569: Zuck + Alexandr Wang's $14.8bn Bromance, Nvidia, John Arnold, China's Electricity, Fixing Ozempic Muscle Loss, and Gene-Editing Mosquitoes Into Extinction?

"bleeding-edge, actionable insight on 18th-century instruments"

Patience is a competitive advantage.

In a surprising number of fields, you can find success if you are simply willing to do the reasonable thing longer than most people.

—James Clear

🛀💭 🎹 Okay, this one is esoteric af, but hey, sometimes questions pop into my head and I have to find answers! 🤷‍♀️

If you’ve seen the excellent Miloš Forman film ‘Amadeus’ (1984), you may have noticed that in the infamous masquerade ball scene where he plays upside down and mocks Salieri’s music, he’s playing on what looks like a piano, but it sounds sharper, brighter, more metallic.

That’s a harpsichord (the ‘march’ scene in the film features the fortepiano, the ancestor of the modern piano, invented by Bartolomeo Cristofori in 1700).

What’s the difference between a piano and a harpsichord?

Which technical innovation made the piano sound so much… nicer? More expressive? Just… more pleasant on the ear?

The biggest difference comes from the mechanism of sound production:

• A harpsichord uses quills or plectra (imagine small guitar picks) to pluck the strings when a key is pressed. This plucking gives a more brittle, percussive, and uniform tone, with little variation in volume. Press a key harder, and it doesn’t get louder.

• A piano uses hammers to strike the strings. This gives it a wide dynamic range: You can play soft or loud depending on how hard you hit the key.

Fixed vs. variable dynamics!

Another big difference is resonance and sustain:

• Harpsichords have little or no sustain. Notes decay quickly after being plucked. There’s very little string resonance, and the instrument’s construction doesn't amplify overtones the way a piano does.

• Pianos, on the other hand, have felt dampers that lift when you press a key, allowing the string to vibrate freely, especially with the sustain pedal. That’s what creates the rich, sustained, resonant tone.

I bet you didn’t expect to learn about harpsichords when you woke up this morning!

Only the most bleeding-edge, actionable insight on 18th-century instruments here! 😅

Next week: clavichords vs dulcimers? (Kidding)

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

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

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

Can't upgrade right now? Share something you enjoyed here with someone who might like it too. Drop a screenshot and a link in your group chat! 💬

---

🏦 💰 Business & Investing 💳 💴

🤖💰 Meta Kinda Acqui-Hires Scale AI CEO Alexandr Wang for $14.8bn… and a 49% Stake in his Company 🤔

This one is a bit strange.

Meta apparently made a deal to take a 49% stake in Scale AI for $14.8 billion (!):

The unusual deal will be structured so Meta will send the cash to Scale’s existing shareholders and place the startup’s CEO, Alexandr Wang, in a top position inside Meta, the people said.

(Where’d Wang’s second “e” go, btw? Did his parents work at Flickr?)

That’s a lot of money for a minority stake in a company that focuses on data annotation, labeling, expert & synthetic data generation, RLHF workflows, etc, using *an army* of contractors, reported to be “more than 100,000” in a 2024 WSJ profile. Who knows how many today ¯\_(ツ)_/¯

It feels like Meta would've bought the whole enchilada if they thought regulators would allow it.

Meta would put Wang in charge of a new “superintelligence” lab, along with other top Scale technical employees [...]

That will put Wang, 28, in competition with some of his customers and friends, including OpenAI CEO Sam Altman.

That could spook Scale’s biggest clients, who are all the biggest names in AI. They now face the awkward reality that their vendor is also competing with them.

In other words, it’s a bit like if TSMC suddenly decided to design its own chips and compete with AMD, Nvidia, and Apple.

Wang is also not going to be involved in the day-to-day of Scale anymore:

Scale faces important questions following the deal. It removes Wang from the day to day of the company he founded in 2016 [...]

Jason Droege, a former Uber Eats executive who became Scale’s chief strategy officer last fall, has been in discussions to become Scale’s new CEO, one of the people said.

Losing the founder has to make the company less valuable going forward, no?

Is this a signal that Meta’s own AI efforts were going off the rails, as I mused about recently, and that they’ve decided that the best way to shake things up was to inject outside top-quality leadership?

At this price tag, a big part of what Meta must be paying for is *speed*. Because for $30bn, they must be able to buy/build a lot of data assets, but that would take time... ⏳

But here's the key question: Is Wang the right person for this? 🤔

He has focused on a very important part of the AI pipeline, yes, but is this the area where Meta has been falling short?

It’s very possible that Wang is also talented when it comes to other parts of the AI stack, but so far he hasn’t had a chance to demonstrate running a frontier lab. I’ll be curious to see if he can make the transition to running what sounds like it may be an AGI-focused lab.

💵 2.5 Years After ChatGPT, AI “Startups” Revenues Hit $15 Billion Run Rate 🚀

Power laws! Always power laws!

OpenAI is generating two-thirds of the revenue in the chart—$10 billion in annualized revenue—though the startup must give Microsoft 20% of revenue as a thank-you to the software giant providing it with $13 billion-plus worth of cloud servers.

Anthropic, which is no. 2 in the revenue rankings, at $3 billion in annualized revenue, has to give an even higher percentage of the revenue it generates via Amazon and Google, which resell its models to their cloud customers.

The biggest winners not shown on the chart? The hyperscalers and the picks & shovels merchants like Nvidia, AMD, and Broadcom. 🪏⛏️

The cloud providers are generating billions of dollars a year in revenue from renting out servers to the biggest developers—OpenAI and Anthropic in particular. Anthropic burned $5.6 billion in cash last year, largely to pay for servers and related costs of developing its models.

Source.

🩻🔍🕵️‍♀️ Nvidia Highlights: DeepSeek Optimizations, Access to China, Networking, etc. 🤖🇨🇳🛠️

Executives from the accelerated computing company have been making the rounds on the conference circuit lately, and I thought I’d share a few highlights.

First, Ian Buck, the VP of Accelerated Computing at Nvidia, which includes Hyperscale and HPC Computing Business:

Q: Do you think anything that DeepSeek is doing or what's happening in China as a proxy for let's call it, CapEx constrained computing. Do you think they are able to bend the cost curve in a way that has implications on how much spending needs to happen in this industry?

Buck: No. Actually the opposite. They just talked about what everyone was doing in an academic paper. Computing has always been constrained. Access to compute, dollars of compute, capital expenditure of compute.

The AI race is about: regardless of how much compute you have, how efficiently you're using it, how intelligently you're using it and how much value you bring. That wasn't unique to DeepSeek. It was around the world.

It's just, do you have the engineering talent to capitalize on it to code your CUDA, [you] know your InfiniBand, know your NVLink, optimize your transformer layer.

It’s a good point, though, of course, life is trade-offs.

Many Western labs clearly decided it was a better trade-off to throw more compute at problems and use their scarce engineering resources on other parts of the stack.

On DeepSeek's specific optimization technique: 🛠️

Buck: One of the big innovations that DeepSeek did is they used a new technique called MLA, which actually is a statistical method for approximating the weights and the KV layers of the transformer layer.

It wasn't a new idea.

It’s actually been deployed in image generation, all those fun, drawing a picture of a teddy bear, swimming in Olympic lap. They were using this MLA statistical technique, but it compressed the Jesus out of the transformer layer, made it a lot cheaper by approximating and they were able to apply it to DeepSeek-V3 and R1.

That was the first time [they] had been publicly talked about. Trust me, these methods are being deployed and optimized just not everyone wants [to talk about it].

DeepSeek themselves are doing the world a favor by sharing some of the state-of-the-art research they're doing. But it's happening everywhere.

It was happening even back in the A100 days as well.

He also had a great line about the timeframe mismatch between building AI datacenters and who the customers of these datacenters are:

It's a fascinating business model. 

[D]ata centers are bought with billions of dollars, 5 years of CapEx, and you end up charging dollars per hour or millions per token at [the other] end.

At a separate event, Nvidia CFO Colette Kress discussed how big 'sovereign AI' might become:

Kress: How large is sovereign is always the question in front of us, but it is going to be a very, very large piece.

Look at it in this perspective: every country will need their own ability to have their AI within their country. […] That's the ability for you to have your own language, your own culture, your own data that you will likely want to keep inside of that country. [...]

And all countries, all enterprises, all people, all consumers are all thinking about how AI would work there. [...]

We have been in the Middle East, as you indicated, and we ended up speaking with not only the Saudi Arabia, but also the UAE. And I think it led to what you heard is the tens of gigawatts that would be available through many of those nations. [...]

again, when you look at the size of this, you can be approaching over several, several years, could be close to $1 trillion.

On China: 🇨🇳 💸

Kress: So in the middle of the quarter, we received notice from the U.S. government that we would not be able to ship our H20. Now keep in mind… we developed them [with] a lot of back and forth with the U.S. government, with continuous approval.

And unfortunately, they chose to not allow it to go.

Now that means that we really don't have anything for [the China] market.

They can’t exactly spin up a new China chip overnight:

We've discussed that it wouldn't be appropriate for us to just start a new chip at this point because essentially, the H20 from where it compared to our Blackwell architecture was significantly lower in terms of what we were being able to enable in China. That was about a 25x change from an H20 to what you would receive in terms of a Blackwell [...]

It's a big market, though. China is a very, very big market. We can think about it just today or this year, probably could be about a $50 billion market. That's a great opportunity for us to continue to innovate, continue to build the platform from the U.S. to the rest of the world, and we think that's an important market for us to go and do.

So again, still in discussions with the U.S. government, and we'll see.

She also mentioned that their networking division is performing strongly with an attach rate of about 70% when they sell a data-center GPU.

🗣️ Interview: John Arnold on Trading, Energy, and Evidence-Based Philanthropy 🛢️⚡️🏢🏥🏢

I enjoyed this interview with John Arnold by Tyler Cowen.

They cover a lot of ground, including his early days trading natural gas at Enron, why he decided to retire, and his push into evidence-based philanthropy.

• 🎧 John Arnold on Trading, Energy, and Evidence-Based Philanthropy

Here’s a highlight on what made him a great trader and why he retired:

Number one is this detachment from emotion. There’s a lot of talk about fear and greed driving markets. To the extent that fear and greed change your process, the more you can remove those emotions, I think, the better… There is this notion of being on the perfect point of the confidence spectrum. You have to be confident in order to say the market is wrong, and I’m right… but if you’re overconfident, you’ll blow up quickly. [...]

I ate, breathed, and slept it. I would be thinking about it first thing in the shower in the morning. I would be dreaming about it. After work, I’d go out with people in the industry and talk about it. It was that real devotion to the markets... Then years 15 and 16, I started to feel that the end is coming at some point. By year 17, I didn’t enjoy the game anymore. I could feel like if your drive, if your passion isn’t there, that you weren’t going to be successful. That was part of the reason for stepping away.

---

🧪🔬 Science & Technology 🧬 🔭

🔌⚡️ Electricity Generation: China vs World 🦾🤖🤖🤖🇨🇳

Prosperity, security, and progress are built on energy.

One of the main bottlenecks for AI is going to be energy (especially post-2030).

This graph ☝️ worries me.

Democracies need to relearn how to build (India, that includes you) 🏗️🪏👷‍♂️

💪 Fixing Ozempic: Researchers May Have Solved GLP-1s' Biggest Flaw — Muscle Loss 💊🚀

GLP-1s are the wunderdrug of the era, but they have a big downside:

“Up to 40% of the profound weight loss induced by these agents can be due to loss of lean mass.”

In other words: sure, you’re melting, but you’re also losing precious, hard-to-rebuild muscle. 😩

It can still be worth it, and you can take measures to mitigate muscle loss (lift MOAR, eat MOAR protein! 🏋️‍♂️), but what if we could avoid this side effect entirely?

This seems to be the promise of a new class of antibodies (GDF8 and Activin A) that block muscle-wasting signals:

Importantly, the Sema + α-MSTN/α-ActA group – which had similar body weight reduction but increased lean mass compared to Sema alone – also had markedly greater losses in fat mass, almost twice that seen with Sema alone (Fig. 1c). 

These data indicate that semaglutide alone results in weight reduction involving both fat and lean mass loss, while adding α-MSTN/α-ActA to semaglutide results in similar weight reduction but with profoundly beneficial effects on body composition – not only by preserving/increasing lean mass but also by markedly increasing fat loss.

In plain English: when animals received both the weight loss drug AND the muscle-protecting antibodies, they had similar total weight loss but preserved and even gained muscle mass. This means that they lost almost twice as much fat compared to the weight loss drug alone and had improved blood sugar, cholesterol levels, and liver health. 👩🏻‍⚕️🩺🫀

Part of this is probably because having more muscle mass increases energy expenditure and is a great sink for glucose.

It also seems like these antibodies — apart from any weight-loss effect when combined with GLP-1s — could potentially be useful to fight sarcopenia, the loss of muscle mass as we age.

This could be a HUGE breakthrough. Let’s hope that it translates well for humans and there are no show-stoppers.

Early signs are good:

In an accompanying manuscript (Gonzalez Trotter et al.), we have extended these findings to human volunteers without obesity and further showed that the profound muscle increases induced by blocking both GDF8 and ActA in humans are also accompanied by a loss of fat.

No serious safety concerns were identified, though there were some ulcers, headaches, and muscle spasms.

Here’s the paper.

🦟💀🚫🧬 Should We Gene-Edit Mosquitoes Into Extinction? (Please Let Me Press the Button)

I hate HATE hate mosquitoes (and for some reason they just LOVE me 😩):

When so many wildlife conservationists are trying to save plants and animals from disappearing, the mosquito is one of the few creatures that people argue is actually worthy of extinction. Forget about tigers or bears; it’s the tiny mosquito that is the deadliest animal on Earth.

The human misery caused by malaria is undeniable. Nearly 600,000 people died of the disease in 2023, according to the World Health Organization, with the majority of cases in Africa.

On the continent, the death toll is akin to “crashing two Boeing 747s into Kilimanjaro” every day

This isn’t just about itchy bites!

In their labs, the scientists have introduced a gene mutation that causes female mosquito offspring to hatch without functional ovaries, rendering them infertile. Male mosquito offspring can carry the gene but remain physically unaffected. [...]

Now, some doctors and scientists say it is time to take the extraordinary step of unleashing gene editing to suppress mosquitoes and avoid human suffering from malaria, dengue, West Nile virus and other serious diseases. (Source)

It wouldn’t be easy, as there are many species, but I think this is one of those global projects that could unite the people of Earth, like going to the moon or eradicating smallpox.

Yes, I know, potential unforeseen consequences, etc.

But we keep hearing about how thousands of species go extinct all the time…

One more can’t be that bad, right? 😈

---

🎨 🎭 The Arts & History 👩‍🎨 🎥

🕰️ Guess Who Was Alive at the Same Time! 🗓️

The “Arts & History” section is mostly the former and rarely the latter, but today let’s switch it up a bit.

The author of the graph is ‘profound_whatever’ on Reddit.

If you can’t read the names, click on the graph to get the high-resolution version.