Theo says T3 Code's nightly is "almost fake" fast
A second party says he can feel the difference, which most PR-count brags never get.
Julius's agent spent 30 hours merging close to 400 pull requests aimed at web performance, and now someone other than Julius says the result is noticeable. Theo, posting from the t3.gg account, wrote that T3 Code already felt fast but that the latest nightly's performance is "insane" and "almost feels fake, like the data is hard coded or something."
Mobile next. Would be a shame if I had to freeze merges…
The numbers Julius reported were, as covered earlier, 86% fewer renders when another thread streams, up to 35% less JS CPU, and a roughly 10% smaller heap. Mobile is next, he said, adding that it would be a shame if he had to freeze merges.
Why the second voice matters
Merged-PR counts are the easiest thing in the world to post and the hardest thing to evaluate. Four hundred PRs could be four hundred renames. What makes this one different is that the claimed effect is the kind a user feels without being told where to look, and a user did. Theo did not cite a trace or a benchmark run, he said the app feels different. That is weaker than a profile and stronger than a changelog.
It is still one app, measured by the person who ran the agent, confirmed by one person who uses it. Nobody has published before and after traces, so the 86% and 35% figures remain self-reported. The honest reading is that an agent grinding on a narrow, measurable objective over 30 hours produced something a human noticed, not that 400 PRs is a number worth copying.
The part that is actually transferable
Web performance is an unusually good fit for this kind of run. Render counts, CPU time and heap size are cheap to measure, the agent can check its own work on every change, and a regression shows up as a number rather than an argument. That is the condition under which long autonomous loops have been producing receipts, and it is not the condition most feature work is in.
Julius has not said what the agent setup is, how many of the 400 PRs were reverted, or how the measurements were taken. Until that lands, what exists is one strong anecdote with a corroborating witness.
T3 Code already felt fast but the latest nightly's performance is insane. Almost feels fake, like the data is hard coded or something.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.