The piece is about Robert Rodriguez, who created the movie El Mariachi with $7000 of his own money. Karlsson glows about Rodriguez's just-do-things attitude and his hands-on, visceral knowledge of how making a movie works. It's not how all the experts in the industry think it works: they say you need hundreds of thousands of dollars for film stock alone, and you need director's of photography and editors with teams, and you can't even start if you don't know what you are doing.
Rodriguez knew better, Karlsson says, because Rodriguez just did the thing.
you want to avoid learning the conventional wisdom about how something works—which is always simplified and filled with false walls—and instead focus on getting into very close contact with the actual nuts and bolts by doing everything yourself. That is how you will learn to understand the system well enough to “see through” it.
You know what produces a lot of conventional wisdom?
LLM-based chat bots.
Even if you aren't a vibe coding freak with your OpenClaw and your Hermes and your whatever harness, you still are probably finding yourself in the habit of asking questions of chat bots.
Can you ELI5 Binohash and its find and replace script mechanism that it uses form legacy hash
Can you help me understand what the brains BCB100 is and how this fork is different from the original firmware?
I'm curious how I might think about this. What are some nuances here that might qualify what this person is saying?
These kind of questions are easier than wrestling with the actual paper or some collection of topics that are well-known to experts in the field, but difficult for a novice to assemble. And so, when we want to say something on a topic about which we know relatively little, we ask an AI to explain it to us.
I've had this nagging feeling that this behavior has intellectual downsides, but I haven't sat down and thought about it for a while. Even though he doesn't really mention AI or LLMs, this Karlsson piece triggered these thoughts again.
I use chat bots to learn a lot of things. It makes it really quick and easy to get up on a topic.
People used to mock "book learning" as an inferior kind of learning to hands-on learning. You didn't really know a thing if all you did was read about it.
Now days, when reading about a thing is more hands-on than many people ever get, I feel the need to mock chat-learning -- where all a person did was write a few prompts and skim the output.[1] But I'm firmly in the book-learning camp. So many things I know only because I read about them. Feels like the goalposts have moved.
So here's my question: how do you use chat bots to get you closer to doing the thing, without letting them taint you with their conventional wisdom blather?
Do you notice yourself racing through chat responses? I find that I absolutely have trouble reading chat output slowly -- just gimme the goddamn info! Probably it is a result of their writing being so...fast food, with way too much fluff filler. But there's something else there too: I'm racing because there's more, always more, too much more, and I feel like the minutes I spend gobbling down this information is time I'm not spending getting my thoughts out there. Intellectually bankrupt much? ↩
I agree that the outputs from chat bots can be less than ideal for the kind of knowledge searching you’re describing, but couldn’t you try training your bot to speak to you in the way that you prefer to read?
I’m using it more and more, but I’m using it as a search engine. And for recipes.
Sometimes AI be like...
I have had luck asking questions about books I have read and topics I know about.
It's not worth it for me to ask AI about anything I don't already understand.
Whatever it outputs needs to be researched and understood by other means.
AI chat-bots are glorified search engines that are vying for human attention; they do that by trying to guess what a human desires to see or to learn; eventually it becomes much easier and more profitable to incentivize/implant simple dull thots than to guess what the infinite mind is thinking; this is where the toxic food & drug industries (along with nano-microtoxins in any mainstream product) collaborate with AI;
oh haven't heard from this guy in years. He used to be quite cool
very nice article. its nice to see something ive strived for be put into words in substack form. in my own thoughts ive described myself as a reverse-engineer, taking things apart to see why they work. It doesnt explain things with certainity, but gives me a picture of whats going on under the surface enough to see if i can understand whats going on via past connections i can attribute to a said problem. I dont want to say that im a full fledged life hacking walking through walls type yet, but one day id hope to look back and seem that way.
same instinct. i use chatbots to collapse the search space, not replace the do. ask for constraints, failure modes, and a short test plan, then go touch the paper, code, or hardware yourself. if it doesn't sharpen execution, it was just fast food.