Versteeg’s experiment reveals the tensions between coherence and serendipity, between visual and thematic, and between fungible and non-fungible.

Since March 13, 2023, an algorithm has been generating digital collages, or ‘moments’, every 15 minutes using trending stories and images gathered from the internet. A moment is available only for a limited time (2 hours 15 minutes to be precise) before it is lost, unless it is ‘claimed’ by a collector who can make it permanent as part of an NFT collection. About 35,000 moments will be generated by the end of the project. 750 of them can be claimed. This is the summary description of the ongoing NFT project For a Limited Time (henceforth FALT) by Siebren Versteeg, launched through Arsnl. We are nearing the end of the one-year claiming period, which gives us plenty of insight into the behavior of the generative algorithm and factors that shape crowd curation. Some reflection is thus warranted.

Below are my reflections as a collector who has followed the project and claimed many moments throughout 2023 (full disclosure!). I will first briefly describe the general visual style and will then interpret the project in terms of generative art and dynamics related to crowd-sourced curation. There is much to cover, so let's start.

A FALT moment starts as a blank canvas that is gradually populated with layers of images scraped from the internet. The images go through various combinations of cropping, scaling, rotating and silkscreen effects before being placed onto the canvas. A digital brush paints over the canvas, sometimes tracing the visual edges, other times forming freestyle strokes.

The blank canvas that forms the background looks exactly like a linen canvas. The digital brush paints at the pace of a human painter and in a way that resembles physical paint. There is nothing that prevents all this collage work from happening at once, but the act of creating each moment is turned into a 15-minute performance viewable via livestream. This is digital emulation of analog media carried to the extreme. It is almost ironic. Cropped images are the main contrast to all this emulation. Their jagged edges are visibly digital and occasionally pixelated. Their crude digital processing goes counter to the recent trends in image recognition and gives the work a hand-coded feel.

Various combinations of silkscreens, cropped images and brushwork lead to a broad range of outputs. Some moments feature only brushwork over an empty canvas, others are dominated by a single image, and still others combine cropped images and brushwork in a way that leads to abstract textures. The pairing of found visual material with abstract brushwork evokes Rauschenberg, but the collection as a whole is hard to describe in terms of a single artistic intent. Some moments look like Dada commentary collages. Some moments reveal the painterly potential of photographic images by blending them with brushwork, as in Gerhard Richter's Overpainted Photographs.

Imagine you are tasked to fill an empty canvas with 100 circles. If you completely randomize the process, you will have many outputs that look similar, with circles evenly spread out across the canvas—just like flipping a coin 100 times would lead to roughly the same amount of heads and tails. To prevent most of your outputs ‘reverting to the mean’, you can code in some conditional rules, such as by partitioning the canvas into certain regions with higher circle density. But hey… in the process of parameterizing these conditionals, you introduced many prior assumptions about what variety is supported or excluded. You precluded the occasional interesting output that falls into the long tail of the distribution.

This tradeoff between providing structure and leaving space for serendipity is well-known to generative artists. ‘Long-form’ generative art has to provide sufficient guardrails to support enough visual variety and coherence with any random seed. Short-form works such as QQL, on the other hand, deliberately remove some of these guardrails and rely on crowd curation or other means to explore edge cases.

FALT clearly leans towards serendipity. Sure, one can observe some conditional factors. There are palettes that specify the brush colors and silkscreen effects. Brush strokes can loosely follow the image edges. Yet it is also possible to watch the algorithm embellish a location with images and delicate brushwork, only to cover all this detail by slapping a whole image on top. The lack of extensive parameterization creates many outputs that revert to the mean: clustered at the center of the canvas, a haphazard collage with a random mix of colors. Yet it also creates many surprising outputs that allude to some happy coincidence. Collectors’ task is to find these occasional gems. What one regards as a gem, however, is highly subjective and goes beyond the visual.

FALT is of course not just a visual exercise. One can argue that the project’s situatedness at the present time is the whole point. Now, you can state where you stand on the scale of visual formalism, but the problem is real for collectors who need to decide on the relative importance of visual and thematic criteria on the go: Do you claim a ‘great’ visual composition with an ‘uninteresting’ theme or a ‘mediocre’ composition with an ‘interesting’ theme? Trending stories cover pretty much everything: sports, politics, celebrities, health, and so on. Suffice it to say that there is a clear bias towards Elon Musk, Donald Trump and crypto.

It would be too quick to assume that curation choices are made purely based on personal taste. FALT claims happen within a social context and there are different communicative affordances to the visual and thematic. Visual material is sensually rich, but the resulting aesthetic appreciation is hard to get across to another person in words (i.e., literally). In contrast, it is relatively straightforward to state that a moment features Elon Musk. It even becomes hard to express it in visual terms once his face is recognized; the literal has some tyranny to it. Landscape and abstract art partly emerged to free form and color, but the literal finds a way. Consider the problem of deciding what parameters to encode as part of metadata in abstract generative art. Once a visual parameter is literally encoded, it also dictates a common meaning and a rarity ranking. Why is Ringers #879, aka ‘The Goose’, so famous? Partly because the figurative goose makes it memorable and easy to get across. It has become a meme.

Do people claim moments featuring Musk and Trump partly with meme potential in mind and even with expectations of future value? We cannot be totally sure, but can at least identify different factors at play when a collection is created entirely by an artist or by different collectors who are motivated not only by their individual preferences but are trying to guess what other collectors will deem valuable.

NFTs make art work available to anyone with an internet collection. Yet this availability also separates the artwork’s artistic form from its commodity form, elevating financial considerations over aesthetic ones when it comes to collecting. Deeper collector engagement, for instance through crowd curation, brings some rebalance. There is certainly an aspect of joy and personal connection to curating moments from FALT’s diverse outputs. The same aspects, however, make the claimed tokens much less fungible.

Now, don’t let the term ‘non-fungible token’ confuse you. NFT market LOVES fungibility. More fungibility translates into better liquidity, and arguably into a better store of value. In general, more uniform and coherent a collection is, the more fungible its individual items are. Coherence often comes built-in with long-form or artist-curated work. Crowd-curation introduces uncertainties.

All this matters as FALT is also situated in an economic context, subject to the ups and downs of the NFT market. FALT collectors need to decide not only on what to claim but also on whether to claim or sell. Claiming a moment irreversibly turns an open canvas, which is fungible and full of potential, into a specific moment. Fungibility explains why open canvases have generally commanded a premium over claimed tokens, and why collectors have been slow to claim their moments. This optionality is now decreasing and claims are picking up as we approach the claiming deadline. The project timeline and financial considerations influence claiming activity just as much as the news cycle.

FALT is framed as an ‘experiment’, but it is not conducted in a lab. There are no simulated test environments. Collectors engage by seizing the moment instead of adjusting parameters. The fleeting nature of the moments and the project deadline brings a lifelike time pressure. Visual criteria competes with thematic. Crowd curation in FALT is shaped by this particular setting. There will be a diverse set of moments once the claiming period ends. This diversity might frustrate a visual formalist or those who particularly value fungibility. A more faithful approach would be to view all that is claimed—and not claimed—as a record of the process. After all, this is an experiment and it is fascinating.