398: GPT-4 is Here, X-Ray of Google, Buffett & Banks, Reddit, Kevin Kelly, Andrew Huberman, Product Placement, and Beatallica

"This digital tool was more or less grown"

Not only may questions remain unanswered; all the right questions may not even have been asked. —Seth Klarman

🎞️🏦 Ok, I gotta pull this clip from 'Sneakers' (1992) out of the archives and give it its moment to shine:

Cosmo: In prison I learned that everything in this world, including money, operates not on reality...

Bishop: ...but the perception of reality.

Cosmo: Posit: people think a bank might be financially shaky.

Bishop: Consequence: people start to withdraw their money.

Cosmo: Result: pretty soon it is financially shaky.

Bishop: Conclusion: you can make banks fail.

Cosmo: Bzzt. I've already done that. Maybe you've heard about a few? Think bigger.

Bishop: Stock market?

Cosmo: Yes.

Bishop: Currency market?

Cosmo: Yes.

Bishop: Commodities market?

Cosmo: Yes.

Bishop: Small countries?

Cosmo: With luck, I might even be able to crash the whole damned system. Destroy all records of ownership.

You can watch the full scene here (the whole film is worth seeing, I’ve probably seen it 5 times over the years).

🤖 🎸🇬🇧🍏🤘🇺🇸 My favorite novelty-band-that-is-actually-good is Beatallica.

As the name implies, it’s a mashup of the Beatles and Metallica, with the singer doing a great impression of James Hetfield, down to the mannerisms and verbal tics.

The song themselves are great hybrids musically and lyrically.

Here’s one of their EPs on Youtube, you can click around the timeline to hear various songs (warning: there is some swearing):

The angle I’m wondering about is: how long until AI can do this? 🤔

One of the strengths of LLMs and generative AI is the ability to extract the patterns that make a ‘style’ recognizable, and mashing multiple styles together.

We’ve seen a lot of this visually. When will we see it musically?

Brace yourself for a wave of novelty hybrid virtual bands!

💚 🥃 🐇 This is the free edition of Liberty’s Highlights with 15,800+ subscribers.

You can get 2 extra editions/week full of juicy stuff + access to the private Discord 🗣🗣🗣 community by becoming a paid supporter (it’s quick & easy).

Paid post since last week 🐇:

• 🔒 396: Meta’s LLaMA, Baumol Effect, YouTube, Building AI Analysts, Doug Leone, Virtual Kidnapping, Visceral Fat, and Last of Us

• 🔒 397: Silicon Valley Bank, Crowdstrike Q4, Stack Fallacy, Google PaLM-E Robot, DuckDuckGo's Wikipedia LLM, Arena FPS, and Silo

If you click the links above, you can see the intros for free and there’s a link to get a 7-day free trial of the paid version. What have you got to lose?

---

🏦 💰 Liberty Capital 💳 💴

🤖 GPT-4 is here — what we know so far 🤖

Above is the graph that you’re going to see everywhere for a few days.

In blue are GPT-3.5’s exam results for all kinds of tests (ranked against human test-takers) and in green, GPT-4’s ranking on the same tests. (note: “[OpenAI] did no specific training for these exams. A minority of the problems in the exams were seen by the model during training”).

GPT-4 is a large multimodal model (accepting image and text inputs, emitting text outputs) that, while less capable than humans in many real-world scenarios, exhibits human-level performance on various professional and academic benchmarks.

For example, it passes a simulated bar exam with a score around the top 10% of test takers; in contrast, GPT-3.5’s score was around the bottom 10%.

GPT-3.5 isn’t exactly ancient, so the rate of progress is impressive. Depending on what you count as the launch, it came out in March or November 2022, at most a year ago.

What about those pesky ‘hallucinations’? 😵‍💫

While still a real issue, GPT-4 significantly reduces hallucinations relative to previous models (which have themselves been improving with each iteration). GPT-4 scores 40% higher than our latest GPT-3.5 on our internal adversarial factuality evaluations

On the vision side: In the white paper, GPT-4 explains Reddit meme jokes, which show a very subtle understanding of visual data. For example, when shown this and asked what is funny about it:

GPT-4 says:

The image shows a package for a "Lightning Cable" adapter with three panels.

Panel 1: A smartphone with a VGA connector (a large, blue, 15-pin connector typically used for computer monitors) plugged into its charging port.

Panel 2: The package for the "Lightning Cable" adapter with a picture of a VGA connector on it.

Panel 3: A close-up of the VGA connector with a small Lightning connector (used for charging iPhones and other Apple devices) at the end.

The humor in this image comes from the absurdity of plugging a large, outdated VGA connector into a small, modern smartphone charging port.

This is a combination of excellent vision and some fairly sophisticated reasoning about a joke that is conveyed almost entirely through visuals. Understanding it implies a visual understanding of the logical connections between phones and ports and new and old charging cables.

GPT-3.5 didn’t even have visual capabilities! Things are moving really fast.

Speaking of vision, here’s a passage from the white paper that really struck me:

We preview GPT-4’s performance by evaluating it on a narrow suite of standard academic vision benchmarks. However, these numbers do not fully represent the extent of its capabilities as we are constantly discovering new and exciting tasks that the model is able to tackle.

Basically, the model is a giant inscrutable blob of numbers in matrices, and what it can do has to be discovered by trial & error because it emerges from the intelligence embedded in the pattern-matching model. 🤯

This is not a human artifact that was designed and built in the way that most of our tools are built. This digital tool was more or less grown by training it on a gigantic corpus of data (text and images) and it extracts the patterns — and thus the meaning, to some extent — from that. It’s almost like a pseudo-biological process that takes place digitally.

If you want to try it, it’ll first be available to ChatGPT Plug paid subs, but with a usage cap at first.

More here: GPT-4 white paper.

🔭 Great interview with Kevin Kelly 🔮

A few highlights below, but I recommend the whole thing!

First, on the technium:

I call our human made system of all technologies working together, the technium. Each technology can not stand alone. It takes a saw to make a hammer and it takes a hammer to make a saw. And it takes both tools to make a computer, and in today’s factory it takes a computer to make saws and hammers.

This co-dependency creates an ecosystem of highly interdependent technologies that support each other. The higher the technologies the more intertwined, complex, and codependent they become. At this point in our evolution we need farmers to support indoor plumbing and plumbing to support banks, and banks to enable farmers, and round and round 

On demographics:

By their nature demographic changes unroll slowly because they run on generational time. Inspecting the demographic momentum today it is very clear human populations are headed for a reversal on the global scale by the next generation. After a peak population around 2070, the total human population on this planet will start to diminish each year.

So far, nothing we have tried has reversed this decline locally. Individual countries can mask this global decline by stealing residents from each other via immigration, but the global total matters for our global economy.

[…] We will have to perfect a system that can keep improving and getting better with fewer customers each year, smaller markets and audiences, and fewer workers. That is a huge shift from the past few centuries where every year there has been more of everything. 

I still think the way to mitigate lower birth rates is by focusing a lot more investment on expanding healthspans (ie. reducing death rates).

Here’s some career advice:

My generic career advice for young people is that if at all possible, you should aim to work on something that no one has a word for.

Spend your energies where we don’t have a name for what you are doing, where it takes a while to explain to your mother what it is you do. When you are ahead of language, that means you are in a spot where it is more likely you are working on things that only you can do. It also means you won’t have much competition.

It reminds me a bit of supporter (💚 🥃) Josh Wolfe’s “Whenever parents say 'It will rot your brain.’ It ALWAYS presages the next $10 billion+ industry...”

📺 Ad Breaks vs Product Placement (regulation about advertising around the world) 🇰🇷 🍔

Fascinating:

In South Korea, TV broadcasts aren’t allowed to have advertising breaks. The way brands reach TV-viewing audiences is by product placement.

Somehow, I didn’t know that.

I wonder how many other countries have this regulation, and how it makes their popular culture evolve differently than in countries like the US 🤔

In some ways, product placement can be less intrusive than stopping the whole show for an ad, but in others, maybe it’s a little bit less… honest?

At least when the show stops for an ad, you know what it is. But when a character interacts with any kind of brand or business, one can never be quite sure if this is what the writers wanted, or if this was added because money changed hands.

Also, product placement may work for certain things types of content, but for more “elevated art”, I think it can ruin the experience. The example that came to my mind was watching ‘The Godfather’ and they all go eat at Burger King and chat about their new Hoover vacuums for no good reason. Ugh.

Heavy-handed product placement usually doesn’t age very well and may make things less ‘timeless’ too…

People on Twitter have been telling me about various countries where all the ads are bunched together, or in between the shows…

I suspect that each format creates incentives for different types of ads and content — which is a good reminder that a lot of things that we think are “just the way things are” and normal are in fact created by arbitrary incentives that may have been set long ago, and there’s a lot of path dependency to all this.

Check out friend-of-the-show and supporter (💚 🥃) Trung Phan’s whole post about grit vs quit.

🩻🕵️‍♂️ Google Deep Dive Update 🔦💻🔎

It’s not technically a deep dive, but it kind of is because it certainly dives pretty deep into what makes Google tick, though it focuses mostly on the core search/ad + YouTube + Google Cloud/GCP and doesn’t try to divine the value of every single moonshot and side thing they do.

Let’s call it a Google company x-ray…

It’s a great read, very clearly structured and explained, I recommend it:

• 📄 Alphabet Update on MBI-Deep Dives

Here’s a highlight:

Google likely spent ~$30 Bn fixed costs in 2022 (this is per year, not one-off costs) in running their search, maps, and gmail platforms (likely most of it is spent on Google search). Many investors may need to update their priors that these are capital light businesses

Also, MBI is now on Substack too, so go subscribe (even if just to the free posts at first — they’re good too and will get you hooked! Check out this recent one: “Meta Needs A Decade Of Efficiency”)

🏦 🔃 Upward Reflexivity, Buffett & Banks Edition

Wisdom from friend-of-the-show Jerry Capital (💎🐕):

Buffett owning a bank is not the same as you owning a bank.

If Buffett owns a bank, it can't have a bank run.

Interview: Reddit’s new Chief Product Officer on the future of Reddit 📲 👽

Reddit’s a bit weird and often gets overlooked when social media is discussed, but it’s actually really big, really useful, and shapes a lot of the online conversation (kind of like Twitter — punching above its weight vs the real giants at Facebook and Instagram).

• 🎧 How Reddit is getting simpler — and dealing with TikTok, with chief product officer Pali Bhat

The most powerful line in the interview was “the conversation is the content”. It truly is a differentiation.

There’s pretty much nowhere else where I’ll quickly get through the content at the top and then spend the majority of the time in the comments. Twitter comes closest, but it’s a different animal.

Pali Bhat also talks about how they’re trying to make the site more accessible to new users while retaining the depth that existing users love, the shift to more video creation and discovery on the platform, improving their search (instead of having people append “reddit” to their Google searches).

My takeaway is that it’s great they’re trying to evolve the site, but I hope that they’re careful and don’t accidentally ruin the community. Because that’s where the real value is, and online communities require a very specific pH to thrive — change things around too much, and you risk killing the ecosystem.

🇪🇺 Energy Crisis = Europe is deindustrializing 🏭🚫

The reaper has come for another European aluminum smelter. [...]

“While power prices have retreated sharply from last year’s peaks, Speira Gmbh will shut its Rheinwork plant in Germany this year due to challenges in the energy market, the company said Thursday. That follows a 50% cut in aluminum production announced in September as soaring power and gas prices plunged Europe’s energy-intensive metals industry into an existential crisis,” [...]

European production capacity for aluminum has decreased by over half since the outset of the energy crisis.

These types of heavy industry plants very rarely restart once they shut down for a while. It’s not just aluminum, but also chemical plants, fertilizer plants, etc.

At a time when much of the world seems to want to be less dependent on long global supply chains (and especially on China), Europe is going in the other direction.

---

🧪🔬 Liberty Labs 🧬 🔭

Image made with StableDiffusion 💚 🥃

Scientific Poetry! ✨💥🌍🥇 (Do you have a gold ring?)

Great stuff by Mark Heyer (👈 go subscribe to his Substack, it’ll expand your mind).

I tried to condense it to a shorter highlight, but the whole thing is worth reading:

We know that all the heavy elements in the universe are created exclusively in the massive explosions of very large stars, called supernovas. So rare are these that an average galaxy of 100 billion stars sees only one supernova about every hundred years. For a few days, the supernova may outshine the entire galaxy as it sprays atoms of all the elements on the periodic table into interstellar space. [...]

The nuclei of the forming stars pull in more and more dust, harvesting and concentrating angular momentum, creating spinning eddies around the central star–to–be. These spinning clouds condense into planets, not randomly but following the Constructal Law of physics, which governs the evolution of all matter and energy flows. [...]

One might think that gold atoms would be diffused throughout the mass of the earth and its rocks, virtually undetectable in their rarity. [...]

But in reality, the earth is a giant, evolving refinery, churning, folding, scraping, compressing, and melting, creating trillions of Constructal flows, each creating more access to flows of its own kind until much of the gold (and many other elements) on the earth becomes concentrated in a few specific locations.

The fact that I exist to wear a gold wedding band is testimony to billions of years of evolution of the inanimate physical universe and billions of years of biological evolution

Carl Sagan and Richard P. Feynman would be proud of this imagery!

👀🧠🫀🫁💪 Conversation/Interview: Andrew Huberman & Tim Ferriss 🗣️🗣️🎙️

So Huberman has become a pretty huge name in recent times. I’m sure that’s causing some backlash somewhere, because it always happens whenever someone becomes popular — they stop being judged on merit and become an opportunity to social signal by hipsters or people trying to raise their own status by dunking on someone who’s high-status… (because nobody is perfect, so there’s always something to complain about)

Anyway. All that to say, I don’t care, I like his stuff, I learn a lot when I listen to him and I tend to trust his judgment, as I do with Peter Attia and Rhonda Patrick.

• 🎧 Dr. Andrew Huberman — The Foundations of Physical and Mental Performance, Core Supplements, Sexual Health and Fertility, Sleep Optimization, Psychedelics, and More (#660)

It may be nerdier and go into more detail than you want, but I feel that’s that best way to learn.

You may not remember every detail of his weekly workout schedule, but you get a much better gestalt of how he thinks about things than if he had just given you the 2-minute outline.

📸🌖🤖 Samsung caught faking moon photos on phone camera’s “100x Space zoom”

When do you cross the line between computational photography that uses AI to enhance photos and completely fabricating photos for marketing bragging rights? (“100x Space Zoom!!! iPhone camera is so much worse because it can’t do this!!!!”)

Samsung was caught faking moon photos by a clever Reddit user who put a blurry, low-resolution photo of the moon on his computer screen, zoomed into it with his Samsung phone, and took a photo, which magically turned into a sharp photo of the moon.

You can see a video of what this looks like here, and Inverse Mag also debunked it, even having the Samsung phone put moon texture on a ping-pong ball… 🏓.

---

🎨 🎭 Liberty Studio 👩‍🎨 🎥

‘The Pantheon, Hagia Sophia, Florence Cathedral, and St Peter's Basilica’

‘ all at the same scale in both plan and elevation/section.’ Source.