Epicycles All the Way Down

Rohit Krishnan’s essay "Epicycles All the Way Down," published on the Strange Loop Canon Substack, argues that large language models look like engines of understanding but act more like over-fit pattern matchers. It has been reported that Krishnan opens with a humbling personal anecdote — his failed experiment trying to learn poker purely from outcomes — to draw a line between “knowing” and truly “understanding.” Funny, human, and a little rueful: sometimes the math matters more than the vibes (and yes, beer helped the vibes).

What Krishnan argues

At the essay’s core: we keep bolting on epicycles — tweaks, tricks, and extra constraints — that push models toward better behavior without changing the fundamental generator. Krishnan notes that the space of possible generators far outstrips the space of outputs, and neural nets’ inductive biases pick among many explanations. He raises a theoretical sting: which collections of patterns will reliably rule out the wrong principles and surface the true generative ones? He even invokes Gold’s theorem to frame why positive-only examples might doom certain learning programs.

Why it matters

It has been reported that Krishnan warns these design choices make failures look like flash crashes or market blow-ups — sudden, sharp, and surprising — rather than Hollywood-style emergent apocalypse. Allegedly, that’s a comfort and a worry at once. The essay lands where theory meets practice: demanding new thinking about generators, not just better epicycles. Who wants more duct tape on a jet engine when you could redesign the turbine?

Krishnan’s piece has provoked discussion across Hacker News and the broader ML crowd, partly because it mixes a clear technical question with a very human moment of failure and learning. It’s a tidy reminder: sometimes re-deriving the formula matters. And if you’ve ever lost at poker by trusting vibes over odds — well, you’re in good company.

Sources: strangeloopcanon.com, Hacker News