CGI renders do use a lot of electricity relative to playing back the movie for individual viewers. It's perfectly analogous.

Model training is similar to the creation of the cgi for the movie. Both happen before anyone consumes the output, and represent the up front cost for the producer.

Both a movie and a language model can cost tens or hundreds of dollars to produce.

In both cases additional infrastructure is needed for efficient usage: movie theaters or streaming platforms for movies, and data centers with the GPUs for LLMs. This is also upfront (capex) costs.

At consumption time, the movie requires some additional resources, per viewing, whether it's a movie theater or streaming. Likewise, an llm consumes some resources at inference time. These are opex. In both cases, the marginal cost for inference/consumption is quite low.

Good Bayesian exploration is much, much better than grid search, and does indeed learn to avoid low value regions of the parameter space. If we're talking about five minute experiments (as in the blog post), Bayesian optimization should chew through the task no problem.

On the other hand, we could probably convince Cory Doctorow to write a piece about how fentanyl is really about the enshitification of opiates.

The better choice would have been Biden stepping out earlier and having a real primary, of course.