Canva Brush Edits in Place Without Showing What It Left Alone
Every one-change request carries an unspoken second one: leave the rest alone, because the rest was approved. This week Canva shipped Brush, a model it post-trained to edit designs by writing and running code against their elements. With that architecture, Brush should already know which elements it changed, but Canva doesn't show you. Edits apply automatically, and history just says "Canva AI." The omission is strategic too. Canva built its own editor to own revision while assistants take first drafts, and a self-reporting edit is the thing they can't copy.
Canva Brush Edits in Place Without Showing What It Left Alone
Every one-change request carries an unspoken second one: leave the rest alone, because the rest was approved. This week Canva shipped Brush, a model it post-trained to edit designs by writing and running code against their elements. With that architecture, Brush should already know which elements it changed, but Canva doesn't show you. Edits apply automatically, and history just says "Canva AI." The omission is strategic too. Canva built its own editor to own revision while assistants take first drafts, and a self-reporting edit is the thing they can't copy.
Product Verdict
Runway's App Studio, announced October 26, turns a creator's Workflow into a hosted app. A Workflow is a saved chain of Runway's image, video and audio models. In the app, the creator locks the prompts, models and settings and exposes a few inputs as a form.
Each of these mechanics has a precedent. Only a Custom GPT's builder can edit its instructions and publish updates. Claude's AI-powered artifacts already bill each run to the viewer's own plan. Runway adds a different thing behind the lock: a multi-model media pipeline plus the creator's Brand Kit of logos, colors and fonts. An agency can hand a client a generator that stays on-brand.
The billing rule matters more. Every visitor needs a Runway account and credits, so each shared app recruits a paying customer for Runway (moderate confidence).
Before trusting App Studio with client work, find out who sees a revision. Runway says apps evolve with the Brand Kit but doesn't say whether shared apps change without notice.
Interaction Patterns
Voice in AI products now does three distinct jobs. It records what other people say, as meeting notetakers do. It takes input you mean to send, which is dictation. And it can now conduct: one spoken session that directs agent work, and that work keeps running after you stop talking.
Record and input are merging into one product. Granola, a notetaker, added dictation this week. Wispr, a dictation app, had already added notes. Apps that do only one of the two will keep losing ground (high confidence).
Conducting is the structurally new pattern. Perplexity is the only shipped example. Perplexity owns its agents, but the model that hears and answers you is OpenAI's gpt-live-1. Siri AI shows the same dependency at the operating-system level. Apple's servers and policies decide which countries it works in and how much you can use it.
When you assess a voice company, ask which of these layers it owns: the speech model, the agents, or the servers and rules.