The Prompt Era Is Ending

For the past few years, the public story of generative AI has largely been about making images from words. Type a description, wait a few seconds and receive something that resembles a finished artwork.

That capability remains impressive, but it is rapidly becoming ordinary. The more consequential developments now lie elsewhere: generating coherent spaces rather than isolated frames, using AI to construct complete interactive systems, bringing powerful models onto local machines, and allowing algorithms to act through cameras, materials and physical installations.

Several announcements this week suggest that the next phase of art technology will be less about asking a machine to make an image and more about creating systems through which worlds, performances and objects can emerge.

From images to worlds

The most creatively significant launch was Atlas, a new “world model” from World Labs.

Unlike a conventional image or video generator, Atlas is designed to understand spatial relationships. It can accept images, text, video, camera positions and depth information, then generate new views that remain connected to the same underlying space. World Labs says it can produce controlled camera journeys lasting up to a minute at 1440p, reconstruct environments from a small number of images and output explicit 3D representations such as point clouds and Gaussian splats.

The crucial development is not simply better visual quality. It is control.

Current AI video often feels like directing a dream. A prompt may produce an extraordinary five seconds, but objects mutate, architecture rearranges itself and the camera only loosely follows instructions. Atlas instead treats the camera position as a fundamental part of the input. The artist specifies where the camera is and how it moves through the scene.

That changes the relationship between artist and model. Instead of repeatedly pulling the lever and hoping for a useful result, the artist begins to stage and navigate a space.

For generative artists, this opens an interesting bridge between code-based work and AI-generated environments. A single algorithmic image could become the entrance to a coherent spatial world. A painted mark might become a landscape. A creature generated in two dimensions could be encountered from behind, above or inside its imagined habitat.

Atlas is initially available only through an early-access programme, so it should still be treated as a research system rather than an everyday production tool. Nevertheless, it offers one of the clearest indications yet of where generative media is heading. Explore the Atlas announcement.

AI becomes a production partner

OpenAI’s release of GPT‑6 Astra points to a parallel change.

Astra is not primarily an art model. Its relevance comes from its reported improvements in software engineering, computer use, browsing and visual interface operation. OpenAI says the model can build websites, operate software and test whether a finished interface actually works.

This matters because much of contemporary generative art is no longer a single sketch. An ambitious browser-based work may contain shaders, simulation, responsive sound, animation, recording tools, narrative systems, mobile controls and multiple performance modes. Every addition creates new interactions—and new ways for the work to break.

AI-assisted coding has already made these projects faster to create. More capable computer-using models could now help with the less glamorous but essential work surrounding creation: performance profiling, cross-browser testing, interface repair, adaptive quality settings and the management of increasingly complicated codebases.

The danger is predictable. When adding features becomes almost frictionless, artists can mistake complexity for progress. A model will nearly always find something else to add.

The best use of this new capability may therefore be subtractive: making a work faster, clearer and more stable while preserving the quality that made it worth building. Astra’s real artistic value will not be measured by how much code it can produce, but by whether it can help ambitious interactive work survive contact with actual audiences and imperfect hardware. Read OpenAI’s Astra announcement.

Creative intelligence moves onto the machine

NVIDIA made two connected announcements this week.

First, it agreed to acquire Hugging Face for approximately $12.93 billion. Hugging Face has become one of the central repositories for open models, datasets and AI applications, used by millions of developers and creators. NVIDIA says it will remain open, hardware-neutral and supportive of models from across the ecosystem. Read NVIDIA’s acquisition announcement.

At the same time, NVIDIA provided more detail on RTX Spark, a new class of compact Windows computers expected in October. These systems can offer up to 128GB of unified memory, a Blackwell GPU and enough local computing power to run substantial AI models without sending every task to the cloud.

For digital artists, local AI has several advantages. It removes the anxiety of paying for every experimental generation, allows private source material to remain on the artist’s machine and makes it possible to build installations that continue operating without a permanent internet connection.

It may also help solve a persistent problem in displaying generative art. Most commercial digital-art screens are designed to play video or display static images. They are not built to run demanding WebGL, Three.js or AI-driven artworks continuously. Compact, high-performance local systems could provide the intelligence behind a new generation of displays: devices capable of rendering live work, reacting to viewers and generating new material in real time.

There is a tension here. Local tools promise independence, while NVIDIA’s acquisition of Hugging Face concentrates more of the supposedly open ecosystem within one company. Artists and developers should enjoy the new capability while keeping their projects portable and avoiding unnecessary dependence on proprietary infrastructure. See NVIDIA’s local-AI and RTX Spark announcement.

Algorithms escape the screen

Technology’s movement into physical space is also visible in A.A.Murakami’s New Nature, which opened in Seoul on 1 September alongside Frieze Seoul.

Among the works is Mud Plotter, a machine that uses algorithms based on seashell growth to make paintings from oil and mud. The resulting works are not simply printed copies of digital images. The algorithm acts through physical materials whose viscosity, movement and imperfections contribute to the final object.

This distinction is important.

The most interesting future for physical generative art is unlikely to be endless archival prints of screen-based works. It will emerge when code encounters processes it cannot entirely control: flowing pigment, changing light, mechanical movement, weather, sound, biological growth or audience behaviour.

A digital system can establish the rules, but matter completes the artwork.

This approach also creates a stronger relationship with galleries, collectors and public spaces. The work retains the conceptual power of an algorithm while acquiring physical presence, material uniqueness and the visible evidence of its own production. Read more about A.A.Murakami’s practice.

The camera becomes generative

Camera Intelligence offered another version of the physical-digital hybrid with new generative video tools for its Caira mirrorless camera.

The system begins with footage the user has actually filmed. It can then relight, reframe or extend the scene and generate new camera movement using conversational instructions. Originals are preserved, while generated versions are separately identified.

The current implementation is limited to 720p or 1080p and is being presented primarily as a previsualisation tool. It is therefore not yet a reason to abandon professional editing or production workflows.

Its underlying proposition is more interesting than the current product: generative AI can amplify an act of observation rather than replace it. The artist must still choose a subject, stand somewhere and point the camera. AI then becomes an extension of that encounter.

As synthetic imagery becomes abundant, work that begins with something physically witnessed, built or performed may become more—not less—valuable. See Camera Intelligence’s announcement.

Platforms remain fragile

While creative capability accelerates, distribution remains uncertain.

The recent closure of fxhash is an uncomfortable reminder that even influential generative-art platforms can disappear. The code and tokens may remain, but discovery, community, presentation and the shared ritual of minting are not automatically preserved with them. Read Right Click Save’s analysis of the closure.

Artists should increasingly treat platforms as channels rather than permanent homes. Generators, seed systems, metadata, collector interfaces and high-resolution rendering should be independently preserved wherever possible.

The encouraging counterpoint is that digital and computational art continues to gain institutional visibility. Ars Electronica’s 2026 festival opens this week with work spanning artificial intelligence, image transformation, spatial installation and human-machine performance. Explore the Ars Electronica programme.

What comes next

The strongest signal this week is not that AI can create more content. We already have more content than anyone can meaningfully absorb.

The real shift is towards systems that understand space, operate tools, inhabit local hardware and interact with physical reality. These systems give artists something far more interesting than another image generator: new kinds of material.

The artistic opportunity is to avoid using them merely to imitate existing media. A world model should not simply make a smoother promotional video. A coding agent should not merely add more buttons. A powerful local computer should not become an expensive slideshow player.

The artists who matter will use these technologies to create forms that could not previously exist—and will know when to surrender control so that the system, the material and the audience can complete the work.