AI unlocks ephemeral malleable UI – UI that morph depending on the context.

For this AI-powered malleability, where should this malleability-making AI application live?  

• At the OS-level?
• The browser-level?
• Or the application level?

Before proceeding, we should clearly delineate between the AI resource that might power malleability and the interface where the malleability might appear.

Language model AI resources are making their way to the edge – browser and OS – from rarefied cloud addresses. Where a language model is deployed seems a function of performance, availability, cost and security.

AI is a valuable commodity – available everywhere with tokens soon too cheap to meter – perhaps even marginally free, but for the cost of power.

This is why it’s important to set aside the location of the LM deployment and instead focus on the application of AI – the human interface where its outputs appear.

While there’s AI at

• The OS-level with Microsoft Copilot and Apple Intelligence
• Browser-level with Arc, Perplexity and Poe (we include the latter two who reportedly use headless browsers for scraping)

we say fundamentally all of these are AI applications, bubbling up outputs of AI resources – wherever they might live – on surfaces they own with the attention of their users.

AI introduces important questions of provenance. Since AI inference can make anything everything, much of AI reduces to questions of who owns what.

AI applications then can be divided by their philosophies around provenance or rightful or practical custody. Some use AI deeply

• Microsoft Copilot on Microsoft Word
• Box AI on documents in Box
• Notion AI on writing in Notion
• Apple Intelligence on apps via App Intents

on content and services over which they’ve explicit rights or connection. That is, users have signed terms that explicitly grant these services rights to provide services against uploaded User Content or actions against registered, explicitly legible actions.

Others

• Perplexity and Poe on the Internet
• Arc over sites
• Experimental browser extensions
• iOS18 Safari “Web Eraser”

apply shallow AI over content not explicitly authorized for use this way.

Tldraw’s engineer-in-residence Orion Reed who created this experimental browser extension that

lets you pull apart any website, create views and transformations with LLMs, add new UI, pull in context from knowledgebases and APIs, and mix pieces of multiple websites together

called such application of AI a form of adversarial malleability

I've been calling this "adversarial malleability" as it makes the web more malleable with no action needed on the authors side in a way that is potentially against the business interests of the site + it's hard to stop you from doing it (without locking down the web entirely)

Irrespective of where you normatively land on the Right Way To Operate, it appears apps fall into one of two classes, one of which structurally blocks shallow AI applications from providing value users expect.

‍

Thick and thin applications

We co-opt language from web development.

We’ll say an application is thick if it requires context that is of heterogenous and stochastic state that is expensive or hard to know. We’ll say apps are thin if … they’re not thick.

For instance, Airbnb is a thick app.

It’s composed of inventory that really only Airbnb knows about. Its inventory quality may be changing – a room or a house that started out great might be next door to loud construction that just started.  That quality is given by reviews and Airbnb has the most concentrated and protected collection of them. Getting data about Airbnb inventory is expensive e.g., last year they expanded their Anti-Bots engineering team that was

harnessing the power of AI, state of the art ML techniques, and data for the higher purpose of preventing scraping and protecting data.

On the other hand, NYT is a thin app.

Its inventory – its stories – are “all the news that’s fit to print.” A story’s quality is given by the article’s provenance – it’s on the NYT! No reviews required. Any shallow app – whether a browser, browser extension, or another web app – could technically take an article and refine it for use in another context. No other data is required to make it useful.

Making an email message "more formal" is a thin application. Re-rendering a page in the style of Barney is a thin application. Anyone – OS, browser, or application – could credibly execute it with only subtle nuance at most lost.

Applications that provide rich individual personalization appear to be thick.

But for thick apps like Airbnb, a shallow AI attempting to operate over Airbnb like

• A Perplexity or a Poe
• A browser or browser extension

would struggle because Airbnb is thick. All the requisite context that a shallow AI might need is not just on the page – it’s across pages, projections of what’s in Airbnb’s protected backend. User context it might need is scattered across applications; not visible from actions that such a shallow AI might have happened to already observe.

Apps like retailers with variable inventory are likely thick. Aggregator apps that operate over publicly accessible APIs are thin, as the inventory they operate involves context that’s not hard to know. Just call the APIs they use. Or get the context once – rooms at the Westin Copley Place are probably the same this year as they were last.

Thick apps make it difficult for shallow AI to provide great user experiences because they involve context that’s just too difficult or too expensive to get. Thick apps are best positioned to offer compelling AI experiences themselves. It’s true that AI can change interfaces, but those changes are only surface level.

They do not engage the deeper context that only thick apps really know about. And perhaps that deeper context is really only theirs to wield.

‍

Sovereign representation

Malleable or not – we think pure representations are those defined by the self; not cast onto others by foreign actors.

With software ever easier to create and incredible interest in the opportunities of AI, a lack of malleability in an application may increasingly reflect intentional expression than a lack of interest. If consumers demand AI-powered malleability, competition could drive change in the market. Adversarial malleability may not be required.

While some shallow AI provide aggregations that a single actor could not (e.g., providing a unified perspective from fragmented sources LA Times, Forbes and Washington Post simultaneously), such aggregation may also violate the agency and policy of an individual or organization that might explicitly restrict such activity.

Either way, at Crosshatch we’re interested in a malleable internet made hyper-personalized – making every app a thick app, by incorporating headless hyper-personalized AI in a way individuals and businesses control. We think malleability defined by everyone – a decentralized malleability – is more interesting and vibrant than malleability defined by a few. It also seems likely to be deeper, as the malleability emanates from the self rather than by others.

Decentralized malleability is really core to what Crosshatch is about – we want to make the internet more you. Only you can define who you are, and who each of us is resists attribute-based identification defined by others. Anyone telling us who we are or what we’re about is bound to feel shallow. Personalized AI must be decentralized and emanating from the self.

We believe personalized AI will appear everywhere (and as a result is an identity). While OS and browser-lead malleability are scalable and require little to no cooperation with others, for the open web, we’re most excited about a malleable internet lead by the application layer.

Applications (particularly thick ones!) know themselves best and are best positioned to recast their own identity remixed with their users'. This is why the application layer is such a promising place for rich personalized AI to manifest – it's closest to the people and organizations that create experiences and services we love.

Only you know who you are. Mediating third parties are just guessing.

A decentralized malleability invites everyone on the internet to participate in defining themselves with respect to everyone else, rather than having a few companies telling everyone what and who everyone else is. Decentralized malleability takes more work to come by, but we believe it’s a path to a richer internet and deeper relationships between people and businesses.