473: Zuckerberg King of GPUs, Meta's AGI, Reddit IPO, Netflix Sacred Cows, Apple Vision Pro, EV Charging, China Stocks, and Allulose

"a wonder that it even works in the first place"

The past is never dead. It's not even past.

–William Faulkner

🚙💥🙉 Can you imagine if humanity had invented the internal combustion engine, but had not figured out how to make it *much* quieter with mufflers?

Have you heard how much noise even the tiniest car makes when the exhaust system disconnects? Now imagine rush hour traffic!

Also, isn’t it cool that cars basically have the equivalent of firearm suppressors connected to their exhaust systems?

🛀💭👽💰🏚️ I’m worried about the upcoming Reddit IPO.

It’s such a useful site — the oldest trick in the book is to append “reddit” to your Google searches — I hope it doesn’t get destroyed by its enemies using its new weak points as a public company to pressure it into changing too much, ruining the delicate pH balance needed to keep the constellation of communities alive.

It’s already hard enough to get anything close to right — and I’m not saying Reddit is perfect, it has so many self-inflicted wounds that I’m surprised it’s even still alive — however, it’s a scarce resource online, a wonder that it even works in the first place, like Wikipedia, and others have done even worse…

Because of how precarious the balance is, I get worried any time big changes are on the horizon.

🛀💭 🕰️ 💻 🖥️⌨️ Alternate histories, alternate lives.

Thought experiment: Sometimes I wish I could see what my life would have been like if the internet had never existed. Either because I was born earlier, or somehow it just never became a thing.

I wonder how I would be different, since so much of what I know and of my friends have origins online. What would be better and what would be worse about my life?

If given the choice between that life and my current one, with perfect knowledge of both, which would I choose?

I strongly suspect the current, online one, but I can’t be certain ¯\_(ツ)_/¯

I also wish I could fully put myself in the shoes of someone who is 18 years old today and has grown up fully online.

I’m not saying I’d want to switch places, but I’d like to know what it’s like to have never known the pre-internet world. How does it change your formative years to grow up with everyone else online? (I was online pretty young, but at the time it was an uncommon, nerdy thing to do)

💚 🥃 🚢⚓️ If you feel like you’re getting value from this newsletter, it means the world to me that you become a supporter to help me to keep writing it.

If you think that you’re not making a difference, that’s incorrect. Only 2% of readers are supporters (so far — you can help change that), so each one of you makes a big difference.

You’ll get access to paid editions and the private Discord, and if you only get one good idea here per year, it’ll more than pay for itself!

---

🏦 💰 Liberty Capital 💳 💴

A Meta datacenter.

🤖 Zuckerberg Reveals Big Bet on AGI, Crowns Himself King of GPUs 👑 🐜

Here’s the transcript of what Mark Zuckerberg said on Instagram:

Hey everyone. Today, I'm bringing Meta's two AI research efforts closer together to support our long term goals of building general intelligence, open sourcing it responsibly, and making it available and useful for everyone in all of our daily lives. 

It's become clearer that the next generation of services requires building full general intelligence—building the best AI assistants, AIs for creators, AIs for businesses, and more—that means advances in every area of AI. From reasoning to planning to coding to memory and other cognitive abilities. This technology is so important and the opportunities are so great that we should open source and make it as widely available as we responsibly can so that everyone can benefit.

And we're building an absolutely massive amount of infrastructure to support this. By the end of this year, we're going to have around 350,000 NVIDIA H100s, or around 600,000 H100 equivalents of compute, if you include other GPUs. We're currently training Llama 3, and we've got an exciting roadmap of future models that we're going to keep training responsibly and safely too.

People are also going to need new devices for AI, and this brings together AI and Metaverse. Because over time, I think a lot of us are going to talk to AIs frequently throughout the day. And I think a lot of us are going to do that using glasses, because glasses are the ideal form factor for letting an AI see what you see and hear what you hear, so it's always available to help out. Ray-Ban Meta Glasses with MetaAI are already off to a very strong start, and overall across all this stuff, we are just getting started.

350,000 H100s!

I don’t know how much Meta is paying per unit, but I don’t think Nvidia has to do volume discounting these days since they are supply-constrained — so if we estimate that Meta is paying $30k per H100, that’s around $10.5bn in capex!

I guess the period of belt-tightening at Meta may be over…

And if you add it all up and count other GPUs, that’s 600,000 H100-equivalents!

That’s pretty bonkers 🤯

Of course, a lot of that is going for inference — they have a massive user base and many of their products have GPU-accelerated features — but I wonder how big the training cluster for Llama 3 is 🤔

“Since Llama 2 was open-sourced last year, developers have downloaded it more than 100 million times, and created more than 16,000 derivative models”, Meta told Casey Newton.

During an interview with The Verge, Zuckerberg made it clear that part of his goal with this investment — and with making it public at this stage — is to attract more AI talent to Meta:

“I think that’s important to convey because a lot of the best researchers want to work on the more ambitious problems. We’re used to there being pretty intense talent wars. But there are different dynamics here with multiple companies going for the same profile, [and] a lot of VCs and folks throwing money at different projects, making it easy for people to start different things externally.”

If talent is the bottleneck, the next bottleneck after that is compute. By owning a crapload of it, they are no doubt hoping to create a compute gravitational well to attract AI talent. 🌀

Zuck also frames AI as a piece of the puzzle to make the metaverse possible and compelling.

From the “Her” model of talking to an AI on the go by using connected glasses (so the AI can see what you see) to fully immersive VR where environments and NPCs are created using generative AI. It should be able to supercharge the whole stack.

While Llama 2 didn’t focus on coding abilities, Llama 3 will. Not so much because Meta expects to use them directly in a product, but because they help the whole model be smarter:

“One hypothesis was that coding isn’t that important because it’s not like a lot of people are going to ask coding questions in WhatsApp. It turns out that coding is actually really important structurally for having the LLMs be able to understand the rigor and hierarchical structure of knowledge, and just generally have more of an intuitive sense of logic.”

On-demand intelligence in a box is undoubtedly one of humanity’s bigger levers on the world.

Ranking technologies is always hard because you start to see the dependencies in the tech tree 🌳 Without A you don’t get to B or C or Z.

But the current level of AI tech is up there, and even if we never progressed much past GPT-4, it would still be world-changing. But things don’t seem to be slowing down, and further gains are not going to have a linear impact.

Stronger AI (however you want to name it) has the potential to make the world so much better — amplifying people’s capabilities to learn and create — and allow us to make real progress on all kinds of difficult problems that have stumped us for a long time.

It’s a cliché at that point, but similarly to many intractable problems that became solvable when we got the right knowledge and tools, things like cancer, Alzheimer’s, and the diseases of aging in general could see breakthroughs with these more powerful tools (among other things — I’d be happy with just a universal vaccine for the common cold).

Because the upside is so large, I think we have to take seriously anything that could stop us from getting there. Even if low probability, we should work on these problems.

One thing that worries me is the potential dynamic where all these big AI labs are in a race, and the one that moves the fastest may be the one that is the least thoughtful about these problems. (To be clear, I’m not worried about some LLM writing stuff that offends someone because it isn’t politically correct — I don’t give a crap about that. I’m thinking further down the line to when these things are powerful enough that any idiot could ask one to create the sequence of a novel super-virus and then split the RNA/DNA sequence into 32 parts that each look benign on their own and mail-order them from Twist Bioscience…)

📺 Netflix keeps slaying sacred cows 🐮 (Wrestling Edition)

For a long time, it was “no ads” and then they changed their mind and went big on ads.

Now the “we’re not interested in live content” policy is being overturned:

Netflix Inc. has acquired the exclusive rights to Raw as well as other programming from World Wrestling Entertainment, marking the streaming service’s first big move into live events.

Raw will air on Netflix in the US, Canada, Latin America and other international markets beginning in January 2025, after the expiration of the WWE’s domestic deal with Comcast

What’s next?

To be fair, wrestling technically counts as “scripted” content, but it’s kind of a hybrid between a TV show and a basketball game.

My suggestion for the next big change: stop releasing all episodes of a TV show’s season at once and space them out to create multiple mini-events. This will increase the cultural relevance of shows by having people watch them at the same time and be able to discuss them rather than having everyone on a different episode and being afraid of spoilers 🤐

They’re also preparing the ground for more pricing flex. From their shareholder letter yesterday:

As we invest in and improve Netflix, we’ll occasionally ask our members to pay a little extra to reflect those improvements, which in turn helps drive the positive flywheel of additional investment to further improve and grow our service.

Netflix now has 260.8 million paid subscribers. I wonder how many total users they could get if they added a “free” 100% ad-supported tier 🤔

🇺🇸 How cheap does gasoline have to be to be equivalent to the cost of charging an EV? 💸⛽️🚘 🔌🚙🔋

This one is fairly self-explanatory. It takes into account the average electricity price in each state to figure out the break-even price of gasoline vs charging an EV using electricity.

They use the Hyundai Kona EV and Kona ICE to make the comparison as apples-to-apples as possible.

Important point: This map doesn’t take into account variable electricity rates and EVs tend to be charged overnight, so the real-world numbers are lower in many states, particularly in places like California where there’s a big difference between peak rates and off-peak rates (it can be less than half).

Source.

🇨🇳 China’s stock market has been flat since 1992 😬

Jake aka EconomPic:

THE STOCK MARKET IS NOT THE ECONOMY

Chart of the year so far... the MSCI China Index is about to go NEGATIVE since its 1992 inception

The Chinese economy is ~13x larger and an investment in it has gotten you the same return as if you put your cash under your mattress

The S&P 500 is up ~18x in that time frame

I’m not saying I have any idea about what it will do in the future.

Maybe this is the bottom, or maybe it goes way down if Xi does something stupid/evil.

But taking the long view provides a good reminder of shows how special the US system has been.

💳 💳 💳 💳 💳 Credit Cards in circulation globally 🌐

Things look different when it comes to debit cards. The top 4 by volume are all from China, with CCB taking the top spot with almost 1.5 trillion USD in volume in 2022.

Source.

---

🧪🔬 Liberty Labs 🧬 🔭

🥽 Apple Vision Pro manufacturing process 🍎

The ballet of machines, it’s beautiful!

I’ve always liked “how it’s made” videos. This is an ad, but it’s the same principle.

🤰🏻 China’s population FELL for a second year, birth rate keeps falling — end of one-child policy didn’t help 👩🏻‍🍼 🇨🇳📉

China's birth rate keeps going down, and the end of the one-child policy hasn't really changed the trajectory, so much so that combined with the pandemic, China’s total population has fallen for the past two years:

China's population fell for a second consecutive year in 2023, as a record low birth rate and a wave of COVID-19 deaths when strict lockdowns ended accelerated a downturn that will have profound long-term effects on the economy's growth potential.

The National Bureau of Statistics said the total number of people in China dropped by 2.08 million, or 0.15%, to 1.409 billion in 2023.

That was well above the population decline of 850,000 in 2022, which had been the first since 1961 during the Great Famine of the Mao Zedong era. (Source)

Decades of doing things one way is plenty long enough for it to become part of the culture and hard to revert quickly.

With that being said, birth rates are dropping almost everywhere else, so it’s not purely a problem in China.

But a shrinking population that is also rapidly aging can lead to a lot of problems and zero-sum dynamics. Japan is facing some of those issues, but it does so with almost 3x the GDP/capita of China.

🍪 Allulose Production Breakthrough Could Make It Plentiful and Affordable 🍩

Back in the Paleoproterozoic era of Edition #62, I wrote about Allulose, the best sugar substitute that we currently have (afaik).

I won’t go over everything again, but here are some highlights:

allulose is mostly absorbed in the small intestine without being converted into energy: at least 90% is excreted by the kidneys without being metabolized. This means that in a functional sense allulose has 95% fewer calories than sucrose and is why the FDA determined in 2019 that it does not need to be listed under total or added sugar.

But it’s maybe even better than just “not being a negative”, it looks like it could potentially even be a positive:

Data from animal studies suggest that compared to fructose and/or glucose, allulose may lower blood glucose, reduce abdominal fat, decrease insulin resistance and fat accumulation in the liver, and prevent or delay the onset of type 2 diabetes. In a recent meta-analysis of human trials, when allulose was given with carbohydrate-containing meals, it was found to decrease postprandial glucose by 10% (noting that the quality of evidence is moderate).

Attia wears a Dexcom G6 real-time blood glucose monitor, she he’s an interesting N-of-1:

I can anecdotally support that when I put allulose into black coffee, my blood glucose goes down. Usually black coffee would be neutral for my blood sugar, so this suggests that allulose is pulling glucose out of my body via my kidneys. In my experience, it also doesn’t leave me with that weird, slightly astringent aftertaste left by many sugar substitutes

A few years ago, I was able to easily buy big bags of allulose for a reasonable price.

We baked with it and loved it, as it looks & feels & tastes pretty much like sugar in the finished product (even that browning effect).

But that didn’t last — allulose rapidly became hard to find and/or very expensive. I went back to my #2 choice, Monk Fruit, but while it’s pretty decent, it’s not as good…

I’ve been hoping that allulose production would ramp up and catch up with demand, making it cheap and readily available. If CPG companies started cutting the sugar content in their stuff and replaced some of it with allulose (even if just 10% or 25%), few would notice, unlike with other sugar substitutes, and it could have a positive effect on public health.

Well, maybe the time is coming:

cientists at the University of California, Davis (UC Davis) have made a “significant breakthrough” in allulose production, with a method that offers both a high-quality yield and viable scalability, setting it up as a viable and healthier sugar substitute. It also rewrites the book on how allulose is currently sourced, which could swiftly advance its commercial production. [...]

Allulose – which is also known as D-psicose – is considered a rare sugar as it only exists naturally in minute amounts in a few plant foods, such as wheat, figs and raisins. When extracted, it has the texture and mouthfeel of sucrose [...]

Currently, it’s extracted using enzymes D-tagatose-3-epimerase (DTEase) and D-psicose-3-epimerase (DPEase) to catalyze its conversion from fructose, but limitations within this process has meant yields are stuck at a maximum of 50% and are of low purity.

The new production method uses the ubiquitous E. Coli bacterium:

The team, in collaboration with the Mars Advanced Research Institute, edited the microorganism’s metabolic processes, so when the cells were fed glucose, they converted it to allulose. It immediately resulted in yields of 62% (and, importantly, a purity level in excess of 95%). [...]

Essentially, the scientists took the E. coli, which naturally possesses the right pathways to produce allulose from glucose, and tweaked the design in order to flick the switch over to a specific metabolic outcome.

As well as being able to produce allulose, this new method is also sustainable and cost-effective to scale up production, using existing infrastructure and biochemistry techniques. (Source)

I can’t wait to be able to buy bags of the stuff on Amazon for a few bucks. Good work, U.C. Davis crew!

---

🎨 🎭 Liberty Studio 👩‍🎨 🎥

Award-winning Japanese Novel was co-written by ChatGPT

After Japanese author Rie Kudan won one of the country’s most prestigious literary awards, she admitted she’d had help from an unusual source — ChatGPT.

“I plan to continue to profit from the use of AI in the writing of my novels, while letting my creativity express itself to the fullest,” said the 33-year-old, who was awarded the Akutagawa Prize for the best work of fiction by a promising new writer on Wednesday.

The author then confirmed at a press conference that around 5% of her book “The Tokyo Tower of Sympathy” — which was lauded by committee members as “practically flawless” — was word-for-word generated by AI.

Source.