pages, HTML files on my laptop, to Claude’s “artifacts” or any other proprietary format.
I’m now reviewing mine, and already I’ve discovered something new to me, that I’m sure I wouldn’t have found on my own: the banzuke, a graphic grid used to display sumo rankings, widely adapted (and parodied) in Japanese graphic design.
Is this a world-shaking discovery? No. Yet my view towards research in books has always been that one single, gleaming, new idea — or scrap of language — or salient image — justifies a book’s entire purchase, and the time it took to read it. That’s clearly a writer’s calculus; we build new work from those scraps. In that view, I’ve already gotten my salient image from this report, and there’s still more to digest, so we’ll call it a smash success.
The “creative” part of the report fits a familiar pattern: there’s nothing I’d ever actually use, but I still find it productive to consider the agent’s work … the way any combinatorial or even quasi-random exploration can be productive. In your assessments — “That’s so dopey … ”—you discover your own opinions and ideas.
Goes without saying, any of these reports should be treated as definitionally incomplete — plenty of the really good stuff isn’t available online. Even so: what a cool magazine!
Gemini 3.7 Flash is now deprecated and no longer available as an independent model in the Gemini API and Google AI Studio. All traffic is automatically redirected to Gemini 3.8 Flash.
I could be wrong, but I don’t believe there was any advance warning about this.
I assume the justification here is that 3.8 Flash is, as they say in trading card games, “a strict upgrade”—i.e. there’s no situation in which 3.7 Flash is better than 3.8 Flash. And yet … how would they know?? The whole value of these models is that they are useful in a breathtaking variety of contexts … so, who’s to say there isn’t some niche application that just broke, because 3.8 Flash has a subtly different opinion of pictures of radishes than reliable old radish-loving 3.7?
I will note that 3.7 was introduced in early August. It was barely two months old. (Update: I just received the same notification about 3.5, which now redirects to 3.6.)
I don’t know … I wonder sometimes if any of these companies really expect anybody to build anything on top of these models at all. If all you’re doing is yakking with a coding agent: easy enough. But if you’re using them programmatically, as steps or gates in a larger process, their organic complexity makes the concept of “a breaking change” basically impossible to reckon. Who knows? Who can know?? Nobody.
I suppose the really AGI-pilled response is, so what? If it breaks, get 3.8 Flash to fix it.
Forty years after its imagining, cyberpunk has arrived, emphatically, totally for real. The configuration of global power is different than the cyberpunks imagined — in a way, it’s less interesting — but the tech is even better. Ghosts of language in the machine; recalcitrant bot swarms; gleaming nodes of compute, dense with possibility.
I’m still running the “deep learning rig” I built back in 2016. Its three Titan X GPUs (purchased used on eBay; hot shit at the time) can run some of the smaller language models, barely, and very slowly, so I don’t bother, allocating my precious 36 gigabytes instead to more playful applications. Art stuff. My rig, called clash—get it? … of Titans? — is big and black, as functional and unlovely as I could make it. I invited an AI agent to poke around the server the other day. I didn’t really understand how many digital-archaeological traces remained littered around the system; the agent’s reaction was, basically, “Damn, you’ve been up to some weird stuff in here … ”
Slabs of compute running hot in offices, basements, dorm rooms … not a new thing, exactly, but their capabilities are new — their inhabitants are new — and the whole scene is just great. Exciting and strange. Gleaming and haunted. Pile on the adjectives, the vibes: it’s somehow all richer and weirder than anybody ever dreamed in 1984.
P.S. Somehow Oxide fits into this too, even though that particular gleaming node isn’t for AI …