548: Buffett's Tangled Web, Amazon Retail, Musk vs OpenAI, YouTube, Demis Hassabis on DeepSeek, Nuclear AI, Tiger Eyes, Cosmological Evolution, and Bob Dylan

"The walls you've built might be made of paper."

I live in my dreams — that's what you sense.

Other people live in dreams, but not in their own.

That's the difference.

–Herman Hesse

🧙‍♂️🪄📚 Great books combine telepathy and time travel, dreaming and awakening.

They compress decades of life experience, thinking, and research into something that fits in your hands and can be downloaded into your brain in a few hours.

Photons bouncing off the pages conjure images and words on the inner screen of your consciousness.

Sometimes they alter your mood for a moment, sometimes they change your life forever.

This is still one of humanity’s coolest technologies, even after all this time.

💔❤️‍🩹🥚🍳🏋️‍♂️🏗️ If you think you’re fragile, you are.

That belief becomes a self-fulfilling prophecy.

Thinking you're fragile makes you fragile.

I’m not saying you’re invincible, but you’re probably stronger and more resilient than you think — especially if you’ve been taught that you’re fragile!

To find our limits, we must go near the edge, do difficult things, and fail a bunch of times. If you never fail, you’re not doing enough hard things.

You might believe your redline is at 37, when it actually lies at 77 (made up units to illustrate my point).

There could be incredible achievements waiting for you in the 50-60 range, but if you've convinced yourself that 37 is your ceiling, you'll never explore that territory.

Challenge your assumptions about your own strength. Don't let untested assumptions about your capabilities hold you back.

The walls you've built might be made of paper 📑

✨😊✨ Life’s simple delights:

A tasty hot sauce 🌶️ , a good espresso with rich crema ☕️ , freshly grated parmesan cheese 🧀 , a good book 📖, a hug from a loved one 🤗 , a great song 🎶, the sun on your skin ☀️, making someone genuinely laugh 😄, a walk in the woods 🌳 🚶‍♂️ 🌳, fresh sheets on the bed 🛏️

What else?

📺🤠👀 My favorite TV show of all time is probably Deadwood.

Here’s a simple flow chart for you:

1. If you’ve never seen it, watch the first three episodes. That should be enough to know if it’s for you. It’s not for everyone, but if it’s for you, it’s probably *really* for you, as there’s nothing else quite like it.

2. If you’ve seen it and would enjoy a discussion of the show, its writing, themes, language, characters, etc, listen to the podcast I did about it with David Senra (📚🎙️): 🎧 David Senra on Deadwood (also: founders, entrepreneurs, civilization, Henry Singleton, Walt Disney, David Ogilvy, etc)

Yes, I’ll keep trying to convert people to Deadwood until I end up with Wu’s pigs 🐖

¯\_(ツ)_/¯

💚 🥃 🙏☺️ If you’re a free sub, I hope you’ll decide to become a paid supporter in 2025:

---

🏦 💰 Liberty Capital 💳 💴

🕸️ Buffett and Munger’s Tangled Web of Cross-Ownership

A recent tweet by friend-of-the-show Trung Phan (check out our podcast together) about OpenAI’s complex ownership structure reminded me of this flowchart from Alice Schroeder's The Snowball about Buffett and Munger’s complex Matryoshka doll ownership of various companies back in the Blue Chip Stamps and See’s Candies era.

This complexity led to an SEC lawsuit in 1976: SEC v. Berkshire Hathaway Inc., et al.

They alleged that Buffett, Munger, Berkshire Hathaway, Blue Chip Stamps, and Wesco acted as a "group" without properly disclosing their combined ownership and intentions, “violating Section 13(d) of the Securities Exchange Act of 1934”, which is the part that requires investors to disclose when they own more than 5% of a company's stock.

The SEC's concern was that this complex web of ownership and cross-ownership, particularly with the involvement of Buffett and Munger in multiple entities, obscured the true control and intentions behind the See's Candies acquisition.

The aftermath? After settling the case with a consent decree, this web of holdings was simplified, partly by having Berkshire acquire Blue Chips in 1983. Wesco was also eventually fully acquired in 2011.

🛒 Amazon… Retail 🚚📦📦📦

While AWS is universally acknowledged as a great business….

37% operating margins in the past 12 months, etc

…I’ve heard many investors advocate for Amazon to spin it off — or maybe even that the government would break them up — allowing them to just own the “good” asset without the low-margin retail attached.

My friend MBI (🇧🇩🇺🇸) wrote a great overview of the most recent Q, and made this observation that I tend to agree with:

Amazon ex-AWS

Despite North America segment being more than 2.5x the size of International segment, it grew by 10% whereas international revenue increased by 9% (both FXN).

For the 8th consecutive quarters, both North America and international segment experienced YoY operating margin expansion. Amazon reported its highest operating margin in North America segment at least since 2013. Moreover, please note that they currently expense majority of their costs associated with development of the satellite network which will be capitalized when services reach commercial viability. Therefore, the “actual” retail margin is almost certainly even higher.

While most investors are usually more excited by AWS’ prospects and infatuated by their lofty margins, I may be in the minority in being more optimistic about Amazon retail’s long-term profitability. AWS has a couple of pretty capable competitors and thanks to Nvidia being the key bottleneck, the industry value chain may not evolve in a favorable way for AWS. But when I think about Amazon retail’s long-term future, the gap between Amazon and the competitors may keep growing. [...]

Amazon continues to indicate that they are far from done in optimizing their logistics footprint as well as improving speed of delivery

I agree with him that retail keeps getting underestimated. The reasons are a bit similar to why Costco was misunderstood for so long (though not anymore 📈🚀).

Investors often have a simplistic view of what makes a good business, focusing on margin and forgetting other factors, like duration, competitive dynamics, and underlying ROIC when you try to take out a lot of the noise or discretionary bets vs actual maintenance spend.

Not to mention that Amazon’s advertising business is largely based around the retail business, and is a high-margin and fast-growing way to monetize it.

While AWS is a great business and everyone agrees, the retail business is also great but there’s a lot more divergence of opinion about it, which probably means that it’s where there could be positive surprises and upside. To be clear, I’m not saying Amazon stock is currently mispriced, just that over time, it’s in retail where there’s most likely to be positive surprises.

‘Elon Musk-Led Group Makes $97.4 Billion Bid for Control of OpenAI’ 💰💰💰💰💰💰🫳🤖💸

More drama, always more drama…

A consortium of investors led by Elon Musk is offering $97.4 billion to buy the nonprofit that controls OpenAI, upping the stakes in his battle with Sam Altman over the company behind ChatGPT.

“It’s time for OpenAI to return to the open-source, safety-focused force for good it once was,” Musk said in a statement provided by [Musk’s lawyer] Toberoff. “We will make sure that happens.”

It didn’t take long for Altman to respond on Twitter: “no thank you but we will buy twitter for $9.74 billion if you want” 🐦

There’s clearly no love lost between these two…

The bid is being backed by Musk’s own artificial intelligence company xAI, which could merge with OpenAI following a deal.

What is confusing to some about this is OpenAI’s weird structure and the planned for-profit conversion of the subsidiary. In short:

1. OpenAI Structure: OpenAI has a unique structure. It was originally established as a non-profit (OpenAI Inc.). Later, a for-profit subsidiary (OpenAI LP) was created, but it is capped-profit, and it is controlled by the non-profit board.

2. Musk's Bid: Elon Musk's $97.4 billion bid is specifically for the non-profit entity (OpenAI Inc.). This non-profit entity is the controlling shareholder of the for-profit OpenAI LP.

3. Control, Not Direct Ownership: By acquiring the non-profit, Musk's group would gain control over the entire OpenAI structure, including the for-profit operating business (OpenAI LP). It's not a direct purchase of the for-profit entity's assets or operations, but rather an acquisition of the entity that governs it.

That’s my understanding, anyway — this thing is confusing!

Here’s how things are being valued:

OpenAI’s nonprofit board and Microsoft, a large shareholder in the for-profit arm, are negotiating what the size of their respective stakes will be after the proposed conversion. The nonprofit is supposed to get at least 25%, we reported. But even if it got, say, 45%, that’s worth only $67 billion based on the most recent implied valuation of OpenAI ($150 billion in a financing that closed last fall)—$30 billion below Musk’s offer.

That said, if the most recent talks valuing the OpenAI business at $260 billion come true, then a 45% stake for the nonprofit would be worth nearly $120 billion. (As a side note, Musk’s team estimates that the nonprofit’s assets—control over the for-profit and rights to some of its future profits—is essentially equivalent to a 51% stake in the for-profit, meaning Musk is valuing the for-profit arm at about $190 billion.)

What seems to be happening is that Musk sees two potentially positive outcomes from this bid (for him):

Either he gains control of OpenAI, or he accomplishes several strategic goals: he slows them down, increases scrutiny (especially by the attorneys general of California, where OpenAI is headquartered, and Delaware, where it is incorporated ⚖️), and forcing the for-profit conversion to be more expensive.

Was Altman intending to lowball the value of the for-profit subsidiary when it is ‘purchased’ from the non-profit controlling entity? If so, the price likely just went up, potentially by a large amount.

📺 YouTube is the new TV 📺

From friend-of-the-show Eric Seufert (“Everything is an ad network”):

TV has surpassed mobile and is now the primary device for YouTube viewing in the U.S. (by watch time), and according to Nielsen, YouTube has been #1 in streaming watch time in the U.S. for two years […]

YouTube is larger by total revenue than Netflix, free of the risk of content production bets, and generating the majority of its revenue from advertising that is mostly immune from consumer subscription fatigue. What’s more, Google can apply the machine learning infrastructure it has developed over decades to optimize ad targeting and engagement with the benefit that the television surface area and context implies: more distilled user attention and more reliable device identifiers for measurement.

Source. (🔐) h/t MBI (🇧🇩🇺🇸)

🇫🇷⚛️🤖 France Leverages Nuclear Energy for AI Arms Race

Good to hear about Europe actively trying to compete, building things, and using its nuclear power rather than shutting it down (for a change):

France is making a bid to catch up in the artificial intelligence race by leaning on one of its strengths: plentiful nuclear power.

The French government plans Monday to pledge a gigawatt of nuclear power for a new artificial-intelligence computing project expected to cost tens of billions of dollars, according to its private-sector backers and the French government. [...]

The nuclear project, which aims to have a first tranche of 250 megawatts of power hooked up to AI-computing chips by the end of 2026, rivals the Stargate project in the U.S., backed by SoftBank and OpenAI. Stargate is starting with a campus in Texas initially fed by 200 megawatts of power, with plans to expand to 1.2 gigawatts.

Macron couldn’t help being cheesy about it: 🧀

“I have a good friend on the other part of the ocean, saying, ‘Drill, baby, drill,’” Macron said in a speech at the summit, referring to President Trump’s efforts to expand American oil drilling. “Here, there is no need to drill. It’s just plug, baby, plug.” (Source)

😬

What say you now, Germany? 🇩🇪

Your move! ♟️

🤖🩻🔍🧠 Demis Hassabis of DeepMind on DeepSeek’s Efficiency

While he’s a competitor, so take what he says with a grain of salt, Hassabis knows what he’s talking about and appears unimpressed by DeepSeek’s breakthroughs.

He said that DeepSeek “seems to have only reported the cost of the final training round, which is a fraction of the total cost… We don’t see any new silver bullet technologies, DeepSeek is not an outlier on the efficiency curve. Gemini is more efficient than DeepSeek in terms of its training to performance or cost to performance. We just don’t talk about it very much.” 🔥

He also said that DeepSeek seems to “have relied on some Western models to distill from.”

My personal experience with Gemini 2.0 Pro Experimental and sometimes Gemini 2.0 Flash (in Perplexity) has significantly improved my assessment of Google's models. I used to barely think about them, but the 2.0 line is very strong AND very fast. Gemini Pro’s 2 million token context window is currently unrivaled.

---

🧪🔬 Liberty Labs 🧬 🔭

🐅 👁️👁️ Do you know Why Tigers and Cats Have Different Pupil Shapes 🐈

I don’t know if you’ve ever noticed, but house cats have vertical slit pupils while tigers have round pupils, more like you and me.

At first glance, it’s a bit weird, right? They’re all felines. If these slit-eyes are good for cats, why aren’t they good for tigers? What created the selective pressure to change that? Did they all start out with similar eyes and then branch out at some point?

I was curious why, so I did a bit of research:

• Tigers (and other big cats like lions, leopards, and jaguars) have round pupils. This is primarily linked to their hunting style and activity patterns. That’s because they are mostly ambush predators, and while they can hunt at any time, they are not strictly nocturnal. Round pupils provide a good balance of light intake and depth perception for hunting in a variety of light conditions.

• House cats (and many other smaller cat species) have vertical slit pupils. This is an adaptation for nocturnal hunting. Vertical slits can close more tightly than round pupils, protecting the sensitive retina in bright light. They also open very wide in low light, maximizing light gathering. The vertical slit also enhances depth perception for objects in the horizontal plane, which is crucial for judging the distance to prey that is often moving horizontally.

Cool, isn’t it? So which came first?

• It's highly probable that the ancestral condition for felids (the cat family) was round pupils. This is based on several lines of evidence: Outgroup Comparison: Looking at closely related carnivore families (like viverrids - civets, genets; and euplerids - Malagasy carnivores), round pupils are more common. This suggests the common ancestor of these groups, and thus early felids, likely had round pupils. This suggests the common ancestor of these groups likely had round pupils. Ecological Niche: The earliest felids were likely more generalized predators, not specialized nocturnal ambush hunters like many small cats today. Round pupils would have been suitable for this lifestyle. 👁️

• Vertical slit pupils evolved multiple times independently within the felid family. This is a classic example of convergent evolution, where similar traits evolve in unrelated lineages due to similar selective pressures. In this case, the selective pressure was the adoption of a nocturnal or crepuscular (dawn/dusk) ambush hunting strategy.

Even with small cats, there’s some in-between: the sand cat and the black-footed cat, have pupils that are more intermediate or rounded.

But that’s not all! Feline eyes also have other cool aspects:

• Tapetum Lucidum: Both tigers and house cats possess a tapetum lucidum, a reflective layer behind the retina. This is what causes the characteristic "eyeshine" in many nocturnal animals. However, the effectiveness and color of the eyeshine can vary. House Cats generally have a very strong tapetum lucidum, contributing to their excellent night vision. The eyeshine is often green or yellow. Tigers also have a tapetum lucidum, but it might be slightly less developed as they are not exclusively nocturnal. Their eyeshine can be a range of colors, including green, blue, or even reddish.

• Third Eyelid (Nictitating Membrane): Cats have a "third eyelid," a translucent membrane that moves diagonally across the eye from the inner corner. It provides extra protection and helps keep the eye moist. You might see it when a cat is sick or sleepy.

• Cats are somewhat farsighted. They can't focus well on objects very close to their face (less than about 1 foot). They rely on their whiskers and sense of smell for close-up investigation.

I grew up with cats in the house — I wish I had known all this at the time. I could’ve run a few amateur feline vision experiments 🤔

OpenAI Taps Google’s Former TPU Chip Designer to Build Its Custom AI-Training Chips 🐜🛠️👨‍🔬🤖

OpenAI plans to submit its custom-designed AI chip to TSMC in the coming months for refinement and design finalization, aiming for mass production using 3nm process technology in 2026. (Source)

The development of OpenAI's custom AI chip is being led by Richard Ho, a former leader of Google's Tensor Processing Units (TPU) chip program, in collaboration with Broadcom (which was also a partner for Google on the TPU. OpenAI now has a team of about 40 engineers working on this (which is relatively small because they can use a lot of IP blocks from Broadcom and other suppliers).

Richard Ho spent 9 years at Google and was Senior Director of Engineering. He was involved in co-inventing methods using machine learning for chip architecture design. Sounds like he’s the ideal person to help OpenAI with this project.

The chip will be training-focused, and I wouldn’t be surprised if its main contribution — at least for the foreseeable future — is to give OpenAI more leverage when negotiating with Nvidia. 😎🤝💰💰💰💰💰💰

It’ll take a while to ramp up production, for the chip and its software to mature, and meanwhile, Nvidia is a fast-moving target with incredible chip design resources and a decade+ buildup of software libraries and frameworks (which can be a bit of a two-edged sword at times, but is mostly a positive for them).

Interview: Julian Gough on Evolved Universes via Black Holes 🌌✨🕳️💥🌌

Cosmological natural selection! What a mind-bending concept.

Julian and Jim discuss about all kinds of interesting things, but the truly 🤯 thing that I encourage you to check out is the discussion of Lee Smolin’s theories (and Julian’s expansion of them, as well as his prediction that could be verified by the James Webb Space Telescope):

• 🎧 Julian Gough — The Egg and The Rock (EP.249)

Cosmological natural selection, as envisioned by Lee Smolin and philosophically expanded upon by Julian Gough, posits that our universe is the result of an evolutionary process spanning a vast multiverse. The theory draws a crucial connection between the Big Bang singularity that initiated our universe and the singularities found within black holes. 

It suggests these singularities might be two ends of the same phenomenon: a black hole's singularity in one universe could be the Big Bang of a new, "daughter" universe. 

Crucially, the fundamental laws of physics would be passed on from the "parent" universe to the daughter universe, but with the possibility of slight variations in the fundamental constants. This "inheritance with modification" is analogous to biological evolution, driving a selection process where universes that are more prolific at producing black holes (and thus, new universes) become more common, leading to a multiverse dominated by universes finely-tuned for black hole formation. 

Gough's interpretation further speculates that the emergence of intelligent life capable of harnessing micro-black holes for energy generation could be a significant factor. Such civilizations might intentionally create black holes, thereby increasing their universe's "reproductive fitness" within the multiverse, adding another layer to the cosmic evolutionary process.

I had a call with Julian a few months ago and we discussed this, and I’ve been thinking about it on and off ever since.

While the evidence remains to be seen, the theory's elegance is compelling. It offers stronger explanatory power than most other cosmological origin theories in demonstrating how complexity arose from simplicity (after all, this is precisely what evolutionary processes achieve in biology!).

---

🎨 🎭 Liberty Studio 👩‍🎨 🎥

🎥 ⚔️ Breakdown of Hector vs Achilles’s Epic Duel 🛡️

Here's an excellent analysis of this combat scene, exploring both why it works so well and how the filmmakers crammed so much subtext into it.

🎬 Bob Dylan: Film 🔂 & Album Sales 💿 💿 💿📊

Two quick ones:

I went to see ‘A Complete Unknown’ (2024, Timothée Chalamet) in the theater for the second time. I shared my review here in Edition #544.

My wife had not seen it, and she also enjoyed it, though not as much as I did (but she’s not as familiar with Dylan’s catalogue as I am). Having just read his memoirs on vacation, it helped me see new dimensions in the film. It held up to a second viewing, and I’m sure I’ll watch it again in a few years.

I was curious about Dylan’s album sales. I know how influential he is, and that he had *some* commercial success, but he never really was a hit single kind of guy and he did a lot to stay of the public’s eye for a long time, so I figured that maybe he had sold 5-10 million albums over time.

I even thought that maybe his son Jakob had sold more albums with the Wallflowers because they had such massive singles with Bringing Down the Horse in 1996 (I checked: They sold about 15 million albums cumulatively)

So I looked it up…

Bob Dylan has sold over 125 million albums worldwide 😲🌎

Slow and steady wins the race! 🐢

The breadth and depth of his catalogue is incredible:

• His discography consists of 40 studio albums, 16 live albums, 17 volumes of The Bootleg Series, 26 compilation albums, and he has contributed to 7 soundtracks.