Courtesy Joan Westenberg, another way in which outputs are not enough:
Bach copied Vivaldi for years before the Brandenburgs. Picasso painted in classical mode for two decades before cubism. Joni Mitchell played other people's standards in coffee houses for years before Blue. Hunter Thompson typed out The Great Gatsby and then spent fifteen years writing journalism that no one could ever mistake for Fitzgerald.
These are long, long stretches of work that looked, from outside, like nothing was happening. Inside, the inputs were being broken down into their components, sorted, and rebuilt as something the practitioner could call their own. The temptation, especially now, is to skip this phase by trusting a model to deliver the surface without the years. That temptation should be refused for the same reason a virtuoso refuses to lip-sync: the work that bypasses the apprenticeship produces no apprentice, only an output. And an output is not enough.
When we train a model, the inputs are broken down into their components, sorted, and rebuilt as the model. That process doesn't change us, though. The model's output is just another input for us to consider.
Several colleagues and former students have told me that LLMs have helped them learn a new area more quickly than they might otherwise have been able. One element common to all their stories is that they have deep background knowledge they can use as they process the new information. Another is the time and energy they spend processing the new information, engaging with it — and with more reliable sources — in ways that successful learners always do.
Our new tools may lead to more personalized education, but they will not eliminate the need to do the work.
But Shannon entropy and Kolmogorov complexity are measuring fundamentally different things. Shannon entropy measures the statistics of outputs. Kolmogorov complexity measures the structure of the generating process — the program, the mechanism, the cause.
Deep learning learns the outputs. It does not learn the program.
The quoted passage is from Shannon Got AI This Far. Kolmogorov Shows Where It Stops. It explains — better and more completely than I've ever articulated to myself — why neural networks and statistical approaches have never appealed to me.
Even back in the late 1980s and early 1990s, these systems could be very good at producing answers. Today's LLMs operate at a new level and really do produce amazing outputs.
But even at their most impressive, they interest me only as a possible substrate for something more. The results are cool, but they don't answer the question that motivates me: What's the program?
One thing I like about the Shannon/Kolmogorov article is that it does a reasonable job of describing how the current statistical models, despite their limitations, might help scientists and creators: These systems may well be useful tools for humans operating at higher rungs on Judea Pearl's ladder of causation. In doing so, they will magnify the abilities of those humans best able to operate at the higher rungs — and especially those able to move with some facility among the rungs.
I'd still rather work with other humans myself.
Here are two passages from an old conversation with musician Henry Rollins about writing and creating that stood out to me as I wrap up another year of teaching:
I'm a shipbuilder. I don't want to sail in them. I want you to sail in them. I'm just happy that they leave the harbor so I can have an empty workplace.
And:
That hesitation, that's what holds a lot of people back. That's why I never say, "I'm a writer," because I don't want to shoulder that. I just want to do some writing. "What would a writer do in this situation?" I don't know, man. Ask one. And don't tell me what he said, I'm busy.
Hat tip to Jen Myers, who quoted yet another blunt passage in her recent newsletter, along with this piece of advice, "Sometimes, you just need Henry Rollins to set you straight, you know?"