399: GPT-4 vs Humans, Meta's Cubicles, Waymo, Microsoft, Stability & OpenAI, Stripe, Diversification, Credit Suisse, and Railroads

"won’t that be closer to magic?"

It is easier to disguise ignorance than knowledge. —Nicholas Nassim Taleb

🤖📝🎓 Great follow-up on GPT-4’s various exam results by friend-of-the-show Mark Heyer:

GPT4 scores high against humans ON THE DAY THE HUMANS TOOK THE TEST. Unlike humans, GPT4 would perform better after years - whereas humans get progressively worse beginning on the day after celebrating passing the text.

This is a very valid point!

When humans pass those kinds of tests, they typically ramp up to them by studying for years, and then during the weeks before the test, they have to study extra hard to saturate their memory with what they need to know.

But as soon as the test is over, this hard-won knowledge starts decaying and the median student wouldn’t do nearly as well on the same test a year or five years later (unless they study hard again).

The machines can take the test without special preparation, and then take a variety of different tests in completely different fields, without ever forgetting what they know.

So when we compare AI test results to human results, we have to keep things in this context. We’re talking about very different sets of capabilities.

🐻 🤖 This morning, my Bearly.AI app had a pop-up saying I now have access to the GPT-4 API. That will be fun to experiment with! I’ve already had it generate hypothetical season outlines for Firefly season 2 and Deadwood season 4… 🤠

🪄🏋️‍♂️🎙️ I love this anecdote by Teller (from the Penn & Teller duo) that friend-of-the-show Dwarkesh Patel shared when he was asked about how he came up with interview questions:

Make the secret a lot more trouble than the trick seems worth. You will be fooled by a trick if it involves more time, money and practice than you (or any other sane onlooker) would be willing to invest. 

My partner, Penn, and I once produced 500 live cockroaches from a top hat on the desk of talk-show host David Letterman. To prepare this took weeks. 

We hired an entomologist who provided slow-moving, camera-friendly cockroaches (the kind from under your stove don’t hang around for close-ups) and taught us to pick the bugs up without screaming like preadolescent girls. Then we built a secret compartment out of foam-core (one of the few materials cockroaches can’t cling to) and worked out a devious routine for sneaking the compartment into the hat. 

More trouble than the trick was worth? To you, probably. But not to magicians.

It is quite clear to me that the reason why Dwarkesh is so good at what he does is because he does everything to prepare.

The lesson is that to make it look easy, you often have to work quite hard. Those who win often put in way more effort than most people would guess.

It’s not so much that the competition couldn’t do the same and go as far, but they wouldn’t, and that’s where differentiation comes from.

Here’s the whole Teller interview.

🤖🪄🚙 Here’s a video of two older gentlemen riding around in a self-driving Waymo car, with no one in the driver’s seat and the steering wheel turning by itself:

My first thought was that it would be so cool to be able to give a Waymo ride to time travelers from a century or two back. If it’s this big a shock to people alive today, imagine to someone with no concept of computers, electronics, and such.

It would seem like pure magic 🧙‍♂️🔮

Thinking about it a bit more… I guess most tech is like magic to most people.

They may know the words, but there's not much understanding behind that facade (How do they do that? Oh, it’s computers! It's CGI! It's AI! It's controlled by software! *hands waving*)...

But even without much understanding, people are still familiar with all of it and they know that someone somewhere knows how it really works which is enough to make it “not a real mystery” and so it rapidly loses its ability to impress.

As we create and “grow” more AI models that do things in ways that we don’t understand, we’ll get to a point where no one anywhere truly understands what is going on fully, and won’t that be closer to magic?

Even today, we have shallow explanations like “it predicts the next word” or “it does diffusion from random noise until patterns that it recognizes as being part of the image it’s trying to create emerge”, or whatever, but that’s not true understanding.

If you look under the hood at the giant matrices of numbers, they’re not something that humans can hold in their heads and *truly* understand.

---

🏦 💰 Liberty Capital 💳 💴

My Thoughts on ‘Internal Diversification’/’Look-Through Diversification’

A concept that is not mentioned enough when it comes to portfolio construction and evaluating risk is the internal diversification of businesses. Many people will ask you “how diversified” you are, but they often stop at the level of “how many stocks do you own?”

I like to think of my portfolio as a conglomerate; how many businesses do I actually own?

Someone with a 5 stocks portfolio may actually be more diversified than someone with a 20 stocks portfolio.

They may own 5 companies that each have dozens or even hundreds of business units in various industries/niches, in various geographies, with multiple product lines, while the person with 20 stocks may own mostly single-product companies that cluster in just a couple of industries and/or countries.

Think of it like the root systems of plants and trees. Some may have a single root that is fairly shallow, while others may sprawl in all directions over great distances and be much harder to pull out. They may even survive large sections being cut off.

One is not necessarily better than the other, and even highly internally diversified companies you can have risks that aren’t diversified by all the sub-businesses (ie. if the CEO is horrible at capital allocation or a crook). But it’s something you should still take into account to get a more accurate picture of what you actually own.

Of course, there are other important aspects when it comes to evaluating diversification, but this is the one that I wanted to focus on today.

This is an updated version of something I wrote in the Eoarchean era of Edition #22. I’m making the bet that you weren’t around back then and haven’t seen it before, but if you have, kudos to you, that’s OG! 🎩

🏦🧀 Credit Suisse is Zeno's Bank. 📉🏁