The Tool Is Becoming a World

The most important art technology announcement this week was probably Google Playground with Unity Spark.

I say probably because demonstrations of AI making games are very easy to overrate. You can create something instantly and show twenty seconds of it on X. It can look as if the world has changed. Then you play the thing and discover there is almost nothing there.

Even so, this one feels worth taking seriously.

Google launched Playground on 7 October. You describe a game through conversation and it produces something playable. You can keep changing it, share it and put it into a public gallery.

The involvement of Unity is what makes it interesting. Unity Spark is due later this year and is intended to take these rough games into a proper 3D environment. It will use the Unity runtime and connect to artist-made assets from the Unity Asset Store.

That creates a possible route from a strange idea to a world that can actually be developed. The first generated version may be crude. You can then work on it, shape the behaviour and decide what deserves to remain.

For generative artists this feels quite natural. We already build systems. The system might make a seascape or a creature. Give it space and memory and suddenly it can become somewhere a person visits.

I have been thinking about this with Creatures from the Overgrowth. The creatures become much more interesting if they continue to exist beyond the image. They could have territories, react to visitors and develop peculiar histories. The Wrong Side of a Brushstroke is already based on the idea that a mark has an interior. These tools bring that sort of idea much closer to something one person can build.

Most of what gets made with Playground will probably be rubbish. That is true of every newly accessible creative tool. The real question is whether an artist can push past the default result and make a place with its own logic.

Art Blocks may have found a more important role

For people already working in generative art, ABX may be the most important announcement of the week.

Art Blocks describes it as a free, open-source protocol for documenting anything digital on the blockchain. You bring your own wallet. The work does not need to be an Art Blocks collection.

This gets at a problem that the digital art market has mostly avoided. We have spent years talking about drops and prices while the work itself remains surprisingly fragile. A platform closes. A browser changes. One dependency disappears and the token still exists, although the artwork no longer works properly.

Art Blocks appears to be thinking about becoming part of the permanent infrastructure beneath digital art. I think that is smart. Marketplaces come and go. A useful standard for recording who made something, what version it was and where its parts can be found could last much longer.

There is plenty still to solve. A hash proves that a file existed at a particular point. It does not preserve the experience of an interactive artwork. You may need the code, assets and exact dependencies. There should also be some account of how the work is meant to behave.

ABX will matter if it deals with those awkward details. It will be much less useful if it becomes a neat way of putting a certificate on-chain.

It could also give experiments a proper history. I make a lot of things that sit somewhere between a finished collection and a sketch. Some are abandoned. Others become part of a later work. Recording that development feels valuable and much closer to how digital art is actually made.

Gemini 4 and the model rumours

Gemini 4 Argon was announced at the end of September, although access remains restricted. Google says it is designed for long-running coding work and has a one-million-token context window.

The context window is important because a generative artwork can become an unruly software project very quickly. The rendering system affects the controls. The controls affect recording. Sound gets tied into events and mobile performance starts breaking somewhere along the way.

Current coding assistants often fix the piece directly in front of them and quietly damage something elsewhere. A model that genuinely understands the whole project would be far more useful.

Reports this week say Google employees are already testing newer variants called Barium and Carbon, with Carbon making another jump in coding ability. Google has not confirmed this. It is still a rumour, although a plausible one.

The same applies to Claude Fable 5.5. Anthropic has made no announcement. Its confirmed release this week was Claude Haiku 5.5, a smaller and much cheaper model for fast repeated tasks.

Fable 5.5 is the rumour I would watch. If it improves substantially on Fable 5.1, it may become one of the best tools for working across an entire creative codebase. The model names are arriving at a faintly ridiculous speed, but the underlying change is real.

I want to test these models on something difficult. Give one a complete p5.js or WebGL work that struggles on an older phone. Ask it to improve performance without changing the image or stripping out the strange bits that give the work its character. That would tell us far more than another benchmark table.

The machines are changing too

Microsoft opened preorders for the Surface Laptop Ultra and Surface RTX Spark Dev Box this week. The Dev Box has 128GB of unified memory and is built for running large AI models locally.

The products are expensive. Their real significance is the new category forming around them. A compact machine can now render demanding real-time work and run a serious local model without depending on a distant server.

That has obvious value in an installation. Cloud services create cost and introduce another failure point. A work running in a gallery needs to recover from problems and carry on quietly for the rest of the day.

I would wait before buying anything. We need independent tests showing how these machines behave after hours of continuous rendering. Fan noise and heat will matter. So will the ability to replace the computer without rebuilding the entire display.

Mistral Large 4 also arrived this week. It is a huge multimodal model with one trillion parameters in total and 49 billion active at any moment. Mistral says the weights will be released by the end of October.

It will still require substantial hardware, even when compressed. The appeal is control. An artist or institution could run the model privately and keep a work alive without depending permanently on the company that created it.

A real Android XR target

XREAL has priced the AURA at $1,279 for the base version. These are transparent Android XR glasses connected to a separate computing puck.

I like the split design. Current components are still too heavy and hot to disappear neatly into a pair of normal glasses. Moving the computing elsewhere makes the glasses more plausible for a longer experience.

The price keeps AURA firmly in early-adopter territory. I am also waiting to see how good the field of view really feels and what development access looks like in the UK. Still, it gives spatial artists a concrete device to think about.

The overall direction is becoming much clearer. An image can become the entrance to a persistent system. That system can live on a screen for now and move into the room as the glasses improve.

Making the world is becoming easier. Making a world anybody cares about remains the interesting part.