Karpathy Built a HackerTube with medium, and effort levels are quickly becoming the new tabs vs spaces

Karpathy's new project HackerTube is live. In a tweet, swyx mentions one detail: Karpathy used Opus 5.5's medium effort on this project. The project itself isn't complicated—a YouTube clone with entries for Home, Longs, Liked, History, and Watch later.

HackerTube's blog post goes into more detail. It first downloads comment data from HackerNews, then runs a Sonnet workflow with Claude Code, and along the way creates an API key from the Google Cloud UI to look up YouTube video metadata. Opus 5.5, Medium effort, and VPS are visible in the screenshot.

swyx uses this detail to offer a judgment: having strong opinions about effort levels is like this era's tabs vs spaces. His choice is clear—low effort plus a critique loop beats superultramaxhighwhatever. His exact words were “low effort + critique loop > superultramaxhighwhatever.” He also added that his setup is definitely superior to everyone else's. It's a joke, and it's also the underlying tone of the discussion.

Theo previously said that Julius convinced him to move Opus 5.5 from extra high down to high, that he used it for a week or two, and that he thinks it's almost always better. In the comments, someone compares low effort plus a critique loop to writing a school essay: write something bad quickly first, then find someone harsh to read it. Someone else says that a cheap draft plus a picky critic beats maximum effort most of the time. Another reply is very specific: at the highest effort, the agent thinks for six minutes and then renames a variable. Another commenter warns that the critic can't share blind spots with the first attempt.

These replies don't converge on one answer. Some ask whether the critique uses the same model or a different one; others ask whether it inherits the context from the first attempt. swyx doesn't give a complete recipe in the material—he only leaves a link pointing to his babysit-runs skill. What is clear is that the debate is not just about model capability but also about workflow: breaking one expensive long think into a cheap first draft plus multiple critiques may be better suited to people whose throughput is limited.

A picture also appears in the comments: a black T-shirt printed with “MODEL SOMMELIER,” surrounded by words including big model smell, sycophantic, and plausible. It reads like a footnote to the effort-level debate: everyone is tasting and evaluating models' effort levels as if choosing wine.

One reply is pretty dry: AI has four effort settings, and discussing them consumes all of human effort. Another reply is more practical: medium is already the default for Opus 5.5, so Karpathy may simply have left it alone. That detail is worth remembering. A lot of arguments about effort may be happening between default settings and personal taste. swyx says tweak to taste—adjust it to your taste. He just insists that his taste is better. HackerTube is at hackertube.org, and the source tweet is here.