We shut down Crosshatch.
We built a “Plaid for personalized AI”: a portable context layer that let users bring their complete context to their favorite apps. We believed consumer AI wouldn’t be interesting without it.
I think we were right.
Go to your favorite consumer website –- has it changed much since ChatGPT launched now three years ago?
Some still imagine a portable context layer––with identity, auth, context and AI all rolled into one. We shipped it! Users could link their past purchases, reservations, schedules, and fitness and grant apps access through a unified API. Apps could register webhooks and get notifications when the user’s permissioned context changed.
We believed that context representation came down to a dutiful ontology, rejecting special little tricks of compression that would get in the model’s way. Coding harnesses’ embrace of search over embeddings seems to validate this approach. “AI memory” felt like a seductive psyop––choosing arbitrary representations of events as some AI innovation that better served VC narratives than good philosophy or engineering.
So why did Crosshatch fail?
I think the biggest reason is that businesses rightfully understand context as a legitimate moat, and that opening the gates to its portability –– while great for the consumer –– could commodify their offerings. Understanding your customer requires investment and businesses make it with the expectation of differentiated return.
And the value proposition of “log in for more personalization” never hit like “connect your bank to pay your friends.”
And perhaps philosophically, it may be that the further away context gets from its source the less useful it actually is. The context of the context is usually the important part, but it’s hard to see or make sense of anywhere but at its origin.
Context rots when removed from its source.
Context constitutes an economic dark matter that provides legibility into the supply and demand of most things. If an end state to AGI is the provision of intelligence for the coordination of commerce, it makes sense that companies would want to keep this dark matter illiquid.
If portable context indeed is a mirage, the alternative must be situated awareness –– intelligence deployed on context observed at the point of action, not reconstructed or reconstituted from distant signals.
I started Crosshatch to unlock an adaptive internet.
The promise of adaptive scaled services remains alluring. I’m not done chasing it.
More soon.