Large Language Models by Andrej Karpathy

An excellent introduction to Large Language Models. Two interesting point

  1. Stuart Kauffman in At Home in the Universe talks about getting “order for free” but does not say how it happens. Andrej Karpathy at minute 25:43 says, "Algorithmic progress is not necessary… We get more powerful models for free because we get a bigger computer to train a bigger model for longer. Maybe this explain why the Universe has to be so big and last so long. In other words, it’s not really ‘free.’ it just looks like it to us.

  2. LLM security concerns at minute 45:43

The Captain


AI: Grappling with a New Kind of Intelligence

A most interesting panel discussion about the current state and the future of AI. Perhaps the most interesting part is about its dangers. I think they get most of it right but they miss the really big picture.

“We Have Met God and He Is Us” Pogo 2023

The Universe has no morals, it blows up stars and wipes out dinosaurs without one whit of regret. Morals are an emergent property of human intelligence. We have two strains, creationist attribute morals to gods while atheists think it’s a human construct. In nature mutations are random and the selection process is simple, what works survives, what doesn’t work dies. There is no morality to natural selection.To make AI safe we need to train model with safe data. We are the gods who create AI morality!

End of The Captain’s crazy musings.

The most interesting part of the panel discussion, AI’s dangers and what to do about them at time 1:08:49

The Captain


Sean Carroll: AI Thinks Different

I’m part way through a Sean Carroll uTube video and I have to point out a fallacy in his thinking. He claims that Large Language Models don’t have a model of the world. Having been trained on just about everything ever written they create new sentences that sound plausible at first sight but are not necessarily accurate or true. He is right but that’s how human innovation actually works. AI does NOT think different!

Colonel John Boyd gave an excellent explanation about how creativity works. From SoftwareTimes, my website:

Boyd described his method of innovation in the briefing titled “Destruction and Creation”

-Imagine that you are on a ski slope with other skiers-retain this image.
-Imagine that you are in Florida riding in an outboard motorboat-maybe even towing water-skiers-retain this image.
-Imagine that you are riding a bicycle on a nice spring day-retain this image.
-Imagine that you are a parent taking your son to a department store and that you notice he is fascinated by the tractors or tanks with rubber caterpillar treads-retain this image.

Now imagine that you:
-Pull skis off ski slope; discard and forget rest of image.
-Pull outboard motor out of motorboat; discard and forget rest of image.
-Pull handlebars off bicycle; discard and forget rest of image.
-Pull rubber treads off toy tractors or tanks; discard and forget rest of image.

This leaves us with:
Skis, outboard motor, handlebars, rubber treads

Pulling all this together
What do we have?


That’s what LLMs do and do it rather well.

What about the claim about truth and accuracy? The LLM output is no different from what scientists do, they come up with hypotheses which must be confirmed by experiment. Most of what scientists dream up is neither true nor accurate. It is put to test and a Darwinian selection happens, what works survives, what does not work is discarded.

What LLMs do is very human like in m view. Algorithms must output correct results. Human intelligence outputs plausible hypotheses that must withstand the scientific method. Sean Carroll’s GAI would be an omnipotent god.

The Captain

At around minutes 20 to 30

o o o o o o o o o o o o o o o o o o o o o

PS: after a few more minutes I lost interest, it was mostly about how to fool LLMs


Fortune cookie says:

“You learn from your mistakes. You will learn a lot today.”

The Captain


“Intro to LLM’s”. That was fascinating. As was this, and also possibly alarming.


An AI model is just a mathematical function (a neural network) that takes some input plus some known parameters (also called weights) and produces some output.

output = function( input, parameters )

So AI “thinking” (if that’s what some want to call it) is, literally, a mathematical function.

One of the interesting things that Richard Feynman said is “Just because you know the name of a bird does not mean you know anything about the bird” or words to that effect. My point is that what interests me is how the output of AI mimics the output of HI (Human Intelligence). The mechanics of it are interesting for a different reason. That’s what Andrej Karpathy covered.

My take is that the "mathematical functions’ a.k.a. the LLM neural networks produce strings of words that sound plausible just like scientists produce theories or hypotheses that sound plausible but which are not always correct or true. That’s where the Scientific Method, a human construct, comes in to verify or falsify the string of words. Algorithms, by contrast, are designed to give the right answer.

Another way AI mimics HI

Malcolm Gladwell Demystifies 10,000 Hours Rule

Data size and iterations matter to train AI well just like training humans takes lots of repetitions.

The Captain


There is a bit of a leap going on here. For example equating LLMs and what lots of people think is next, such as some AGI in robots. No one wants a robot building stuff in a factory to make lots of mistakes (i.e. making junk that has to be thrown away) just because they are iterating a million times before they “get it right.”


1 Like

Think back to the middle ages…

Guilds were organized so that workers would learn skills from others connected with the guild. Members traditionally advanced through the stages of appren- tice, journeyman, and finally master. An apprentice was a young person, most often male, who learned a trade by working for a guild master.

One important difference between robots and humans is that humans learn one by one. Robots get their learning via over the air updates. You train the robots before you let then get to work but once they are working they feed back more stuff for robots to learn.

Tesla’s FSD Beta is a good example.

The Captain


I would agree that “mimics output” might be an appropriate comparison of AI to HI.

1 Like

An apprentice watched and helped with easy tasks, watched and helped over an over and progressively did more and more difficult parts of various jobs. In general people aren’t talking about AGI robots that assist a worker and slowly learning, are they.
Stationary type robots in factories today do difficult, dangerous, strenuous or precision jobs, exactly as programmed to “help” on an assembly line with no thought that they will watch and learn other tasks.


The Future of Artificial Intelligence

by Melanie Mitchell at the Santa Fe Institute

The Santa Fe Institute is a very special place, founded by the Los Alamos Manhattan Project scientists. It was initially focused on Complex Systems.

Melanie Mitchell recounts the early days of AI and explains how LLMs work. So far so good. Where I disagree with her is about the future. She puts too much emphasis on what AI gets wrong and misses the solution which is scale. The most powerful current supercomputers are tiny by comparison to human brains.

Tesla gets it, build Dojo, a modular machine designed to run deep learning and feed it with real world data from 5 million EVs now, many more in the future, plus millions of Optimist Humanoid Robots, and computer simulations. Size matters! With AI, the BIG PICTURE matters.

To understand AI we need to understand Human Intelligence (HI). How come that AI produces Christianity in the West, Islam in the Middle East and Buddhism in East Asia? If intelligence were perfect it would produce the one perfect religion or atheism. How come HI produces hundreds of languages, not the one perfect language? Nature is wasteful in the extreme, random mutations appear and natural selection keeps what works best at any particular time, the rest is recycled. Rinse and repeat until the universe ends. We can improve and accelerate the process but we cannot change the universal paradigm. At least not yet.

The Captain

1 Like