Logic and Reason with Robotic Agentic Artificial Super Intelligence: Storytelling and Life’s Mysteries of Fact and Fiction

Disclaimer: Everything below is a mix of what I observed and heard during the event. The goal isn’t to pinpoint "who exactly said what," but to share (usually) an outsider's view and overall perspective on these industries. I’m not here to act as a definitive firsthand source—readers should do their own research. I hope this inspires you to attend events, explore new industries, and hear what leaders are presenting. These notes combine my observations with thoughts on how things could run smoother and how ideas connect (IMO). I’m not an expert, you know? Just hanging out in the room with them. Enjoy.


These blogs have been a’lacking (though your girl has NOT been slacking) - but it’s time to go back to some lectures and events and learn from others. This event in particular is super spot on to what I’m doing with Kelly Tutors work + the conversations I’ve been having with trusted minds. Like… AI only knows what it’s told to be true. It can’t sometimes use common sense because it only knows what is input. For now. And this talk is about how AI responds when something feels a lot like storytelling. Actually, I’m not sure what the point of this lecture will be, but plenty of it sounds SUPER related to my work, and I’m excited to go

Why Attend: Well, attending lectures at the local university - both - feels like a lifehack and a (what’s the word? Retribution? Making up for my lack of attending lectures with focus when I was back in school) - so I feel like I’m a master of it all with my lifelong learning passion, which so naturally emerged. Then once the lecture is done, I’ll jump onto a video call with two amazing women - working on our plans to host kids’ events - then I’ll go to another event tonight. Should be super super super fun and fascinating!!


Photo Collage + Commentary:


Notes from the Event:

Okay, it took a lot of effort to get myself to get up and go to this event, but I’m typing now from the lite rail train. This event is hosted in a building donated by Bill Gates. Lame. I’m super not a fan of him. But I don’t think I’ve ever been in that building, so it’ll be interesting to explore.

I arrive… The speakers awkwardly talk about how there aren’t many people in the audience and maybe they marketed it wrong by not including “AI” in the title.

The presenter says he’s trying to imagine what AI would be like without Leon and that Leon hardly needs an introduction because of his influence on AI. To just give some examples, his favorite work was neural nets and structured prediction - we don’t talk about this as much anymore. We’ve learned about document compression; he was an early thinker in reasoning. A big thing is about terms we use every day… for example “scaling laws”

The presenter is wearing his hat, i forget what it’s called. The small jewish hat. Is it a holiday?

He said Have you heard of Socrastic gradient descent? Stochastic.

Leon changed the way we think about generalization on a budget and SGD.

They met in Lisbon, and he was so delighted to hear about the world. Thought it’d be good to bring him here to Seattle.

I got hte audience to start clapping.

The origin of this, though, was his own idea, from a couple of years ago. Tech was changing rapidly, and he wanted to understand what research should be.

He wants to share this long journey. IT sa bit strange but let’s start with the short story of AI

This starts with Homer (some people laugh) in BC with the golden handmaiden - the third one. He’ll pass on the Golem of praying. But the one you never get to. An engine in a wooden machine “there moved swiftly to support their lord handmaidens wrought of gold in the semblance of living maids. In them is understanding in their hearts and in their speech and strength, and they know cunning handiwork by gift of their immortal god.”

Also, Gulliver’s Travels 1726: “Everyone knew how laborious the usual method…”

Then in 1950, I, Robot: “Well then, a few thousand workers of Tientsin are temporarily out of a job”

Snoop Dogg is a smart investor: “Well, I got a mutherfuckin AI right now that they did for me. This nigga could talk to me. I’m like, man, this thing can hold a real conversation? Like real for real? Like it’s mind blowing cause I watched movies about this. Are we in a fuckign movie right now?”

  • Are we in a fucking movie right now?
    - suddenly bosses want AI everywhere

  • Scientists have lost control of hte narrative

  • Talented engineers are working to make AI of science fiction real.

  • AI must sell briskly to recoup costs

AI = achieve more than you deserve. You dont’ need the list assistance, study to do things… what could go wrong?
AI engineers understand each component of hteir AI systems. Why these pieces add up to some kind of intelligence that resembles what was compromised by science fiction literature nd centuries of popular culture.

Why should an LLM fulfill the expectations of centuries of literature and decades of scifi

-Skeptics speak of stochastic parrots, able to fool a naive observer with mimicry yet doomed to regurgitate the countless fallacies found in their training data.

Many engineers workign to make AI real are racking up successes.

So let’s first go into Recipes:

  • DEEP LEARNING: training data

  • This guy is talking super super fast an in a super strong french accent. THey have subtitles on the screen - but i do appreciate his speed and pace and he’s cool

If you take a machine and ask: Why is the sky blue

  • you may get a response like: Why is the grass green? Why are flowers pretty? These questions no longer worry me cause your eyes are green. My sweet, sweet, sweet….

Post-training, a responsive chatbot learned to say: “The sky is blue becausethe scatered light….”

  • Post-training, we can fine-tune the model with this training.

The magic only works cause of pre-training.

The trickiest part is making sure post-training does not move the weights too far away.

Chatbot trainers use post-training to make the machine resemble what people believe it should be

INDUSTRIALIZED:

Leaders decide AI should be able to do X

Team collects data for X to find how to simulate X.

Team creates a training environment for X

Team adds to post-training routine

Repeast

So how far can this take us?

  1. None of this would work without the pre-training step?

    1. Not enough data to directly train anything like this

  2. Pre-training makes this model fluent 100% of the time. Post-training is a hit-and-miss

    1. Models still hallucinate answers, forget instructions, etc.

Old-style machine learning used to process billions of checks every year for 15 years, starting in 1996.

  • End-to-end differentiable model for the whole task, trained by optimizing a suitable cost function.

Then look at language: Shannon from Bell Labs in 1948; they modeled sources. In telegraphy, for example, the messages can be transmitted as sequences of letters. These sequences, however, are not completely random. In general, they form sentences and have the statistical structure of say, english. The letter E occurs more frequently than Q.

The number of words they use, doesn’t look the same at all - the frequency.

English is a mix and distribution of languages. Look at English versus Chinese versus French versus German. It doesn’t make any sense at all.

Folklore is that if you train a machine to replace the general distribution of English, it's going to look smart, but that doesn’t work.

One of the most important functions of language is to express new ideas. If you keep expressing the same ideas, then mankind wouldn’t go very far ahead, right?

In language, you have a structure, and you can shuffle and apply transformations and change it to new things. This is what we call compositional models. And we tried to learn compositional models for a long time, and we essentially failed, right? In the 90s and the early 2000s, there was a lot of work in unsupervised grammar. IT was frustrating.

But LLMS can now change the sentence and the form. It knows the verbiage. There is no agreement on hos this is happening. People sy “just scale,” and invoking scale is denial. Not knowing is no option; I want to know.

Understanding is the highest priority.

Let’s look at Borges (1941) - its a good way to look at language. It can help us look at LLMs in a different way. This novel was called The Garden of Forking Paths. IT’s a German spy in England who was Chinese.

  • It’s not just one story; it’s the branches of the story.

Not just all texts that have been written but also all texts taht make sense.

Find an occurance of hte words printed in the tapf in teh garden of all texts. Read the word that follows that occurrence and prints it on the tape after another. Then repeat. That’s an idealized language model. Always generate text that “makes sense”

  • but each word added to the tape narrows the subset of possible continuations in oru collection. Each addition constrains the story, the characters, their roles, their ideas, their future, and the same time…

The only driving force of the idea machine is not necessity.

  • this guy is hard to follow cause he kinda writes stuff on the PPT differently than what he says, but not always.

The aiblity ot recongize the demands of a narrative if a flavor of knowledge distinct from the truth.

What is true in the world of the story need not be true in our world?
- As new words are printed, the machine borrows facts from training data (not always true) and fills the gaps with plausibly invented details - not hallucinations but confabulations.

True of False: Sherlock Holmes lives on 221 B Baker Street?

  • Sherlock does not exist,

  • yet this is better than claiming he lives in Italy

  • So, in a fictional world of the books, yes.

But if a kangaroo had no tail, would it topple over?

A perfect language model is a machine that prints fiction on tape - but we’ve been sold an artificial intelligence worthy of science fiction soon, with perfect reasoning abilities, encyclopedic knowledge, humane manners, and guaranteed reliability.

Is this bait and switch?

  • It's complicated. Can we align a machine to be similar to what is promised?

Alignment:

  • hte garden of all factual texts - not all texts written, but texts that are true in our world

The curse of factuality: representing the set of factual texts is much harder than representing the set of all meaningful texts. In the distant tail, the aligned machine can no longer rely on the rich composition of language. The aligned machine only gives what it has been trained on.

  • But can we use deductive logic?

So the good news is that models can apply reasoning patterns (even if wrong), but incomplete specification undermines informal logic.

You need to be able to make stories before you can reason. Children need to imagine before they can reason.

You need to understand the sequence of events and be able to imagine how events could have unrolled under different conditions or decisions. We often formulate alternative timelines using common sense and knowledge. We often read battles described as fictional battles.

SO, we humans are fiction machines. Think of the last thing you thought about - was it a raw fact or something you made up?

Weather is an expression of the mood of the gods. Right? Or is the weather a dynamical system obeying some flavor of Navier-Stokes? Or the weather is a mass of warm air coming from the south and a meeting of a mass of cold air from Canada, which causes nasty thunderstorms.

  • Typically, the weather is one week different between the USA and the UK

  • Which story do we want to take? Belive?

  • Imagine a machine that tells a story about a machine that tells stories. The outer machine can produce a story that names certain facts nad decides which of the inner machine's stories best predicts these facts. So “i think because I am” is true?

When you ask AI about why it can’t talk about Tiananmen Square, and it’s not allowed. It has policies to follow and stay within guidelines.

  • So many AIs have these guidelines. So it has to figure out how to be useful but follow its guidelines.

So what is the long-term impact? What we are going to learn about ourselves is far more consequential and lasting than business changes within a few years.

What are we? We make a story from fragments that come from our culture. We are fiction machines. We use our minds. They’re manifestations of our stories and our culture. We are far less individual than we think we are.

Do not use AI to avoid thinking; use AI to become smarter.

A story may help us think - the talk here is only a story.

AI is a manifestation of our culture. Every discipline from our culture has an impact on it. It’s a challenge, but it's the only rationale we can have.

Understand that making up stories is the path to intelligence. Understanding “why” is more important than making incremental improvements.

  • But what are we missing (asks the audience) - our understanding of intelligence is exhibited by LLMs. We generalize better than LLMs? To what extent? LLMs draw from many fields of mathematics and develop complex solutions without tiring. They are not always elegant, but they exist. We are better at generalizing differently. But this can not be forever. Look at the pace of AI’s successes; it’s clear. Or not clear at all.  

  • They are able to find structure in the data. If we have a structure to find the parameters of the structure, we know how to do it essentially. Structure was always the problem.

  • The Garden of Forking Paths is very interesting and most direct.

Why did we stop telling stories? Every few years we hear eggs are good or eggs are bad - we restore that story instead of a richer data set about eggs. We compress things into labels.

What is bad, egg whites or egg yolk?

  • To him, that’s an example of how our reasoning patterns are more driven by the set of stories around us than by the rational view. People who change their minds in the face of evidence are rare. Did you have a fact or a story in your mind? Sometimes it’s a crazy story with no facts or science.

  • My own mind is influenced by what I read, but there’s a limit. I try to read more. But I get tired.

I spoke with the speaker and his friend after the event. We talked about how restrictions on AI can always be broken, just depending on how you prompt - which made me think about David R. Hawkins research on Power vs Force, coming to the same conclusion too - but those guy hadn’t heard of that. Instead, they brought up Asimov’s Laws and how all the laws can be broken.

  • This made me even more hopeful for the future.


Until next time, I wish you the motivation and success to search for opportunities around your area. Search and explore: Who is out there giving talks? There are new things happening all of the time.

Find relatable or interesting topics you like and check them out! Maybe even something hosted at a cool venue, if there’s no other reason to go. Let’s see what you can learn and discover not too far from home. 😊

Previous
Previous

Sales Intelligence + Customers + Data Organization

Next
Next

Completely Burning Out from Hosting Nine Events in Five Days: Seattle AI Week with Kelly Tutors