506: Jensen's Worry, Apple AI + Google TPUs, Big Tech, S&P 500 as a VC Fund, Passive Investing, YouTube, Antibiotics Resistance, and Contact

"going clickety-clack on the keyboard"

Liberty

Jun 19, 2024

I don't believe anything, but I have many suspicions.

–Robert Anton Wilson

☕️ I ordered my first espresso machine.

A new thing to learn about!

I mostly drink green tea (loose-leaf Japanese sencha), but I’ve long enjoyed a bold espresso. For some reason, I never pulled the trigger on a home setup… until now.

The Breville Bambino is apparently a decent entry-level machine. It was on sale and I couldn’t resist (20% off on Amazon — my value investor side comes out whenever something that I want anyway has a decent price drop).

I’m sure some Espressophiles will look at this and say it’s garbage compared to their Aviatore Veloce TurboJet which costs about as much as a Tesla, but that’s fine.

Gotta start somewhere!

Whatever the field, there’s usually a sweet spot where you get the best bang for your buck and beyond which you enter diminishing returns territory. I’m not sure if this little Breville is quite on that frontier, but if I like homemade espresso enough, maybe someday I’ll upgrade to the forever one.

We’ll see. I have no idea how deep into this rabbit hole I’ll go ¯\_(ツ)_/¯

🗣️🗣️ A reader drove over 2 hours to have coffee with me (hey J.K. 👋). It was a lot of fun!

After all these years, I’m still impressed by the power of sitting down in front of the computer and going clickety-clack on the keyboard to make things happen in the real world.

💚 🥃 🚢⚓️ If you’re getting value from this newsletter, it would mean the world to me if you become a supporter to help me to keep writing it.

You’ll get access to paid editions, the private Discord, and get invited to the next supporter-only Zoom Q&As with me.

If you only get one good idea per year, it’ll more than pay for itself. If it makes you curious about new things, that’s priceless.

🏦 💰 Liberty Capital 💳 💴

🚀 🐢The S&P 500 Index as a VC Fund

I’m sure you’ve heard VCs talk about their model.

They say that most of the companies they invest in likely won’t do well, but that a few big winners will create a power law that makes the whole fund. They can only lose 1x at most, but their upside is uncapped if they can find big winners. One 1000x win can make up for a lot of 1x losses.

These days, it looks a bit like the S&P500 is acting like a big VC fund.

To be fair, it’s *always* a small subset of big winners that generate most of the return in the market, but it seems particularly stark right now, with just 6 stocks having a combined market cap of close to 30% of the total index.

Note that the graph above is from June 12th. It’s already out of date — Nvidia now has a larger market cap than Apple and Microsoft, making it the biggest company in the world.

Source.

Indexing isn’t necessarily “passive” 📊 🥚🪺🍳🐣

This will probably be very obvious to you, but it isn’t for everyone.

Some of the most active investors only invest in index ETFs.

Even at the level of an individual index, what does it mean to be “passive”? There are inclusion criteria and committees that decide which stocks get added and removed. Some indexes have more mechanical criteria, but others make subjective calls about what fits.

That’s not good or bad. It’s just that passive investing and indexing are often conflated and shouldn’t be. At least not without some nuance.

If you buy 20 individual stocks and largely don’t touch them for 10 years, are you more “active” than another investor who constantly buys and sells from a mix of 10 index ETFs, trying to time the market and change allocation based on macro sentiment or whatever?

Not that "passivity purity" matters. Everyone can do whatever they want with their money. But I often feel like the discourse about this is a bit more tribal than rational.

🍎☁️🤖 Apple leverages Google's Cloud for AI model training: Is it time for their own infrastructure?

In the same way that Apple is trying to abstract away and commoditize AI models by being the user layer on top, it has long been doing the same for hyperscalers: iCloud data is stored across a mix of Apple data centers and AWS and GCP in a way that is transparent to users (they also used Azure in the past, but as far as I can tell, they haven’t been using them recently).

Apple’s own AI models, which will be used for most of the features and services that they announced at WWDC, were apparently trained on Google’s cloud using TPUs:

When it came to training the AI models that power the iPhone maker's new Apple Intelligence, the company asked for additional access to Google's Tensor Processing Units for training. TPUs are chips specifically designed for artificial intelligence, and Google rents these out via its cloud service as an alternative to Nvidia graphics processing units.

Google had some problems but ultimately delivered:

Apple's request created a scramble inside Google in April, when Googlers became aware of technical issues that could've stopped them from delivering what Apple wanted on time.
It was known internally as an "OMG," a Google term for one-off urgent incidents that don't quite warrant a code red. A war room was convened inside Google, according to a person with direct knowledge of the incident.
The team delivered for Apple after a few very long days, the person said. But it was a close call that could've been bad news for Apple, which has earned the nickname Bigfoot among Google Cloud employees because of how much of Google's data centers it uses. (Source)

It’s not clear if Google is Apple’s only partner for AI training or if they also use AWS. Or maybe someone else, but that info just hasn’t leaked yet.

Because Apple *isn’t* trying to train frontier models that require tens of thousands if not hundreds of thousands of GPUs or TPUs linked together in a cluster, it makes perfect sense for them to train models on other people’s clouds.

The question is: Will they decide that they need to be in the frontier model business themselves, or are they happy to let others spend tens of billions of capex and then reap most of the benefits by plugging their models into their ecosystem?

If they decide that they need to compete with OpenAI, Google, Anthropic, Meta, etc, then they will have to build their own training clusters because anyone who builds a cutting-edge cluster intends to use it, not rent it out to competitors.

😎🐜 Jensen is worried Nvidia’s customers will run out of data-centers and electricity ☁️⚡️🏗️👷‍♂️🚧

Huang told colleagues he was worried cloud server providers such as Amazon Web Services and Microsoft, which collectively have been buying about half of Nvidia’s AI server chips in recent quarters, weren’t moving fast enough to set up new data centers and power sources to accommodate the chips they had ordered

Nvidia has apparently considered going into the data-center business itself to self-host DGX cloud:

Last fall, Nvidia even considered leasing its own data centers for DGX Cloud […] Such a move would have cut out the cloud providers entirely. Nvidia also recently hired a senior Meta Platforms executive, Alexis Black Bjorlin, to run the cloud business. It isn’t clear whether Nvidia plans to move ahead with its own data centers for DGX Cloud.

Divide and conquer:

As he tries to build new software businesses, Huang is attempting to maximize the growth of hardware sales while ensuring no single customer gets leverage over Nvidia.
That might explain last week’s announcement that Oracle, a relatively small cloud server provider, would obtain a large number of Nvidia chips and start renting them out to Microsoft and OpenAI by early next year. Microsoft would have preferred to purchase those chips directly, though one person with knowledge of the deal said it might have struggled to find space and power for that many new chips.
Huang also has maintained a special relationship with CoreWeave, another small cloud provider [that] got generous allotments of Nvidia GPUs at the expense of bigger cloud providers. Microsoft also had to rent GPU server capacity from CoreWeave to meet its needs. (Source)

Forcing Microsoft to pay rivals for access to GPUs and no doubt paying a lot more for the privilege than if they owned the GPUs themselves. 😅

It’s a good strategy, but there may not be much Nvidia can do to fix construction and electricity bottlenecks.

Large data centers can’t be erected overnight. Even if they could, it takes a long time — too long! — to build new transmission and generation capacity (though chances are that big tech can just outcompete other customers and still find power — but this will lead to higher electricity prices for everyone).

If the growth rate of compute requirement remains anything like it has been recently, we should hit the construction/electricity wall relatively soon unless something changes to streamline building stuff.

📺🔎🐦 YouTube experiments with crowd-sourced fact-checking, taking cues from Twitter's Community Notes

YouTube is introducing a new experimental feature that will allow viewers to add “Notes” to provide more context and information under videos […]
The Google-owned company says the feature can be used for things like clarifying when a song is meant to be a parody, or letting viewers know when older footage is being portrayed as a current event.

This is probably a good idea.

Savvy internet users probably have a very finely tuned BS detector and will often reflexively check top comments to see what the wisdom of crowds has to say about any piece of media.

Maybe this more prominent annotation will help everyone else get additional info or context on this that could be misleading. In fact, even the most skeptical and experienced of us don’t have a perfect batting average — and it’s worst when we don’t even realize we’ve been fooled — so I’m sure it’ll save everyone’s bacon once in a while.

The main question is the implementation: will they have a voting mechanism like on Twitter where you need some kind of consensus with votes coming from users across the spectrum of opinions? (ie. users that belong to different sub-groups in the networks, and not just people coming from some tribal group)

🧪🔬 Liberty Labs 🧬 🔭

🧫🦠🔬💰Longitude Prize to detect antibiotic-resistant bacteria has been won after a decade

Here’s some context on the historical inspiration for this prize:

The Longitude Prize - with a financial reward nearly 10 times that of a Nobel - was inspired by a contest in 1714.
It was won by John Harrison, who developed a precise clock allowing sailors to pinpoint their position at sea - one of the greatest challenges of the 18th Century.
Fast forward 300 years - and the modern blight of superbugs was the focus.

It’s estimated that more than a million people die from infections that resist antibiotics each year. This is partly because we misuse antibiotics by prescribing them when they don’t help (ie. viral infections) or the wrong ones.

Having a reliable and rapid way to test 1) Is the infection bacterial and 2) which antibiotics work against it would be a game-changer.

This is what the new PA-100 AST System by Sysmex Astrego does:

The prize-winning test, however, reveals within 15 minutes whether the infection is bacterial - and within 45 minutes, which antibiotics work.
It starts with a urine sample, which is squirted into a cartridge containing tiny doses of five different antibiotics and placed inside a machine.
Any bacteria are forced - in single file - into one of thousands of channels inside the cartridge.
The machine then uses microscopy to monitor how any bacteria grow in the presence of the different drugs.
And the antibiotics that lead to little or no growth are the most likely to work.

They use “nanofluidic” technology to separate the bacteria. Very cool stuff.

More detail here.

🪨⛏️ It takes 16 years on average to build a mine (and how many tears? 🥲)

ESG is likely to increase that number in many jurisdictions, though the boom in construction and industrialization may force some changes if we are to meet the increasing demand for electricity/data centers/manufacturing capacity/electric cars/etc.

All these things we want aren’t made of magic. We need those atoms.

Never forget: Speed has value in itself.

A lot of the 20th century’s greatest triumphs took place in a few years. All of WWII’s mobilization happened in less time than it takes to build one mine today. The US and France built dozens and dozens of nuclear power plants in about the time it takes to build one mine.

Mines that shouldn’t be built at all should get denied quickly, but those that can be built should get approved quickly so that the time it takes is mostly the actual construction time, not years and years in paperwork to finally be given a green light anyway.

h/t Otavio Costa

🥶↔🥵 The Land of Heat Pumps 🇸🇪

there are 130m boilers in Europe, and that every year between 7m and 9m of them break down.
Although in chilly Sweden there are 438 heat pumps per 1,000 households, in France, Germany and Britain there are only 157, 50 and 13, respectively (Source)

Anyone who thinks that modern heat pumps can’t handle cold weather should look at Sweden.

Here’s an obvious thing we should do: Every air conditioner should be built with a reverse valve so that it can also pump heat inside buildings (rather than just take heat from inside and pump it outside).

This would add minimal cost, especially if it was done at scale. Most of the price difference between heat pumps and AC units is due to the fact that we make fewer of them and they tend to be larger, not because there’s much mechanical difference between them.

This would be a cheap way to increase energy efficiency in buildings. Even if an AC unit isn’t large enough to provide all of a building’s heating throughout the year, if it can do so for a large portion of the year and the existing furnace only needs to be used on the coldest days, that’s still a big win.

🎨 🎭 Liberty Studio 👩‍🎨 🎥

🎥 🎞️ How they got that shot in ‘Contact’ (1997) 📡📡📡

If you’ve seen the film, you probably know which shot I’m talking about. It’s pretty mind-blowing the first time you see it, and over the years, I’ve seen it highlighted in compilations of cool shots on YouTube and elsewhere.

Here’s how they did it. Revealing the trick does take away some of the magic from it, so if you don’t want to have it spoiled, don’t look at the video. But if you’re curious and want to see behind the curtain, it’s a pretty good example of what can be done with enough ingenuity and preparation.