(

note on title: Viewpoints for AI was me trying to use Viewpoints (a theatre/dance thing) as inspiration for finding a vocabulary to have better conversations about AI. Reading McLuhan recently, thought...some of his things might do what I was trying to do even better.

Mclu-points for AI: Using Marshall Mcluhan-isms to try to talk about AI

)

Mclu-point #1: LLM as Extension of Human

Just as Mcluhan talks about media as things that allow humans to extend themselves, in what ways do LLMs extend us? What can we do today that we couldn't do five years ago? What could we do five years ago that we can't do today?

Things we can do today:

  • We can (practically) excrete software

It's not quite so fast that we can Matrix-style download the knowledge of how to fly the helicopter as we walk towards the helicopter. But I could potentially ask for a custom museum guide before I drive 20 minutes to the museum and then have it waiting for me when I get there? Sure can.

  • we can find needles in haystacks

We now have a tool that lives between reading things ourselves and grepping for a keyword. Our eyes can be bigger than our stomachs.

I used to time my CS coursework for the hours just before my TA's office hours. I did this because I while never ran out of hours in the day to complete my assignments, I would inevitably make some stupid mistake during my coursework because the tutorial / instructions made assumptions about what I knew that didn't hold.

LLMs can get you unstuck when you're stuck because the person who gave you your instructions didn't realize you had some gap in your knowledge.

Things we can't do today:

  • Our work is no longer rock-solid proof of our competence

Five years ago, some neat hacker-news-bait Kubernetes diagram-explainer-sandbox-whatever would be proof that you really knew Kubernetes. Now? Not so much!

  • When we see things, we have no idea how long it took the person to make it (or if a person even made it!)

The basic heuristic of, this person's website is kinda decent, they must know a few things about websites is broken. The basic heuristic of, this email doesn't have any typos the writer of it must have a competent grasp of the english language is broken. The basic heuristic of, someone just sent me a photo, that photo represents reality is broken.

It's a legibility crisis, baby! And while the things you can do today you have to participate in AI to reap the benefits of, everybody pays the costs of what is no longer true.

Mclu-point #2: Hot and Cold media

I am not going to sit here and pretend to completely understand precisely what Mcluhan really meant by hot and cold media. That said, I think the lens is interesting and useful for LLMs.

  • Mcluhan said the written word is hot. I think LLMs cool text down a bit

When Mcluhan says text is hot what I take him to mean is that if your ideas have to be transmitted orally you need to account for small digressions + alterations as the ideas pass from speaker to speaker. You can't count on your message, three jumps down the chain, to be exactly word-for-word the same.

Text "solves" this because you print the pamphlet once, and it's the same words, with the same subtle points & phrasing, for every reader.

But with LLMs, a chunk of text becomes less canonical and more of a thing to skim + start from. Maybe I take your pinecone's technical blog post and feed it to my personal LLM and say "implement this but use postgres + the postgres vector extension instead of pinecone." I think that's Mcluhan-cool. The user / the user's LLM is filling in a lot of details.

  • LLM-generated HTML is "hotter" than LLM-generated markdown

When you let the LLM generate html for you, you are allowing it to be more expressive, specific, etc. You are hoping to be shown, not told. With HTML, the LLM can create an artifact that reaches towards you a little further than markdown can.

  • The "Stochastic parrot" criticism of AI suggest (maybe?) AI is Mcluhan-cool to a fault

When media is Mcluhan-cool, that means the user is filling in the details themselves rather than having their sensed "filled up" with the experience.

To say AI is all in the eye of the beholder is saying AI is extremely cool media that sparks a strong reaction in the mind of user.

I think there's definitely something to this, for the record! I, too, see faces in car headlights and implicitly talk to LLMs like there's a person on the other end.

If this is true, this is maybe an argument for "heating up" your AI-related-media. Put the burden on the LLM to prove that it understood what you asked for by creating a "hot" media that allows as little room as possible for misinterpretation.