Today we launched a personalized inference provider for Vercel’s incredible AI SDK.
Vercel’s AI SDK enables developers to generate UI like
- chat
- dynamic carousels
- reranked elements
- relevant tools like weather or maps
typically in response to unstructured inputs.
Developers love the Vercel’s AI SDK.
Indeed, we envision a future where every interface is personalized to us individually
The internet’s version of rolling out the red carpet
So making a secure personalized inference provider for Vercel’s Typescript SDK is a natural step.
We’re really excited about the possibilities this unlocks.
App context windows are limited
App user “context windows” are constrained.
To be clear, that’s probably a good thing. We don’t need a panopticon where every app knows everything about us.
And practically app context windows are constrained not because the AI models they use are limited. It’s rather because they just don’t have that much context.
App interfaces into user context are just a function of app-induced user labor
- A quick search
- Some clicks and scrolls
- A purchase that one time
Adaptive interfaces won’t be powered by first party data.
They need a new data layer. We interact with applications across the internet. [Our full context probably resists digital representation.] But our interactions scattered across the internet represent a lot of it.
Apps as permissioned AI views of context
With Crosshatch, developers specify views into the user’s complete context
- Recent restaurant reservations
- Hotel spend over the past two years
- Runs for the last three months
and instruct a user’s AI how to
- interpret
- view
- activate with (private?) tools
that context to be in the app view.
That is, with Crosshatch application behavior can vary independent of user interaction with a given application!
Of course, some of the most exciting applications require coordination inter or intra users or businesses
- Scheduling
- Product / place recs
where we see developers adding their own context to Crosshatch calls let AI in Crosshatch-secured VPC mediate context
- Users
- Developers
share.
In these applications, Crosshatch acts as a trusted compute layer between multiple parties, enforcing respective egress requirements.
How Crosshatch keeps context secure
You can learn more about our security model in our docs.
For most applications, users link data under an “Access while using” model. Apps can only access our context while we’re on the site. We execute this via short-lived single-use access tokens available only while users are on the linked web app.
For our forthcoming Webhooks feature, the permission model follows Apple’s “Always allow” permission setting. Here users can permission apps to subscribe to AI views of changing context. This is useful for logistics companies who help users save time based on changing underlying context.
Context sharing occurs under a Crosshatch-operationalized “contextual integrity” framework – where data is agreed to be shared for a particular purpose.
We log calls for ex-post verification that calls satisfy sharing agreements. Next year we expect to make this real-time.
Though we’ve seen Apple, Google, and Bill Gurley express contempt against cross-app context sharing, we’re encouraged both by first principles and even Apple movement to the contrary.
Toward adaptive interfaces
AI unlocks lower-friction interfaces to our context.
Instead of filtering inventories
> gender: men
> clothing: sweaters
> size: L
> color: grey
we can now just search in natural language.
We’re excited by
“chat to interact”
interfaces, but we’re most excited by interfaces that vibe to us – interfaces that use complete user context – to deliver experiences that delight and save us time and money.
If you’d like to start building adaptive interfaces in your app, create a free developer account or jump on our onboarding form. We’d be thrilled to chat.
