Andrej Karpathy @karpathy, 2h
Scaling reasoning at inference: depth vs width is a clean separation nobody is exploiting yet.
Ask Grok about this
›
Grok in the Timeline does one thing: turn Grok into a layer over the X feed. Everything else - the prototype, the visual board, the connectors, the steerable algorithm - is opt-in.
"We need to stop talking about product design in absolutes. There are no rules. Everything is made up. Do what makes sense and feels right."
Benji Taylor, design lead at X, xAIGrok shouldn't be a tab. It should be a layer over the timeline. Every page on this site argues some part of that idea. The whole proposal collapses to this thesis: the right place for an AI inside a social product is anchored to the artifact that prompted the question, not in a separate room.
The standard pattern for an AI feature is a tab in the top nav. That model fails because every interaction with the AI becomes a context switch. The user has to leave the artifact they were curious about, copy intent across, lose their place. The layer model anchors Grok to each post, and the user never leaves the scroll.
Grok's answer starts appearing within 400ms of the tap, and completes inside 1.5 seconds. The user begins reading before the model finishes generating - perceived latency drops by roughly 60%. Same trick Cursor and ChatGPT use. Works because reading speed is approximately equal to token generation speed.
Once the answer lands, source chips fade in one by one and a confidence indicator renders explicitly. Hidden uncertainty is dishonest. Showing it builds the calibration loop users need to know when to trust the system - and when to double-check.
For live events, the answer doesn't stop at text. Game stats animate in from typed data feeds. Prediction-market odds render side-by-side, sourced from Polymarket and Kalshi. A connector is not an embed. An embed is a third-party iframe with different brand, latency, and trust. A connector keeps the answer inside one monochrome surface.
The same Grok layer fires across every content domain - stocks ($NVDA earnings with Bloomberg connector), the Super Bowl (NFL Next Gen Stats), the Oscars (Box Office Mojo, settled markets), the World Series (Statcast), elections (live presidential odds), or just the general feed. One interaction model. One design language. Different data underneath.
Letting users steer their feed by talking to Grok is honest in a way no social product currently is. The mechanism is the diff - the user sees exactly what Grok will change before they commit. Letting Grok silently update the algorithm is the same product mistake as the black-box ranking everyone resents.
A list of refusals is the cleanest possible signal of taste. Every line below is a pattern I declined on purpose. The strikes fire as you scroll past.
Pick the entry point that matches what you're trying to evaluate.
The whole site lives at grok.davidtphung.com once DNS finishes propagating (see DOMAIN.md). The prototype is the fastest way to feel the design.