PWV Portfolio Feed

Zed is built from the ground up in Rust, using our own custom GPU-powered UI framework. Owning all layers in the stack allows us to deliver low-latency performance, but comes with the cost of a long beta period as we implement everything from scratch.

Use PlanetScale to operate your sharded database. We support:

- Workflows for going unsharded → sharded
- Scaling your shards to use larger or smaller instances
- Workflows for increasing or decreasing the number of shards (new)

All with no downtime.

My hope is that teams will take this code and run with it. My double hope is that we get a community of developers trading tricks and strategies for generating better outputs, evals to train and tune against, and perhaps even new models to use.

The ai module is MIT licensed. It is designed to let developers write their own integrations with model providers like OpenAI, Google, or Anthropic—and to hack the shit out of it to get better outputs.

So rather than trying to solve this problem ourselves, we've decided to go in a different direction: let's package our work into a "module" for the SDK and then open it up to other developers to try and crack it.

Skip to now, a few months later, and not much has changed. None of the new models seemed natively capable of virtual collaboration on a whiteboard. However we've met lots of teams who think they can make it work... and want help trying.

We spent a few months pretending like it didn't matter, as if we were guaranteed that new models were coming which could produce perfect results. We built the features and identified the patterns that we would need once those new models dropped.

The short answer was, uh, sort of. They could—which was amazing—but even with creative prompt engineering, the models were pretty bad at these tasks. However, the experiments were so compelling and truly weird that we decided to push ahead anyway.

Last year, right after the first multi-modal models like Sonnet and GPT-4 were available, we did a spike on this type of experience.

Could the models recognize what the user is doing on the canvas?

Could the models create content on the canvas?

The seed of the idea was: if a communication channel works for people then it will probably work for AI, too. I like to chat with people. I like to chat with LLMs. I love to whiteboard with people. Maybe I'd like to whiteboard with AIs, too?

If you want an LLM to do stuff on the canvas, you need to:

1. get information from the canvas 
2. send that info to an LLM and generate instructions
3. execute those instructions on the canvas

The module helps you with steps 1 and 3.

2 is up to you

Starting in 5 minutes, come say hi

Happening today: Come speak with us at Visitor Hours, four hours from now.