(One) good AI is here ⊗ Taste as interface behaviour

(One) good AI is here

Featured articles around here are often quite conceptual, aspirational, and/or directional. What could we do, what should we do, how? This one by Anil Dash is an actual working example of what “better” can look like in AI. To be fair to myself, I’ve often linked to projects that hit some of the same notes, like Māori Data Sovereignty inspiring new AI voice models or the work of Abundant Intelligences.

Still, great to see new projects popping up. Dash thinks we may have seen a rare glimpse of an AI that’s “actually” good, and the example comes from an unexpected place. Corridor Digital, a YouTube video-effects studio, built CorridorKey, a tool for green screen keying—the tedious work of cutting a foreground subject out of its green background. Cofounder Niko Pueringer, who’s not a developer, trained a neural network on footage he generated himself, then released the code. A community formed around it within days, made it usable for non-experts, and cut the hardware requirements down to ordinary desktops. He later also released the training data. Dash runs through the checklist: trained on consensually gathered data, open source and open weights, able to run on ordinary machines and renewable power, and controlled by “creators, not extractors.”

The wider argument is that Big AI’s worst habits are a choice, not a necessity. If a small team that isn’t even a tech company can make an AI that takes nothing without consent, shows its workings, and phones home to no one, the largest companies in the world can’t claim they have no option. Dash’s second point is that CorridorKey works because it has a specific job, not because it tries to be everything to everyone. A tool built for one community can’t tell a kid to self-harm the way a general chatbot can, and accountability gets easier as scope narrows. The model is under-hyped rather than over-hyped—the launch was buried near the end of a video that doesn’t even name the product.

Dash doesn’t mention Karen Hao, but his “purpose, not one-size-fits-all” makes the same case she does, that smaller, targeted models are the saner alternative to the scale-everything posture of the big labs, with lower costs and lighter environmental load. Both arrive at the same place from different directions: the harms critics name aren’t intrinsic to the technology, they follow from a particular set of business decisions.

The AI zealots have done themselves no favors, by not only dismissing all of these valid criticisms, but by also making increasingly outlandish and extreme claims about the capabilities of the Big AI platforms, while simultaneously scaremongering about the brutal effect they’ll have on people’s lives and careers. It’s no wonder the public sentiment about AI has become so negative. […]

This might be a little tiny bit of a good AI future, and now we just need to distribute the same kind of thing to a lot more people. […]

All of these traits are things that can be replicated in many more fields, by many more passionate people who don’t have to necessarily be experts, but who care about displacing the tech tycoons’ one-size-fits-all platforms with something that is human-scale and accountable.

Taste as interface behaviour

Jay Springett takes a recent Guardian piece by Rachel Aroesti as his starting point and pulls it into a number of topics he‘s been exploring over the past few years. Aroesti worries that algorithmic feeds have stripped us of personal taste, that what feels like preference is residue from being shown the same thing a hundred times a day. Springett largely agrees, but comes to a different conclusion: taste hasn’t been destroyed, it has been “translated into interface behaviour.” Record shops, libraries and radio were interfaces too, with their own taxonomies and recommendation systems, so the nostalgia for pre-internet discovery as a pure encounter with culture doesn’t hold. What separates the feed is latency and aim—the record shop’s feedback loop runs in months and points at a market, while the feed runs in milliseconds and points at a single person, reshaping the next view based on every gesture of “pausing, not skipping, replaying, hovering.” (I have a sudden craving for a High Fidelity rewatch.)

He also pushes back on the idea that we no longer choose. Attention is sovereign, and the choices remain ours even when dark patterns press on them; they’ve just been compressed into a stream of tiny, pre-formatted gestures. Drawing on Emily Segal’s notion of “tasteslop,” Springett locates the failure in the automation of the classifying function: the machine knows the gestures clustered around taste, not the relation between a person and a work. Segal calls taste a socially validated relation between objects, people, histories, scenes and timing, and Jay extends this by arguing that discernment is the one part of that relation the feed cannot perform for you. Being influenced was never the problem; the failure to be discerning about what is influencing you is.

Discernment needs latency, and the feed is built to collapse it. Taste matures in the delay between exposure and judgement, so the practical work is putting that delay back—letting a thing sit long enough for your relation to it to form before the interface asks again. Once the delay returns, discernment can rise a level, past whether a given video suits you toward whether the medium itself does. “If your Attention is Sovereign, then taste is what you do with that sovereignty; and discernment is how you keep it from being spent for you. Otherwise you end up with no taste at all. Only a feed-shaped pattern of exposure and reaction.”

Side note → The first quote below feels very Nietzsche’s “If you stare into the abyss, the abyss stares back at you.”

A window is something you look through, the feed is something that watches you doing the looking. It’s an apparatus that observes you observing, then changes its own behaviour based on that observation, which in turn shapes the future view you see though the window. […]

Inside the feed you are being presented with a ‘culture-world’ shaped by all your previous gestures of choice. The self is still in there somewhere, but it’s increasingly buried under its own captured behaviour. […]

Taste is not just what passes through you. It is what culture becomes after passing through you, via a particular interface. […]

[The feed] can supply the cultural objects, but it can’t do the telling-apart for you, no matter how hard it tries. Discernment is not the whole of taste, but in the 2020s it may be the most important part: choosing one thing over another that the feed serves up. Which in many ways we are doing all the time, acticly choosing. […]

Being influenced is not the problem online, not being discerning about what is influencing you is the problem. It’s the same with advertising, as the same flattened interfaces that serve you culture, also serve you adverts.

§ la machine is a quirky little thing I quite like, kind of this week’s “feed cleanse.” The team includes a few veterans of the mythical Nabaztag. “With her vibrant colors and geometric forms, la machine draws direct inspiration from legendary Italian designer Ettore Sottsass and the Memphis movement — a joyful rebellion against boring, functional design. Part object, part character, she’s a statement piece that refuses to blend in.” If you want one on your desk, Sentiers readers get 10% off, just use the suppasekrit code SENTIERS.

Futures, Fictions & Fabulations

• Our world of futures studies as a mosaic 2. “A new volume exploring the diversity and richness of futures studies and foresight through perspectives from across the globe. Building on the first volume published in 2024, this publication continues to explore how the field and its evolution are shaped by different regional, cultural, and intellectual traditions and contexts.”
• Dator’s Four Futures. “While we can find these four phases in any living organism, we can see them best (particularly Transformation) in human systems, which have have been noticeably self-improving, between generations, since the advent of human intelligence.” (Via the Near Future Laboratory Newsletter.)
• Five very different science fictional takes on space habitats. “Readers will notice most of the five works I am reviewing here were published roughly between the publication of Gerard K. O’Neill’s 1977 The High Frontier: Human Colonies in Space and the 1986 Challenger disaster. This isn’t a coincidence.”

Algorithms, Automations & Augmentations

• Midjourney, the AI image generator, is developing a full-body ultrasonic scanner. This is some LLM-generated shark jumping right here, but supposedly it “is real, and the company is even building spas where you can find the machines and get scanned.”
• “Digital Colonialism”: U.S. demands to access Africans’ data raise privacy, sovereignty concerns. “U.S. officials are demanding access to the health data of millions of Africans as a condition of giving billions of dollars in lifesaving aid to African countries. … Experts said the deals are vague and lack standard language to guard personal data from being exposed, misused or commercialized without people’s consent.”
• Have a Thorny Medical Question? Your Doctor May Be Using A.I. for That. “Dr. Gavin soon learned that they were far from outliers. A third of Mount Sinai’s 9,000 doctors were already regular OpenEvidence users, the health system’s executives found out in a meeting last year with the start-up’s leaders.”

Built, Biosphere & Breakthroughs

• How the Squamish built Senakw. “Just outside downtown Vancouver, the Squamish Nation is building one of the most ambitious and unusual housing developments in the world and getting rich in the process.” I don’t think “getting rich,” after being robbed for centuries is the right framing, but great story.
• The climate friendly city is a bullseye. “In urban regions with a single, defined core such as Berlin and Boston, the best place to increase housing density is in a ring around the center where there’s room for infill development but the city center is still easily accessible. In the case of Boston, for example, this zone occurs about 10-21 kilometers from the city center.”
• Mexico just showed off a new extremely cheap, government-backed EV. “The Olinia Uno is a planned six-passenger car that can travel up to 125 kilometers (77 miles) on a single charge. It is expected to go on sale next summer for about 150,000 Mexican pesos or roughly $8,500, according to a press release.” (Via The FLUX Review.)

Asides

• Extreme dynamic symmetry enables omnidirectional and multifunctional robots. “We build a physical 20-leg Argus variant that achieves near-extreme dynamic isotropy and demonstrates orientation-invariant locomotion, agile traversal of cluttered and deformable terrain, rapid self-stabilization, and resilience to partial actuator failures. Its distributed sensing further enables omnidirectional perception and object interaction during continuous motion.”
• Les Étoiles d’Ivry. “The term ‘brutalist’, which has been used to describe the work of Renaudie, seems somewhat hasty to define the urban richness which this project harbors. Nearly forty years after its construction, the freshness of the proposal remains intact, even if people who live it often don’t know the futuristic aim of the project.”
• DNA from 2,000-year-old grape seeds points to origins of modern winemaking. “Found in ancient wells in Tuscany has enabled scientists to map the most extensive genetic history of grapevines recovered from a single site. The findings revealed that vineyards of the Roman era formed part of the empire’s sophisticated agricultural network that might have influenced the development of modern winemaking.”