Good design still depends on making software feel fast, trustworthy, accessible, and delightful, even when AI makes feature development quicker.
2
Design includes every user touchpoint, from the interface and API errors to documentation, emails, and social posts.
3
Small implementation choices, such as server-side rendering, semantic HTML, browser autocomplete, progressive disclosure, and prediction cones, can remove friction from real product experiences.
Summary
John Pham argues that AI has changed how quickly teams can build software, but it has not changed what makes software good. Feature parity is becoming less useful as a differentiator, so teams should pay attention to how products feel across every interaction. He presents four principles: speed, trust, accessibility, and delight. Using SF Compute's onboarding and GPU purchasing flows, he shows how these principles work in practice. The examples include fast first paint, small assets, clear expectations, local time zones, reduced-motion support, browser autocomplete, stable layouts, and forms that adapt to how users think. Delight comes from details such as a San Francisco fog effect, a GPU habitat camera, and an onboarding ending designed around the peak-end bias. He also explains a Vercel nested-menu interaction that uses a prediction cone to anticipate pointer movement.
Pham says design is not limited to polished pixels or animation, and it is not owned only by people with a designer title. Anyone who builds and ships a product participates in design. The product includes every place where a prospect or user interacts with the company, including UIs, APIs, API error messages, documentation, emails, and tweets. Design affects whether the user feels slow and scared or fast and powerful. Since AI makes it easier for competing products to ship similar features, these human qualities become more important to the product experience.
Speed means removing the feeling of being slowed down
Pham breaks speed into perceived time, feedback, and jank. For SF Compute's onboarding, the first paint happens in under 300 milliseconds, which he compares with the roughly 400 milliseconds he says most people can perceive. The team uses server-side rendering, caching, small bundles, and limited external assets. The flow runs at 60 frames per second or more, so browsing does not feel stuttery. Pham also treats LCP, INP, and CLS as scores the team can put on a scoreboard and try to reduce.
A complex visual effect can use a very small amount of code and data
SF Compute's fog effect begins with one transparent cloud image. The team duplicates it, offsets the transforms, varies opacity, and changes the speed of each layer. Fog density is controlled by changing the opacity and the number of stacked layers. The result uses one 14-kilobyte image and no third-party JavaScript. Pham presents this as a way to keep an expressive visual fast, browser-compatible, and sharp on retina screens.
SF Compute has to collect information such as a social security number and address because of US export laws, even though users may expect a cloud provider to ask for less. The onboarding flow reduces that friction by saying that it will take three steps and under a minute. Creating a business account is described as one extra step. The progress indicator updates as users move through the form. For required onboarding calls, the product says the call is not a sales call, explains that engineers will conduct it, and promises to suggest another provider if SF Compute is not a fit.
Accessibility is part of making the product usable
Pham says repeated motion can make some people seriously ill, so the fog animation respects the browser's reduced-motion preference and pauses when that preference is enabled. Forms use the browser's autocomplete attribute to fill addresses and other inputs. The design avoids layout shifts that can move a button just before someone clicks it. Progressive disclosure splits the onboarding into three steps instead of one large form, which reduces cognitive load. Pham also calls for labels on every input, semantic HTML, and testing with screen readers.
Delight comes from care and the details people remember
Pham says SF Compute uses colors, metaphors, and other devices to make onboarding feel less rushed. The flow ends by revealing a beautiful San Francisco scene, using what he calls the peak-end psychological bias. He also shows a GPU purchasing page with a live camera feed aimed at the GPUs a buyer is selecting. The effect is built from stacked divs, a small looping video, and browser primitives. Buyers can choose units such as days, hours, or weeks instead of doing conversions themselves. The product may recommend a cheaper order even when that reduces immediate revenue.
Prediction cones make nested menus easier to control
Pham's final example comes from Vercel. In a nested menu, a naive implementation can switch menus while the user moves the pointer toward a submenu. A prediction cone tracks the likely direction of the user's pointer and delays the switch when the movement suggests that intent. This lets users move through the menu without the unusually precise pointer path that a basic implementation demands. Pham presents it as a small interaction detail that makes the menu feel much easier to use.
"We should basically build all of our products where everyone can use them because we're building products to be useful."04:13
Who should watch
Teams building AI products that are adding features quickly but have not reviewed the full experience across onboarding, forms, loading states, and errors.
Engineers who need practical examples of performance work, accessible web interactions, and lightweight visual effects.
Product designers and developers working on flows where legal requirements or complex choices create friction for users.
