The Multiplex

An place for my unorganized things

This is a place for me to put thoughts that I haven't fully digested or that I haven't found time to write about.

LLMs for curious NICU parents

The speed with which I became very knowledgeable on neonatal hypoglycemia care in the intensive care unit was astounding, and I attribute it to asking Claude.ai to use HTML, CSS, and JS to create an infographic presentation on everything I should know about the process, the treatments, the different underlying causes, etc. I came away thinking hospitals should probably be using this tech to cheaply generate educational material for patients.

Crucial: have another LLM session fact check everything produced, and link sources.

The case for letting them train the models on your content

Newton wrote to Robert Hooke "if I have seen further it is by standing on the shoulders of Giants." From a certain perspective we can generalize this to "everything is built on all that came before." To what extent can an author say that their work is the product of everything they read up to that point + their lived experience? To what extent is everything produced today some permutation of components that preceded it?

This logic applies to large language models; they're fed all the text their creators can get their hands on, some high-dimensional matrix of floating point numbers compresses all of it, and everything the resulting model produces text. It sees what it sees by standing on the shoulders of all the text (avoiding multi-modal stuff here) that came before it. I suppose if this is the direction we're going, I'd at least like to make up part of that foundation.

Writers as agile methodology practitioners

Ray Bradbury was talking to some university students about failing fast in his earlier writing years, preferring short stories over novels because he was afraid of spending 1-2 years on a project with nothing great to show for it. With short stories you get rapid feedback. On top of that, if you write a story a week, at the end of the year the odds are low that you've got 52 bad stories. Example of distinct and distant fields having something to teach one another? Writers like Bradbury were failing fast and practicing agile well before these terms existed.

The terribly inefficient cycle of compression and decompression

One of the things that profoundly disturbs me in this age of LLMs is the vast quantity of generated text that simply never gets consumed (by a human). In this case I am referring to cases like

Person A: "write an email to Person B that covers these bullet points"

Person B: "turn this email from Person A into bullet points"

Of course reasoning models generate text that never gets consumed by a human all the time via reasoning tokens. That's not the stuff I'm talking about here.