JetSpec asks whether a frozen autoregressive transformer can draft an entire future token tree in one pass, keep each branch causally honest, and turn larger draft budgets into lower wall-clock latency on long generations.
The base model stays autoregressive. The draft path changes shape.
Use the steps to watch the tree shift from root state to accepted prefix.
The claim lives in the mask: path-conditioned, but still parallel.
The JetSpec bars use the exact spoken values where the transcript gives them. Baseline bars are realistic reconstructions that preserve the episode's ordering.
The high-budget story is not just wider trees. It is longer valid prefixes per verify round.
Click a method. Top-left is the sweet spot: coherent branches without too much draft overhead.
JetSpec improves one corner of the space rather than ending the debate.
This heatmap visualizes the episode's skepticism: once drafting gets wider, the next wall can become verification, KV bandwidth, or scheduling.
Compact source list from the episode prompt, plus the related serving episodes mentioned on-air.
DDDD.DDDDD found no extra arXiv IDs beyond the known source paper: 2606.18394.