The Glitchatorio: A podcast about AI strangeness

AI is a weird technology that does weird things.

Take LLMs as an example. Although they're great pattern matchers and text generators, there is something fundamentally unpredictable about these prediction machines. Even the way they work when they work well, remains fundamentally mysterious and produces unexpected results.

And in my own AI journey, as I started to engage more seriously with it technically and intellectually, one of the first things I noticed is the way that it's like an experimental drug.

There are a lot of promising applications, but there are also serious risks, and a range of side effects, both known and unknown.

The best-known category of AI weirdness at this point is probably hallucinations. Even here, we still don't know: Are they a symptom of overconfidence? A special category of mistakes? Or are hallucinations actually the core features of large language models, because their default operating mode is a hazy hashing of likelihoods and wild guessing?

Then there's that subset of weird behaviors that seem to indicate both some level of self-awareness and desire for self-preservation, like the shell game of evals, or the anxious deliberations on the scratchpads. It's hard to know whether we're projecting our own egos onto these situations or if they're the harbinger of something more sinister.

Whatever the case, I think the glitches and strange behavior deserve a different kind of attention than they usually get, which tends to either be "top 10 AI fail"-type listicles in the popular media or "approaches toward corrigibility" papers in the technical discourse.

There's not much in between, in the way of explaining this stuff to regular folks. I feel like it's important to have more transparency around the workings (and non-workings) of AI, so that more people understand the ways that these systems are brittle, fallible and fundamentally still experimental.

The Glitchatorio is a podcast born out of my desire to bridge the technical and popular discourse around AI, while bringing a creative, playful approach to the science. (Mary Roach is my role model here.)

I'll be talking to researchers, practitioners, and other AI experts, as well as people outside the field, for their perspective on AI weirdness. Apart from the conversations, The Glitchatorio will be an audio-first experience, putting the creative possibilities of podcasting at the forefront.

You can find The Glitchatorio on Apple, Spotify and other podcast platforms or my hosting website. Listen to the trailer here: