325 | Alvy Ray Smith on Pixar, Pixels, and the Great Digital Convergence
What this covers
Sean Carroll hosts Alvy Ray Smith, a computer scientist and co-founder of Pixar, for a conversation spanning the history of digital imaging, the emergence of computer animation as an art form, and the philosophical implications of what Smith calls the Great Digital Convergence. The core argument is that around 2000, all forms of analog media—film, videotape, printed pages, analog audio—were replaced by bits in the form of the pixel, making information independent of any physical medium. Smith traces this convergence through his own career, from his work on the first 24-bit pixel memory to Pixar's founding alongside Ed Catmull, when Steve Jobs took the company public on the strength of Toy Story and eventually became a billionaire.
The conversation moves through several distinct technical and philosophical territories. Smith explains the Sampling Theorem—the mathematical proof that a continuous image can be perfectly reconstructed from discrete samples if done correctly, meaning no information is lost in digitization—and traces how misunderstanding this principle produces artifacts in digitized film and animation. He argues that Turing's Halting Problem implies that computation's future is fundamentally unpredictable; each step is deterministic but outcomes cannot be foreseen without running the program. The episode addresses contemporary anxieties about image-generating AI by reframing an artist's role as curator and critic using technology as a Tool/Lever Framing of Technology, and contests whether current large language models deserve the label "intelligent." Throughout, Smith insists on careful definition—both as a historical method and as a way to dissolve disputes about who invented what first.
Alvy Ray Smith argues that around the year 2000 a 'Great Digital Convergence' replaced all analog media with bits (the pixel), and that—contrary to intuition—correct application of the sampling theorem means no information is lost in this conversion, even as the future the convergence enables is fundamentally unpredictable.
- All old media types converged into bits/pixels around 2000, making information independent of its medium
- The sampling theorem proves a continuous image can be perfectly reconstructed from discrete samples if done correctly
- Turing's results imply we cannot predict where computation/AI will lead; you have to run it to see
Large language models excel at remixing, remembering, distilling, and searching human creative work.
- Large language models and machine learning are trained on the products of human creativity and excel at remixing, remembering, distilling, and searching; there is no fundamental reason computers cannot be creative, but that is not what the systems we have built right now actually are.
“the large language models we've had such enormous success with... is trained on things that are the products of human creativity. And they're very good at remixing and remembering and distilling and searching”
The host suggests setting the stage with biographical context before discussing larger ideas.
- The host suggests setting the stage with some biography before getting to the bigger ideas.
“I think this is a case where maybe a little biography would be good to set the stage a little bit.”