DEV Community: aditya

AI Trends to Watch in 2026

aditya — Wed, 02 Sep 2026 07:06:38 +0000

Artificial intelligence keeps moving fast, but 2026 feels different from the last couple of years. The conversation is shifting from "can AI do this?" to "can we trust AI to do this reliably, at scale, without blowing up our budget or our risk profile?" Here's a look at the trends actually shaping the industry right now.

1. Agentic AI Grows Up — But Stays on Probation

Agents were the buzzword of 2025. In 2026, the hype is being replaced by scrutiny. Organizations are learning that agentic AI only delivers real value where the underlying data and processes are already solid — messy inputs still produce messy, unreliable agent behavior. Expect less "give the agent everything" and more targeted rollouts in well-defined workflows: coding, customer support triage, research assistance, and internal ops.

The bigger shift is architectural: instead of a dozen disconnected AI tools, companies want a single "front door" — one place to kick off a task, with agents then operating across your browser, editor, and inbox in the background. Whoever controls that front door controls a lot of the user relationship.

2. From Answering Questions to Actual Collaboration

Early generative AI was mostly a really good search-and-summarize tool. The 2026 shift is toward AI as a collaborator that understands context, not just instructions. In software development this shows up as AI that tracks the reasoning behind code, not just the syntax. In research, it looks more like a lab assistant that can run experiments and interpret results rather than just fetch papers.

For engineers, this matters a lot: the tools you use for coding, debugging, and reviewing are getting less like autocomplete and more like a teammate that keeps context across a whole task.

3. Reliability and Security Take Center Stage

As agents get more autonomy — reading files, taking actions, touching real systems — the cost of mistakes goes up. Reliability (staying on task, recovering gracefully from errors, behaving predictably over long-running tasks) and security are becoming the actual bottleneck, more so than raw model capability. If you're building anything agentic, this is where most of your engineering effort should go in 2026: guardrails, monitoring, and graceful failure handling, not just prompt design.

4. Smaller, Specialized Models on the Rise

The "bigger is always better" era is fading. Open-source, domain-specific models are catching up fast in narrow tasks while costing far less to run. Expect more multimodal reasoning models that are smaller, easier to fine-tune, and cheaper to deploy close to where the data lives — which also helps with latency and data residency concerns.

5. The AI Bubble Question Won't Go Away

Alongside all the technical progress, there's a real economic debate: is enterprise AI spending outpacing the value it's generating? Analysts are watching closely for signs of a correction, alongside continued (if slower and more skeptical) enterprise investment in agentic AI. Practically, this means ROI is under real pressure — a "results over experiments" mentality is replacing the earlier "let's try everything" phase.

6. Compute, Energy, and Geopolitics

Processing power is becoming a strategic resource, not just a technical one. Export controls, sovereignty regulations, and energy availability are increasingly determining who gets to build and deploy at scale. If you're evaluating AI infrastructure providers, expect more emphasis on hybrid, regionally distributed compute rather than a handful of mega data centers.

7. AI in Knowledge Work Keeps Redefining Skills

The "one-person unicorn" idea — a very small team producing outsized output with AI — is gaining traction. This is reshaping what's expected of individual contributors: less about doing every task manually, more about knowing how to direct, verify, and combine AI outputs effectively. For engineers specifically, this reinforces the value of understanding systems deeply enough to evaluate AI-generated code and designs, not just accept them.

The takeaway: 2026 isn't about flashier demos — it's about operationalizing AI responsibly. The trends that matter are governance, reliability, and getting real ROI out of agentic systems, not just chasing the next capability jump.

Six Benefits of a Museum Interactive Installation, and the Systems That Produce Them

aditya — Mon, 17 Aug 2026 13:21:38 +0000

Six Benefits of a Museum Interactive Installation, and the Systems That Produce Them

A wall label waits. It shows everyone the same text and works only if someone reads it. A responsive exhibit detects an input, decides what it means, and changes state. Only the second closes a loop with the visitor.

That loop is the point of a museum interactive installation: the value sits not in the hardware, but in what it lets a visitor do that a sealed case cannot.

1. Interaction turns a passing glance into a stop

Beverly Serrell's tracking-and-timing work, aggregating 65 studies on top of 110 earlier ones, measures how much of an exhibition gets used, through the sweep rate index and the percentage of diligent visitors, against benchmarks of 300 and 25 percent. Visitors are casual, not thorough.

An interactive element earns a stop by offering a cheap action with a visible consequence the visitor can attribute to themselves, and attribution depends on timing. Nielsen’s response-time limits apply: under roughly 0.1 seconds a response feels like direct manipulation, and past a second the causal link weakens. An exhibit with 400 ms of latency reads as sluggish however good its content. A stop is not proof of learning, though, only its precondition.

2. Visitors set their own sequence and depth

A printed panel offers one level of detail in one order. An interactive system separates content from presentation: text, media and object records become addressable nodes, each input selecting the next. Two people cross the same dataset by different routes.

At Mumbai’s Chhatrapati Shivaji Maharaj Vastu Sangrahalaya, the Networks of the Past exhibition set over 250 artefacts alongside four digital experiences, including an LED table for tracing trade routes back to objects in the cases.

Branching multiplies what must be written, translated and approved, and depth without editing produces a database browser.

3. Multisensory interaction widens who can reach the interpretation

Can interactive exhibits improve access? For blind and low-vision visitors, yes. 3D scanning and printing produce a tactile replica of an object that cannot be handled. A camera above it tracks fingertip position, software maps that to a region of the model, and the matched audio description plays. Reichinger and colleagues demonstrated this on tactile reliefs; embedded NFC tags do similar layering.

Smithsonian accessibility guidance stresses that replicas belong in the main gallery, not a side room: separating them removes the shared experience. Handled replicas wear, so reprinting is part of the design.

4. The visit can continue after the visitor leaves

Cooper Hewitt’s Pen documents this well: every object label carries an NFC tag holding that object’s database ID. The Pen holds a battery-powered reader, pairs at ticketing with a visit ID, and uploads its stored IDs on return to build a personal web page. The museum reported over 400,000 acts of collection in five weeks.

The cost is documented too: power and physics forced the team to abandon readers-in-labels for a fleet of powered pens tied into ticketing, collection management and the API. A cheaper route is an NFC sticker on the label, read by the visitor’s phone.

5. Instrumented exhibits produce evidence a static case cannot

A vitrine tells you nothing without an observer and a clipboard. An interactive installation emits events: session started, node selected, time in state, completed, abandoned. Those logs show what people choose, what they never reach, and where sessions die.

Handle the data carefully. A depth sensor producing joint coordinates is a different privacy proposition from a camera classifying faces. Discard raw frames, log derived events, keep retention short. A log records behaviour, not understanding.

6. Content changes without rebuilding the gallery

Printed graphics are fixed at fabrication. Put media and text in a content management system behind the runtime and a curator can swap an artwork, fix a caption or add a language without touching hardware. The Cleveland Museum of Art rotates featured works in ArtLens Exhibition every 18 months, so returning visitors find something new.

That pipeline also makes newer sensing usable. ArtLens Reimagined, opened in July 2026, includes pose and hand-tracking experiences matching several visitors at once against the collection. The direction is interaction needing no device and no instructions.

How does a museum interactive installation create these benefits?

Every example above runs one loop: visitor action → sensing → interpretation → application state → output → feedback → next action.

Sensing captures input: depth cameras give joint positions, capacitive panels give touch coordinates, NFC readers give object IDs, proximity sensors decide whether the system idles or wakes.
Interpretation turns signal into intent: computer vision converts frames into a pose, debounce logic separates deliberate action from passing traffic.
Application logic holds state, picks the response, and drives the output, with a CMS feeding it content and an event stream flowing out.

Total loop latency decides whether the result feels responsive or broken, which is why the sensing and rendering budget matters more than resolution.

What should museums consider before building one?

Reliability before features. An installation that fails one time in ten does not deliver ninety percent of the benefit: visitors who watch someone fail rarely try.

Environment. Ambient light defeats projection and interferes with infrared depth sensing. Sunlit atria are hostile.
Calibration. Sensors drift, so write a routine an attendant can run in minutes.
Throughput. A two-minute experience in a busy room creates a queue, and queues suppress the engagement being measured.
Failure behaviour. A black screen is worse than a printed panel. Design a degraded state that still says something, and add watchdogs and remote monitoring.
Ownership. Name who updates content, and fund maintenance past opening day.

How can museums measure whether it worked?

Choose metrics from the objective, not from a dashboard. Useful indicators: attract rate (passers-by who stop), initiation rate (stops producing a first input), completion rate, content coverage and repeat interaction.

Learning and accessibility outcomes need other instruments: cued observation, short exit interviews, and testing with the intended audiences. Logs say what happened; people say what it meant, and only both together tell you whether the installation is doing its job.

10 Museum Interactive Installation Concepts That Start With the Visitor

aditya — Mon, 17 Aug 2026 13:16:46 +0000

10 Museum Interactive Installation Concepts That Start With the Visitor

A visitor stops at a carved stone fragment. The label gives a date, a findspot, and a material. Two steps away, a screen invites her to set that fragment back into the building it came from. She stays four minutes instead of forty seconds, and leaves knowing what the object was for.

That gap between reading and doing is where a museum interactive installation earns its budget. An interactive installation is any exhibit element that responds to what a visitor does, whether a touch table, a projection, or an AR layer. The concept, not the hardware, decides whether the gap closes. Each idea below starts with what a visitor does.

1. See the Object in the World It Came From

Visitors aim a phone at an artifact and see the workshop or landscape it once sat inside, rebuilt at scale around it. A column base makes sense once the column stands above it. Archaeology and architectural fragments suit it best. Markerless AR keeps costs low. Show where evidence ends and inference begins.

2. Reveal What the Surface Hides

Some objects hold more than the eye can reach. Give visitors one slider that moves between a photograph and an X-ray or multispectral scan. The discovery lands instantly: an earlier composition beneath a painting, a repair nobody mentioned. Keep it to one control. Add a second slider and people disengage.

3. Walk Through the Timeline

Run the chronology along the floor rather than a wall, so visitors move through it. Standing on a decade brings up its objects and voices. Social and industrial history suit this format, because change is the argument. Pressure mats can drive it. Plan for queues, and add a hand height version.

4. Follow One Object Across the Map

A shared map table lets several people trace where a collection travelled. Touching a port lights the route, the goods, and the people who moved them. Provenance stops being small type and becomes geography. Trade, migration, and species distribution all fit. Round tables invite groups; wall maps make queues.

5. Find Something Without Knowing Any Names

First time visitors often cannot name a single artist. Let them draw a shape or trace a curve, then return the works that match. No reading and no vocabulary are required, which helps children, second language visitors, and anyone unsure where to start. Image similarity search does the work.

6. Match Your Own Body to the Collection

Body based interaction turns looking into comparison. The Cleveland Museum of Art reopened its ArtLens space in July 2026 with experiences that match a visitor’s hand position or pose to figures in artworks, tracking several people at once. Visitors start reading gesture. Cameras raise privacy questions, so state plainly what is stored.

7. Let Visitors Collect Their Own Visit

Give people a way to keep what caught their attention. Cooper Hewitt, Smithsonian Design Museum, issued visitors a Pen from 2015 to 2020: touch a label and the object saved to a personal page opened at home. Selection is the whole interaction, and the difficulty is operational: distribution, charging, returns, staffing.

8. Build Something Together on One Wall

Contribution changes the register of a gallery. Visitors add a drawing, a memory, or a remixed collection image, and the wall shows what everyone else made. Cleveland’s reimagined gallery includes a community mural built around prompts that change regularly. Moderation is the real design work, so decide who reviews submissions before building it.

9. Interpret an Artwork Through More Than Sight

Sight is not the only route into an object. Tate Sensorium, shown at Tate Britain in 2015, paired four paintings with sound, scent, taste, and touch. The constraints matter as much as the results: small groups, staffed sessions, consumables, allergy checks. Treat multisensory work as a set piece, not a floor.

10. Make a Dataset Something You Can Stand Inside

Numbers rarely land as text. Drive light or sound from a dataset (a century of river levels, a city’s growth, a collection’s acquisition record) and let visitors change one variable. The abstraction becomes physical. Keep the source visible, refresh it on a schedule somebody owns, and resist adding a second layer.

How Do You Choose the Right Museum Interactive Installation Concept?

Start with the sentence you want a visitor to repeat afterwards, then test candidate concepts against the conditions you actually have:

Exhibition objective, and the story the zone carries
Audience: ages, languages, school groups, repeat visitors
Collection type, and what the object cannot show alone
Floor area, sightlines, throughput, and what a queue does to the room
Accessibility: seated use, contrast, captions, one handed operation
Maintenance: who reboots it, who updates the content
Privacy, if cameras or personal data are involved
Budget across five years, since content ages faster than hardware

Concept choice sits inside the wider museum design process, where the zone story and visitor flow are settled before technology is specified. The Smithsonian guidelines for accessible exhibition design cover label contrast, viewing heights, circulation, and instructions for interactives, a sensible floor for anything new.

What Makes a Museum Interactive Installation Work?

Successful installations share a few traits. Visitors understand the action without instructions. The system answers immediately, so cause and effect are obvious. The interaction says something the label cannot, and it sends attention back to the object. It survives a busy Saturday with one member of staff, and it still rewards a second visit.

Interactive is not automatically better. A good label and good lighting often beat a screen. The honest test is whether the installation changes what visitors notice.

Conclusion

The strongest museum interactive installations begin with a meaningful visitor experience, not a technology shopping list. Find the moment where understanding breaks down, decide what the visitor should do there, then choose the simplest system that delivers it. Most disappointing interactives are not badly built. They answer a question no visitor asked.

Key Takeaways

Define the visitor outcome before choosing hardware.
Match the concept to your collection, space, and staffing, not the best demo.
Keep the interaction learnable at a glance and instant in response.
Design for accessibility and mixed ages from the first sketch.
Budget for content updates and repairs, then measure whether visitors notice more.

12 Museum Interactive Installations and the Reasons They Work

aditya — Mon, 17 Aug 2026 13:16:07 +0000

12 Museum Interactive Installations and the Reasons They Work

Most people move through a gallery the same way: look, read the label, move on. It is a comfortable rhythm and an easy one to forget. Everything changes when the room answers back.

A good museum interactive installation replaces looking with doing. The twelve below are real projects. Some run on serious technology. One runs on printed relief and a pair of hands.

1. Gesture instead of touchscreens

At the Cleveland Museum of Art, projections read body movement instead of fingertips. Visitors copy a sculpture's pose or rotate a 3D model in mid-air. Dropping touchscreens was deliberate: the games sit behind the real artworks, so each round pushes attention back to the object.

2. A forest you walk down

teamLab's Story of the Forest filled the Glass Rotunda of the National Museum of Singapore from 2016 to 2024. Sixty-nine watercolours from the William Farquhar collection became animated wildlife along a 170 metre spiral descent. Singapore's Roots portal noted that visits to the gallery holding the originals rose too.

3. AR that repairs an old hall

The Bone Hall at the Smithsonian's National Museum of Natural History opened in 1881, and staff found most visitors walked straight through. The Skin and Bones app lets you aim a phone at thirteen skeletons and watch muscle and fur return. The historic cases stayed untouched.

4. Drawing on a design museum's walls

Cooper Hewitt's Immersion Room ran from 2014 to 2024. Using the museum's Pen, visitors pulled wallpapers from the collection and projected them floor to ceiling, or sketched a pattern and watched it repeat at full scale. The same Pen collected objects around the building with a tap.

5. Ask a survivor a question

Dimensions in Testimony records long interviews with Holocaust survivors and other witnesses to genocide, then uses natural language processing to match a spoken question to the right recorded answer. Visitors ask what they actually want to know, out loud, and a reply comes back.

6. Every sense but sight

Tate Sensorium, the 2015 IK Prize winner from studio Flying Object, ran six weeks at Tate Britain. Four British paintings were paired with scent, sound, chocolate and physical sensation. Only a few people entered at a time, a reminder that intensity and throughput pull against each other.

7. VR against a crowding problem

The Mona Lisa could not join the Louvre's 2019 Leonardo retrospective without making the show unvisitable. The museum built a seven minute VR piece instead, using infrared and X-ray imaging to reveal the panel beneath the varnish. Technology here answered a constraint rather than chasing novelty.

8. An aquarium made of crayon drawings

In teamLab's Sketch Aquarium, a child colours a fish on paper, hands it over to be scanned, and watches it swim into a projected ocean with everyone else's. Touch a fish and it darts away. The input is paper and wax; the result belongs to the room.

9. One tag, one person's war

Every visitor to The National WWII Museum in New Orleans registers a digitally enabled dog tag tied to one real service member or civilian. Tapping it at kiosks through the Road to Berlin and Road to Tokyo galleries unfolds that person's war alongside the campaign.

10. Paintings you read by hand

Hoy toca el Prado opened in Madrid in 2015 with six relief reproductions made using Estudios Durero's Didu process, plus braille labels and guided audio, developed with blind and partially sighted advisers. The museum also lent opaque glasses so sighted visitors met the works by touch alone.

11. A map of ancient trade routes

Objects tell you what was made; maps tell you what moved. At CSMVS in Mumbai, the Networks of the Past gallery traces India's ancient exchanges with Mesopotamia, Egypt, Greece and Rome. An interactive table lets visitors follow commodities along those routes back to the artefacts in the cases, one of four digital experiences built for the gallery.

12. A room a guide drives

Deep Space 8K at the Ars Electronica Center in Linz projects onto a 16 by 9 metre wall and a floor of the same size, with laser tracking that follows up to thirty visitors at once. A person leads each session rather than a menu.

What makes a museum interactive installation memorable?

Memorable installations give visitors something real to decide.

Meaning before mechanism. Rotating a 3D model teaches you something about a maker's decisions; swiping a carousel teaches you nothing.

Feed the collection, do not compete with it. Cleveland's games point back at the wall. The Smithsonian left its 1881 cases alone.

Match the effort to the reward. Colouring a fish takes two minutes and produces something that is yours.

Reliability is an experience quality, not a maintenance line. An installation that fails on a Tuesday morning is worse than none at all.

How should a museum choose an interactive installation?

Start with the constraint that will not move, usually space, budget or staffing. Worth settling before anything is specified:

What should visitors understand, and can an installation carry that better than a label?
How many people pass per hour, and how long does one interaction take?
Does it work for a wheelchair user, a child, someone with low vision, someone who reads no English?
Who cleans it, restarts it, and replaces the projector lamp in year three?
Is visitor data collected, and how is that explained to the person interacting?
What counts as success: dwell time, completion, repeat visits, or a curatorial outcome?

Interaction complexity deserves its own line: every extra step loses people, and the busiest installations need no explaining.

Start with the visitor, then pick the technology

Answer four questions before choosing a single component: what should the visitor discover, what should they feel, what should they do, and what should they remember on the way home.

Technology chosen after those answers tends to be modest and correct. Chosen first, it arrives looking for a story to justify it. The strongest installations do not make visitors think about the technology. They make them think about the subject.

Textile traceability starts with the data model, not the blockchain

aditya — Mon, 10 Aug 2026 04:51:22 +0000

Textile traceability starts with the data model, not the blockchain

Most textile traceability projects I have seen begin with a platform decision. Which chain, which vendor, which hosting model. That ordering causes problems later, because the hard question is not where the record lives. It is what the record says.

A product is not a person

Verifiable credentials were designed around people. An issuer makes a claim about a subject, the subject holds it, a verifier checks it. Textile traceability breaks that shape immediately. The subject is a bolt of fabric or a finished garment. It cannot hold anything, and it cannot consent to a presentation.

You have two workable options.

Bind the credential to an identifier printed on the item. The physical tag becomes the pointer. Whoever holds the item can resolve the record.

Bind the credential to the producer and reference the item inside it. The weaver or the unit is the subject. The product is an attribute.

The first is simpler for consumer verification. The second is easier when one producer makes thousands of items and you want to avoid issuing thousands of separate credentials.

What actually goes in the credential

The temptation is to put everything in. Resist it. A credential is signed, so anything inside it becomes expensive to change. Anything that changes should live outside and be referenced.

{
  "@context": ["https://www.w3.org/ns/credentials/v2"],
  "type": ["VerifiableCredential", "TextileProductCredential"],
  "issuer": "did:example:textilesdept",
  "credentialSubject": {
    "id": "urn:item:PTH-2026-004182",
    "productType": "Paithani saree",
    "producer": "did:example:cluster-yeola-0114",
    "productionMethod": "handloom",
    "verifiedOn": "2026-07-15"
  }
}

Notice what is absent. No retail price, no stock location, no current owner. Those change. Put them in a system that expects change.

Read the specification before you design the schema

The W3C Verifiable Credentials Data Model 2.0 became a Recommendation in May 2025, and it is worth reading properly rather than skimming a summary. It defines the three-party model, how credentials are secured against tampering, and the extension points you will need for a product schema. Most of the mistakes I see come from people inventing structure the spec already provides.

Issue per item or per batch

This is the decision that most affects cost and credibility.

Per batch is cheaper. One credential covers a production run. It also means a counterfeit only needs to claim membership in that run, which is easy to fake if the batch identifier is printed on a label.

Per item costs more and gives you something a counterfeit cannot replicate cheaply, because each identifier resolves to a record that either exists or does not.

For heritage textiles the per-item route is usually the right one. The Maharashtra textiles department took that approach for Paithani and Karvati Kathi sarees, giving each registered piece its own record rather than a shared batch number. The Vingatha case study sets out how the registration and verification steps fit together, and there is a longer write-up of the design decisions if you want the reasoning.

Decide what the record proves

Be precise here, because overclaiming is how these systems lose trust.

A signed record proves that a named issuer made a specific claim on a specific date, and that nobody has altered it since. It does not prove the claim was true. If a verification step upstream was sloppy, you have an authentic record of a bad judgement.

That is not a reason to skip the cryptography. It is a reason to design the human process with the same care as the schema, and to be honest in your documentation about which part guarantees what.

Where to start

Write the schema before you choose a platform. Show it to somebody who works in the supply chain you are modelling. If they cannot tell you where each field comes from and who fills it in, the schema is not finished, and no amount of infrastructure will fix that later.