Giovanni Kanter goes to at least one art museum whenever he visits a new city. That’s on top of his regular exploration of galleries in New York, where the 25-year-old ICU nurse lives and studies. He typically discovers new artists and gallery openings through word of mouth, though he has recently started curating his social media feed to serve relevant content (he’s currently in what he terms a “surrealism phase”).

He was raised on visual art; he's the son of painter Marianne Gargour and nephew of mixed-media artist Heather Hutchison and painter-printmaker Mark Thomas Kanter. His “exposure-oriented education” featured childhood visits to galleries and shows in and around Dallas, where he grew up, and in France, where he spent time with his grandparents.

But Kanter is an exception. Many other young people have a less immersive artistic education, perhaps shaped by occasional school trips or family visits to museums. If they haven’t internalized an instinct for what Kanter calls “moving through the art and absorbing it,” the museum experience is central to how they learn about and understand art.

Museums take this educational role seriously, and many institutions understand the need to sustain and broaden their appeal. “They used to be these big ivory towers of knowledge and history and thought and art,” said Hilary Knight, director of Change& in London and former director of digital at Tate. “Museums themselves are changing their role to be centers for discussion and encounter and to be much more participatory.”

To aid this transition, museums and other cultural institutions are increasingly using data to inform decisions about how to present the works they house. Amsterdam’s Van Gogh Museum (VGM) has embraced this challenge head-on, instituting a yearslong collaboration with academics from leading business schools to explore how using data and analytics can improve visitors’ experiences.

More extensive data use by museums has enabled a new era of individually tailored experiences, and new waves of technology have impacted everything from how art is displayed to how fragile art is preserved. But this active role in shaping museumgoers’ experiences raises questions about how museums shape and nurture cultural engagement at a time when both consumption habits and the technologies surrounding them are evolving rapidly.

Artistic institutions are amassing more and more information about us, and they can now use it to guide nearly every aspect of modern artistic exposure. Museums have always conferred cultural legitimacy and served as tastemakers, but to what extent have they also become—even inadvertently—more strategic stewards of our evolving tastes and interests? And where do we go from here?

Data-informed layouts

As you navigate a museum, data can be generated on which paintings you visit, how long you spend viewing each one, and even where your eyes linger.

Such data can lead to better layout decisions, according to a new study by Ali Aouad, a professor at MIT’s Sloan School of Management; Abhishek Deshmane, a professor at Georgia Institute of Technology’s Scheller College of Business; and Professor Victor Martínez de Albéniz of the University of Navarra’s IESE Business School. In their paper, published in the April 2026 print issue of Management Science, the researchers show that visitors respond to both the physical location of works and their placement on digital guides.

According to the authors, a visitor’s next move—whether to view another painting, change floors, or even end a visit—can be analyzed in terms of the perceived utility, or benefit, of each option, likely weighed subconsciously.

The researchers randomly selected 25,000 out of 1.5 million visits to the VGM between 2019 and 2021 and extracted data from visitor interactions with digital multimedia guides, along with the location and artistic characteristics of the museum’s collection, which contains over 4,600 works (of which approximately 200 are on display at a given time). They used this data to gain insights about visitor movement and identify “optimal” locations for pieces—those that collectively maximized viewer engagement, as measured by the expected number of pieces seen.

The authors’ model, named pathway multinomial logit or “pathway MNL,” does not reveal causal effects, they caution, but the associations they uncover are informative for museum professionals.

Multimedia guide users largely visited recommended artworks in the suggested order. Lisa van den Bos, product manager for digital education at the museum while the study was underway, expressed surprise that even subtle ordering changes on the guide impacted traffic patterns. “We actually have a lot of power in how we move people through,” she said.

Visitors sought variety in time period and size but consistency in subject and theme, the team found. Time pressure also affected choices; as time elapsed within a given visit, people were more likely to prioritize seeing masterpieces.

Perhaps surprisingly, the data revealed that congestion can increase engagement. “When there are more people, at least up to a certain number...people are more likely to go and view artworks that they probably wouldn't have seen,” said Deshmane. Crowding around a painting may also be seen as a signal of quality.

Rearranging permanent collections can be tricky due to space or thematic constraints, but museums can often experiment with exhibition or gallery spaces, where shorter time frames and lower infrastructural requirements allow easier changes, said Deshmane.

Rather than actually moving pieces, the researchers used a simulation to generate new layouts and predict likely visitor reactions. Based on their simulations, they found that moving highly attractive art from leaky spots (e.g., stairs or lifts) and strategically swapping pieces could lead to visitors seeing more art—and staying for longer—without undermining curatorial aims. Additionally, data-informed planning of digital layouts, even without accompanying physical moves, may help visitors understand art through new perspectives, said Gundy van Dijk, the museum’s head of education and interpretation.

Designing layouts in this way is atypical, according to the authors, but it can enrich the cultural and educational value of visits. When tested on data not used to develop the pathway MNL model, it achieved 63 percent accuracy in predicting transitions between pieces; beyond this improvement over baselines, the authors verified that their results hold even when the modeling approach or assumptions change.

The model was also used to predict visitors’ likelihood of reaching specific pieces (81 percent accuracy) and visitor departure rates (96 percent accuracy). The authors noted that more general forms of the model may be needed for settings with significantly larger collections or more complex physical layouts. Beyond museums, they feel their approach has direct application to other sequential experiences that blend the physical and digital, such as in entertainment, education, and retail.

Following the sequential layouts study at VGM, the research team conducted a follow-on study between 2022 and 2024 that tested different approaches to combat visitor fatigue and information overload, providing digital nudges by experimentally varying actual visitor offerings on the multimedia guide.

A “gold mine” of data

This study took existing but underutilized data, mining it for insights. Audio guides, available since at least the 1950s, remain the main way to supplement a primarily visual experience with minimal distraction, said Sandy Goldberg, who creates content for museums, including the Van Gogh, and described a “pendulum swing” that has seen museums revert to audio after exploring successive waves of more interactive technologies. “Study after study showed that what people wanted was what was called an ‘in-the-pocket’ experience, where they could take their device and just literally put it in their pocket and not think about it anymore,” she said.

Though digital guides—often with multimedia capabilities supplementing audio—are now widely available, most visitors don’t use them, though their uptake has grown over time, said van den Bos. The VGM's multimedia guide at the time of the study was a standalone device with hardware from Imagineear and software maintained by Tapart, a creative tech agency specializing in digital storytelling. Other museums allow visitors to access visit-related content on personal devices.

Throughout 2023, 36 percent of VGM’s average 4,600 daily visitors booked multimedia guides (up from the 25–31 percent typical during the study). That’s high among peer institutions; sector-wide, uptake during the same period was typically closer to 10 percent for paid tours, said Sietze de Jong, creative and managing director of Tapart—which also partnered with The British Museum on the development of its audio app.

Museums rarely use digital guide data—a “gold mine,” said Deshmane—for decision-making. “Every museum requests us to do analytics and include it in their audio guide… but when we ask them if they reviewed the data, they almost never have,” said de Jong, citing limited institutional capacity for data analysis.

He noted that the situation has begun to change in recent years due to investment in in-house capabilities and external collaborations: “Digital teams are increasing in size, which is a positive development.” Museums also seek to ground longer-term strategic decisions more deeply in data and to measure and communicate their progress using data, according to a 2023 report by McKinsey & Company. That comes in response to expectations from trustees, funders, and governments.

In 2025, the VGM announced a redesigned version of its multimedia tour, with hardware powered by Samsung and software built and maintained by NOUS Digital. The design of the guide was informed by the research teams’ insights into visitors’ actions and desires when using digital guides.

A “duty to expose”?

Museums can use data to manage traffic by adjusting opening hours and providing navigational aids, but digital tools also enable subtler interventions. For example, recommendations can be discreetly varied throughout the day based on real-time congestion levels, said Goldberg.

Perhaps the biggest opportunity created by detailed data is the ability to provide personalized recommendations. Supplementing traditional curation with analytical methods, more recently including machine learning and artificial intelligence, has become more common. The use case for personalization is especially clear within a single institution, but cross-museum applications can now also recommend that visitors explore artwork at other museums within a partner network.

The study’s explicit inclusion of visitors’ perceptions of utility, and particularly the idea that visitors seem to actually respond to recommendations, raises the question of how museums should wield this power over the visitor experience. Given their educational role, do they have a duty to expose visitors to a diverse group of works? Would doing so actually influence artistic “enjoyment”?

While the answer is not straightforward, experts in the field generally believe that curators should make intentional recommendations rather than simply predicting preferences. “Showing people artworks that you think they're going to like actually is a disservice,” said Knight. “One of the roles of art is to challenge you and to present you with things that you might not like or might not realize you like.”

There’s also a risk of recommendations becoming self-perpetuating—“reinforcing the canon” of what’s considered important, added Knight.

Certain works of art are so renowned that they enter the consciousness long before young people are exposed to any formal art education. These pieces often occupy premium real estate within museums, receive prominent billing during exhibitions, and are the most sought-out for loans and transfers among institutions.

From a business perspective, this emphasis is understandable; these big-name pieces are huge audience draws, both for locals and traveling visitors. However, as museums increasingly use data to improve operations and introduce more personalized offerings, visitors may see themselves encouraged to explore less popular corners of the catalog, whether due to congestion-induced redirection or preference-based recommendations.

We don’t yet know whether these shifts will be significant enough to result in long-term changes to the process of taste formation for individuals or the crowning of “classics” at the societal level, but it seems clear that when a cultural experience truly resonates with a visitor, engagement doesn’t end at the museum exit.

Decision makers must strike a delicate balance as they redefine their roles as custodians of culture in an era when technology empowers them to do so much. Not all visitors enjoy having their experiences shaped externally, of course. Kanter typically prefers “more of an organic experience,” he said, though he noted that he sometimes uses museum audio guides when seeking to gain a new perspective.

Deshmane argued that the solution is to provide an “assortment” of suggestions that enables informed decisions and caters to various tastes and engagement habits. Depending on the context, these suggestions can be as subtle as presenting one piece of artwork before another on a digital guide or as explicit as generating a push notification or other nudge. Deshmane noted that “suggestion assortments” could be organized by how much time visitors have or by narrative threads.

“You as a museum can do a lot more with one physical layout… if you have these digital assortments of layouts,” said Deshmane.

A single “precision” cultural experience has potential downsides. While too many options can empower visitors, overpersonalizing could also create a sense of loss. “You have to make choices [such] that people don't feel like they're missing a lot if they're choosing,” said Rianne van Dam, project coordinator for Digital Education at the VGM.

Kanter is generally fine with museums using the vast and growing amounts of data at their disposal, as long as they address privacy concerns and think carefully about when explicit disclosures are needed. “A lot of tracking is going on in the world…I think it’s interesting that it applies to art,” he said.

Leading with creativity

Leveraging technology is key to staying relevant amid changing expectations, said Birgitte Aga, head of innovation and research at MUNCH, the Oslo museum dedicated to the work of Norwegian painter Edvard Munch. “If we can't engage and create museum experiences that are both physical and virtual in a way that the next generation of users find engaging, then we're not going to be a museum in the future.”

The most visible uses of technology involve transforming artistic presentation: virtual reality and augmented reality have been used to reveal layers within art, and AR can restore historical context such as the original colors of ancient artworks, said Goldberg. “Seamless” experiences enabled by tools such as smart glasses or wearables can reduce device friction and enable interactivity.

Some applications seem less likely to see widespread adoption soon. While a handful of museums and standalone services have experimented with conversational and generative AI to help create the experience of "hearing from" or “speaking” to artists, the sector appears to be treading carefully with these tools due to concerns about both accuracy and artistic integrity. Institutions are also shying away from AI playing a curatorial role.

Beyond presentation, AI has also been used to analyze collections databases, visitor records (including survey feedback and social media comments) and operational data, said Elena Villaespesa, head of digital analytics at the Thyssen-Bornemisza National Museum in Madrid and former digital analytics lead at the National Gallery of Art in Washington, DC.

AI can also feature in conservation and virtual restoration efforts, said MUNCH’s Aga, and is helping increase accessibility. For example, machine learning models can learn artistic characteristics from collections data to generate automated textual descriptions. Such techniques can also support personalization efforts.

MUNCH’s New Snow exhibition, available in two waves in 2024 to 2025, allowed visitors to draw their own images and then matched them to the closest item among Edvard Munch’s sketches, thereby creating a channel of interactive engagement with a collection of drawings that are brought out infrequently due to their fragility. Over 70,000 visitor drawings were created through the project.

These examples are far from comprehensive; the range of application areas is expanding rapidly given the rate of advancement in AI capabilities over the last few years. Museum directors are taking note—and sharing notes—as they look ahead. Digitization, personalization, and AI strategy have become staple topics on the agendas of leading cultural industry forums.

Museums are applying discretion in evaluating the dizzying array of options; according to Villaespesa, they’ve become “more mature in the use of digital.” She advises continuing to ground decisions in a user-centered perspective. “Experimentation is great... with evaluation behind it," she said. "I think museums should test the different trends and technologies available, but I’m always thinking from the user perspective… make sure that there’s a motivation, a need, a type of user that will use it. It’s not just ‘let's do technology for technology’s sake.’”

For now, given widely varying consumer preferences, museum staff seem to view personalization as an imperative. To van Dam, tailored experiences serve the museum’s goal of disseminating van Gogh’s story. “I don't think every visitor wants the same,” she said. “So the key, really, to reach our goal… is to personalize.”

Ultimately, though, museums don’t want to become optimization machines. “They are creative spaces in the end,” said van den Bos. “It will always be a place where the story is leading and the creativity is leading.”

Researchers say that data-driven and creativity-driven approaches are compatible—and that insights from tools such as digital guides can lead the way. “Use this tool first to create compelling narratives,” said Deshmane. “And then use the data that is generated for future purposes of creating more compelling narratives with the physical space as well.”

Mureji Fatunde-Iloeje is an academic and writer interested in companies, industries, and consumers. Her writing, which has appeared in Bloomberg, WIRED, and MIT Technology Review, explores climate and the energy transition, social and economic policy, and the modern consumer experience.

Read full article

Comments

Right now, sockeye salmon season is peaking in Alaska.

Some fishing vessels in the Cook Inlet have welcomed a new crew member, one who doesn’t complain about swells, eat, or even sleep—yet is able to harvest salmon more rapidly and humanely than any fisherman ever has.

The newcomer is PSDN-S, short for “Poseidon,” a robot built by a startup called Shinkei Systems.

PSDN-S’s process sounds coldly mechanical but is in fact far more compassionate than the traditional methods, which often leave fish to slowly asphyxiate on deck. As crews aboard Shinkei’s partner boats navigate the inlet—notorious for its treacherous tidal waters—they hand feed salmon they’ve caught into the robot, a tall metal box bolted to the deck. Inside, computer vision helps the machine identify the species, locate a half-inch spot on the skull, and drive a spike through the hindbrain before severing the gills. The whole cycle takes about six seconds, depositing the fish in ice before it can thrash.

“We’re basically doing the equivalent of surgery on these fish in a very difficult maritime environment,” Shinkei’s cofounder, Saif Khawaja, said as we watched footage of the process together recently over a Zoom call.

The Shinkei model

PSDN-S replicates ike jime, the exacting slaughter technique long required for fish served at top sushi restaurants in Tokyo, Los Angeles, and New York. The method is hands-down the most humane way to kill a fish. Not incidentally, chefs at high-end restaurants including Daniel and Atelier Crenn believe Shinkei fish simply taste better.

Shinkei is an El Segundo, California-based company that Khawaja founded four years ago in New York, before coming to believe that all the serious engineers are in California. (Significantly, and in keeping with the company ethos, the Japanese word shinkei translates as “sensitive.”) Shinkei installs the PSDN-S robot for free and pays fishermen a premium for their catch. In exchange, it takes full possession of the fish, keeping it off the open market. The company then processes, distributes, and sells the seafood itself, under its own consumer brand, Seremoni.

Until this summer, Shinkei was using this method to process just a handful of fish types that most American consumers don’t eat very often: black cod, black sea bass, red snapper, vermilion rockfish. This was by design, Khawaja tells me: “We chose to avoid the most popular fish in the American diet” to show “what we could do with beautiful, artisanal noncommodity fish.”

Now, PSDN-S is being deployed toward the most popular fish in the country, which Shinkei says will begin rolling out at restaurants and retail partners in the coming weeks. Last year, 200 million salmon were harvested in Alaska during the brief fishing season, and Americans put away 200,000 metric tons, making them the planet’s number one consumer.

The case for PSDN-S

PSDN-S’s predecessor, the full-sized PSDN, has won attention for its innovations in animal welfare and sustainability (including a spot on Fast Company’s 2026 list of the Most Innovative Companies in Food). The robot helps mitigate the guilt that people feel about eating creatures that feel pain. But Khawaja argues that an equally big selling point—and why backers including Founders Fund have invested $22 million so far—is the fish’s better shelf life, and what that makes possible: More people can have access to fresh, high-quality seafood.

When fish are harvested aboard typical commercial vessels, the process causes fish to release stress hormones, and lactic acid leaches into the flesh, priming it for duller flavor, mushier texture, bacterial growth, and a shelf life of five to seven days for most species. Shinkei says that its fish, by contrast, can last for two to three weeks.

Estimates suggest that as much as 35% of the global fish harvest goes to waste because it spoils. The most recent snapshot, calculated in 2023, found the U.S. loss rate to be 23% of the total edible supply. A recent study on reducing seafood loss, conducted by researchers with the UN Food and Agriculture Organization, Cambridge University, and a university in China, explained that the best way to feed our future selves isn’t by catching more fish (we sort of . . . can’t), but by reducing our excessive waste.

Salmon was never meant to travel quite that far

There’s an industry quirk that exacerbates the fish-waste problem: Much of the fish sold in U.S. grocery stores is first shipped to Asia for processing—e.g., the messy, grisly work of gutting, scaling, and filleting. This arrangement has drawn fire in the U.S. from across the political spectrum, rankling everyone from China hawks to labor activists to food-waste watchdogs.

In March, Shinkei more than doubled its domestic production footprint with a 16,000-square-foot facility in Tacoma, Washington. It’s here that both Shinkei and Founders Fund are betting the startup can bring the U.S. fish supply chain back on shore by becoming a vertically integrated harvester and processor, deploying robotics from boat to plate for a virtually inexhaustible roster of species. (Khawaja says more challenging fish shapes, like of swordfish and flounder, are off the table, at least for now.)

Partnering with Alaskan fishers is a big first step in that direction. Alaska accounts for roughly 60% of U.S. seafood, nearly 6 billion pounds a year. Yet about two-thirds of that catch is exported. Meanwhile, as much as 85% of the seafood sold in American grocery stores has been imported, a carousel dance where we ship our fish to other countries, then buy back theirs—about half from Chile, much of the rest from Norway and Canada.

The sockeye now coming off boats, prized for their deep-red color and robust flavor, are an early test of whether the model can handle a fish that pretty much everyone eats. Compared with boats harvesting something like black cod (8 to 10 million per year) or vermillion rockfish (closer to 200,000), Khawaja says, “The volume that they’re pulling out of the water is absolutely mental.”

Who will speak for the fish?

Khawaja likes to trace Shinkei’s humane-slaughter idea back to an essay by the Animal Liberation author Peter Singer. Khawaja grew up in Dubai taking family fishing trips and later dropped out of grad school at Penn after winning a prize that gave him $100,000 to chase what became Shinkei. But it was Singer’s “If Fish Could Scream,” published in 2010, during a cultural reckoning over fish ethics, that stuck with him. Commercial fishing inflicts “an unimaginable amount of pain and suffering” on fish, Singer argued, but it hardly registers for us because fish “cannot give voice to their pain.”

Obsessed, Khawaja tried building a sensor he could attach to a fish’s body that would trigger noise intensifying as its stress rose internally—a technological scream. The problem, he quickly realized, was that nobody wanted to listen to screaming fish. He instead focused on the commercial end of things, the fact that treating fish better also made them taste better and sell for more.

To train PSDN-S, Khawaja and Reed Ginsberg—Shinkei’s chief technology officer and cofounder—asked fishermen to bring in around 50 species of live fish, from little 1-pounders to 25-pound whoppers. They used cameras to capture images of them from every angle, so that Shinkei’s AI models could recognize them if they’re placed into the machine at a funny angle, or even upside down.

Sorting salmon was easy. The hard part was installing a fish-killing robot on vessels the size of a city bus navigating 10-foot swells. “Boats in Alaska for wild salmon are limited to 32 feet, as the maximum size,” Ginsberg tells me. The PSDN-S is the result of Shinkei’s listening to fishermen who were eager to join the fleet, but had a sort of reverse-Jaws request first: “You’re gonna need a smaller bot.”

Ginsberg had joined Shinkei from SpaceX, where he led the Starship Avionics group and designed hardware for Star Shield, the government version of Starlink. His job was to build things that could survive violent shocks, thermal radiation, extreme temperature swings—stuff that “goes really, really far away and you don’t get to access it again.”

Fishing boats, it turns out, aren’t so different. Engines run hot, waves cause intense vibrations, decks bake in the Gulf sun and freeze in the Bering Sea. “Corrosion is even worse than aerospace,” Ginsberg adds, calling the ocean “probably the harshest environment I’ve designed against.” Because PSDN-S also travels to places he can’t access, Shinkei patches in software updates remotely over Starlink. “It’s like having a Tesla,” he explains. “Plus, now they’ve got Wi-Fi on the boat to watch Netflix.”

The ‘boat to plate’ movement

Since Shinkei-caught fish became available commercially two years ago, under the brand name Seremoni, it’s been a hit with chefs across the country; the fish has appeared on plates of several dozen Michelin-starred restaurants. Chef Dan Barber—whose restaurant Blue Hill at Stone Barns is famous for sustainability (it sources scallops from a husband-wife operation in Maine whose team free dives for them and then ships the live catch overnight in cans)—has called Seremoni “the platonic ideal” of what fish should taste like. It’s available to shoppers at Erewhon, Wegmans, and Fresh Direct. Last year, it debuted at Tokyo’s Toyosu fish market, marking the first time wild-caught black cod was sold there in 17 years.

Now Shinkei is pushing its technology farther. It’s preparing a new sensor, dubbed NERA, that scans fish to assign each one its own date when it will reach optimal quality—what Khawaja calls its “ripeness.” It’s a disorienting way to think about an animal pulled from the ocean. But he showed me a chart plotting the data as if we were discussing a piece of fruit. “Think of it like a banana,” he offered. “You don’t want it green, and you don’t want it black. You want it yellow.” For regular fish, this window is incredibly brief. But the Shinkei method widens it considerably.

For now, NERA is still appliance-sized, just like PSDN 1.0 was, and it runs only in Shinkei’s El Segundo factory. But Khawaja’s ambition is to create sensors that can be used anywhere along the supply chain. “If we shrink them down enough, we can put scanners at the dock where fish are coming in, and at the grocery store where the labels are being printed,” he says.

He says that if you go eat the fanciest $500-a-person omakase, the chef will have watched the fish carefully, sometimes for days, to see when the texture begins to stiffen, an indication that the flavor is at its peak. Khawaja sees no reason the same shouldn’t be true for fish sitting in the average American household fridge.

I wanted a clock that, annoyingly, didn't seem to exist.

Since childhood, my bedside clocks have been a series of red, seven-segment LED clock-radio specials from Walmart or Target. They are invariably cheap, simple, and long-lived—but they require manual intervention at the start and end of Daylight Saving Time and whenever the power flickers. After a recent power flicker, as I found myself standing by the sideboard holding down "TIME" and mashing the "HOUR" button, frustration boiled over, and I thought to myself, "We're a quarter of the way through the 21st century. There has to be a better way!"

My perfect clock would be self-setting. It would offer auto-DST adjustment (or not, depending on how this bill fares!). It would manage drift and always show the exact sub-second time. It would show that time on a red seven-segment display—not blue, not green, not yellow, and absolutely not white. And I shouldn't have to install any privacy-destroying garbage apps to make it work.

Simple? No. While many bedside clocks meet one or perhaps two of these requirements, I couldn't find anything that meets them all. Battery-backed self-setting "atomic" clocks that get their updates via the cosmic ether have been a thing for years and get me most of the way there, but damned if I could find one with a red seven-segment display that I liked (maybe someone else's search kung-fu is better than mine?).

For a time, despair won out. But as I closed dozens of browser tabs featuring fruitless searches and close-but-no-cigar product pages, I thought to myself, "Wait a second. I've got a 3D printer. I'm, like, smart and stuff. Why not buy a seven-segment display and make my own clock?"

And so, standing on the shoulders of giants stacked up so high that I could practically touch the Moon, I did.

This is the clock, doing clock-y things. Credit: Lee Hutchinson

O brave new world, that has such clocks in't

For folks who aren't interested in several thousand rambling words about process, here's the finished repo. It contains my bill of materials with prices and purchase locations, the software, and the 3D printer files.

There were two potential paths this hilariously overengineered weekend project masquerading as a clock could have shambled down. One, the path not taken, started with an Arduino or Arduino-like microcontroller. The other began with a Raspberry Pi or Pi-like computer-y thing. I went with the Pi, variously using both a Raspberry Pi Zero W and Zero 2 W.

My reasoning was that a Pi gave me the security blanket of a Debian-based operating system, complete with Wi-Fi and NTP for the "the clock sets and updates itself" requirement, along with the usual Linux remote management routine I already know.

Picking a seven-segment display was easy: Adafruit makes awesome clock-face style LED displays with 1.2-inch high numerals, and it sells a kit that bundles the display I want with a "backpack" board containing the HT16K33 controller needed to drive the LEDs. I ordered three and ended up using all of them for testing, assembly, and figuring out how to solder.

The display, from Adafruit's product page. Credit: Adafruit

Ah, yes, soldering. I'd never done it before, but the seven-segment display had to be soldered to its backpack board, so I grabbed a baby's-first-soldering-iron kit from Amazon and a roll of 60/40 solder. (I also had to buy a desktop magnifying lens, because as I found out when I got in there, these old eyes can't focus up close like they once could.)

Setting aside the matter of the clock's enclosure—I felt sure that someone else had already designed a 3D-printed case compatible with the Adafruit display, and I was right—I sat down with my new Pi Zero and began poking at the software it would have to run in order to speak clock. I quickly realized I was in over my head. As I've said on these pages so often, I put the "ops" in "devops"... somebody else needs to bring the "dev."

A clock past the wit of man

The RPi image loader got me going, and I was able to log into my Pi Zero. After thinking about things for a bit, I distilled my software requirements down to a list:

• The clock host should be LAN-only and not accessible from the Internet
• The clock host should get its updates from a LAN-only apt mirror
• The clock host should get its NTP sync from a LAN-only NTP server
• The clock service should be a systemd service running unprivileged under a dedicated service account context
• The clock service should use the system time, so the host OS handles NTP and keeps us in sync with whatever DST is or isn't doing
• The clock service should be able to turn the display on and off on a schedule so it's off for most of the day when I'm not in the bedroom
• The clock service should also be able to brighten/dim its display on a schedule
• The clock service should have some way of being controlled via the CLI for terminal connections, too
• The display should be controllable via HomeKit, because I live in iOS-land
• The clock service and its dependencies should be installable via a single script
• Once installed, everything should be deployable so I can push updates if needed rather than having to log in and reinstall

Many of these items were easy and well within my typical ops wheelhouse. I fell back on good ol' systemd timers and services for a big chunk of things—I'm actually coming to quite like systemd, God help me. The LAN NTP and apt-mirror sources already existed (I know, I know, I should be using apt-cacher-ng instead). The deployment pipeline would use Gitea actions and would be more or less exactly like one I'd already set up for another project, so I cribbed from Past Lee there. HomeKit integration looked like it was going to basically be a bolt-on thanks to HAP-python.

But I started to worry when I looked up examples of how to communicate with the clock display via I2C. My much-atrophied Python muscles were already straining and would absolutely not be able to meet this challenge. This was the point where the project stopped feeling fun and started feeling impossibly hard.

So I shoved the coding tasks off onto an LLM.

"You taught me language; and my profit on't is, I know how to code"

Seeking an LLM's help when one can't really verify the outputs can be fraught, but fortunately, I recall just enough Python to follow-flail my way through the results, with the help of the inline comments. Claude Code proved more than capable enough to tackle this project—first with Opus 4.8 and then later with the new fancy Fable model, whose world-ending powers I harnessed and used on what is probably in truth an intern-level coding project.

It was a bit like unleashing the full power of the Death Star on a mosquito, but it definitely did the trick. The result was a tidy collection of Python files and a nice little test suite. The LLM did such a good job that I also had it do the HomeKit integration, the install routine, some specifics around the deployment pipeline, and most of the repo documentation.

I know this admission may be anathema to many among the Ars commentariat, but it is what it is—without the LLM, I wouldn't have finished the project. I would have gotten annoyed, angry, or just tired of endlessly reading StackOverflow posts criticizing what I'm trying to do for being dumb and wrong.

The project workspace. This is a slightly different version than the public GitHub repo, with some Lee-specific defaults and a Gitea action. Credit: Lee Hutchinson

The application side is a proper systemd service, and it listens for commands from HomeKit; it can also be controlled locally via a Unix socket if you want to make the display do things from a terminal session. The service runs under a non-privileged service account. Deployment works via a Gitea action, whereby I push a tag to my local Gitea repo and a runner creates a release artifact and shoves it onto the Pi via a separate local service account that can only do deployment-related things. (The deployment workflow is included in the project's GitHub repo as an adaptable template, in case someone out there has my exact setup and wants to use that as well.)

Full fathom five Autodesk lies

On the physical side, I did indeed find a Creative Commons-licensed 3D-printable enclosure designed around the same Adafruit display I was using, but it wasn't quite right.

Modifying the model meant doing battle with the absurdly user-hostile nightmare that is Autodesk Fusion, so I girded my loins and dove in—and hit another wall. Parametric modeling, especially when weighted down with decades of AutoCAD's stupid UI/UX choices, was even harder than programming.

Autodesk Fusion, we meet again. (Thanks to Boosted for <a href="https://www.printables.com/model/550564-adafruit-12-4-digit-7-segment-display-wi2c-backpac">the initial design</a>.) Credit: Lee Hutchinson

But Fusion now ships with an MCP server, so I could potentially let an LLM remote control the application and make the modifications for me. Could it be that easy?

Locally hosted Qwen3.6-35B-A3B-NVFP4 operating Fusion via OpenCode and Fusion's MCP server. It mostly worked! Credit: Lee Hutchinson

The answer turned out to be both "yes" and "not quite." I first tried my modifications with a quantized version of Qwen 3.6-35B (this one, specifically), running locally on a GB10-powered Acer Veriton GN100 that I'm writing a long-term Ars review about.

Qwen 3.6 was almost up to the task, making one of my changes but flubbing the other; I fell back on Claude Code and Fable to handle most of the model adjustments. Still, the local model was intriguing, and I'll be returning to it in a future piece.

Such stuff as prototypes are made on

Once the software began to take shape and the package deliveries were done, it was time to start prototyping. I took over the kitchen table, set up my new soldering iron, and attempted to assemble my first Adafruit display and backpack without destroying them both—and I was mostly successful!

Don't judge my workspace. (And by "workspace," I mean "the kitchen table.") Credit: Lee Hutchinson

Emboldened by not screwing up the soldering too badly and now having a live display to mess with, I pressed on. The next thing to deal with was that while the Adafruit display is dimmable, even at minimum dimness, it still proved too bright for a dark bedroom. This meant I would need something in front of it to block light.

Even at its dimmest, the Adafruit display is hella bright at night. This was me experimenting with combinations of smoked acrylic and NDF material. Credit: Lee Hutchinson

B&H Photo came to the rescue, as it has rolls of neutral density filter material for relatively cheap. This proved fragile and very prone to collecting fingerprints, though, so I ended up pairing the NDF with some smoked acrylic, which meant finding a vendor that would sell me small quantities of cut-to-size acrylic material. (I actually found two—this place and this one.)

One acrylic piece plus one strip of 12 percent NDF knocked the display back to just about the perfect dimness, comparable to my existing cheap bedside clock.

The next issue was iterating through all the model changes necessary to incorporate the acrylic and NDF into the clock case. I ended up (via LLM MCP magic) splitting the existing design into a few more separate pieces and cutting out a pocket for the acrylic face; I also had the LLM add guide pins and holes for each piece. This was all doable without creating any overhangs, so the whole thing still prints without needing supports.

The entire enclosure, sliced and ready to print. Credit: Lee Hutchinson

I iterated through at least three major revisions of the whole thing, and I'm very happy with the endpoint I arrived at. The final version mostly holds itself together, though the front bezel requires either some electrical tape or a couple dabs of superglue to stay attached. I could fix this by altering the guide pins so they snap in instead of merely sitting there, but tape works well enough for me.

The hour's now come

The end result exactly matched my expectations—the best criterion for success that I can think of. With an LLM providing the heavy code lifting and the CAD work, I think I spent more time waiting on supplies to arrive than on anything else—something attributable to my lack of planning and the ease of next-day delivery.

Here's the finished device, first in pieces and then all assembled:

And, look! HomeKit support!

Tailing the clock systemd service's journal while I move the brightness slider in the iOS Home app. The seven-segment display's 16 brightness levels are automatically mapped to the slider's 0–100 percent scale. Display response is basically instantaneous. Credit: Lee Hutchinson

And it deploys!

I run a LAN-only Gitea server because it's fun, and because deploying things via Gitea Actions makes me feel like a real sysadmin. Credit: Lee Hutchinson

For anyone who may be thinking of following in my footsteps and forging their own 3D-printed bedside embodiment of recklessly unchecked horological overindulgence—perhaps because you have no adults nearby to tell you not to—there are many different ways to approach the task. The use of LLM code is a choice, obviously, and you can make a different one. Raspberry Pi units of any flavor are extremely scarce right now, so someone with better coding chops or with a more outsized sense of adventure might try this with an ESP32 microcontroller instead of a Pi Zero. In fact, the ESP32 is probably the smarter choice for controlling the Adafruit display, and it comes with Wi-Fi and I2C support without dragging Debian along for the ride.

Either way, this was a great hobby project. I got to solder stuff, which was both harder and easier than I expected. I used miles of filament while printing and re-printing different iterations of the case. And I learned a ton.

I spent... well, a lot more money than I intended to, between a couple of false starts, the soldering iron and kit, and extra supplies for redundancy and do-overs. And I could have compromised and gotten a regular clock that does most of what I want. But the experience was fun, and the joy of having exactly what I want is priceless.

For folks wanting to see how the code works or to adapt anything in it to their own needs, here's the repo. Enjoy! I'll just be over here, doing CI/CD with my bedroom clock, which is totally a normal and fine thing that normal people do!

Read full article

Comments

Some U.S. lawmakers from both parties and Israeli officials have expressed opposition to such a plan, fearing that the kingdom could use a civilian nuclear project to eventually develop nuclear weapons.

Charles Ross spent 50 years building “Star Axis,” a naked-eye observatory in New Mexico. Now his masterwork is ready. How to share it with a changed world?

China consumes more eggs per capita than almost every other country, so a recent spike in costs is touching a nerve.