487: Apple's AI Strategy, OpenAI Goes to Hollywood with Sora, Temu + Meta & Google, Poisoning AI Training Data, China's Coal, and IMAX

"There are horrors in everyone’s past."

Ornate theories impress people at dinner parties, but common sense impresses fate.

–The Stoic Emperor

Made with Midjourney V6 💚 🥃

🐍🦍🔥🌳🏚️🏠🏛️ Listening to my friend David Senra’s (📚🎙️) excellent podcast about ‘The Lessons of History’ by Will and Ariel Durant (they could really turn a phrase!) made me think about what my family’s long-term history must be like, in the context of what the world has been for most of its existence.

As one of the lucky ones, living in comfort in a peaceful wealthy country, with antibiotics, vaccines, and very low child mortality, it’s easy to forget what life has been like for most of humanity’s history, and for many around the world today.

We think it’s others.

There are horrors in everyone’s past.

We all have countless ancestors who lived through wars and famines, were hunted by predators or roving bands of marauders, broke bones that never set right or lost loved ones to random infections that today would be dealt with by a quick visit to the drug store. During the Black Death in the 14th century, an estimated 30-60% of Europe's population died. And it wasn’t just the natural world: In a zero-sum world of scarcity, the way to amass wealth and power was to take from someone else so warlords and conquerors raped and pillaged at intervals — today we read their biographies, but back then our ancestors were on the receiving end.

Our family trees are by definition made up of the lucky ones, those that lived long enough to procreate. But that relative luck doesn’t mean that their lives were easy.

We feel apart from the natural world, but for most of humanity’s existence, for thousands of generations, there was no separation.

Friend-if-the-show Jason Crawford (🚀👨‍🔬) reminds us of how things were not so long ago:

"Philippe Ariès and Lawrence Stone, in a landmark study of the English family, suggest that in periods of high mortality parents protect themselves against the emotional pain of a child's death by remaining affectively aloof. From this perspective, it is 'folly to invest too much emotional capital in such ephemeral beings.'"

The concept of children and childhood as precious is a modern sentiment made possible by the conquest of infectious disease

When half of children died before adulthood, it was emotionally impossible to become too attached to them.

I don't mean that parents didn't love their children, or that they weren't emotionally devastated when the children died. But there is *something* about the way we value children today that is very different from the past.

Let’s not take for granted the civilization that has been built for us by our ancestors. Let’s protect *and* improve it for our kids and their kids.

There’s a lot to fix and upgrade, but I’m glad that my kids are safe and I can earn my next meal by thinking about ideas and manipulating abstract symbols.

🌡️🥵🧠 After years of research, planning, and dreaming… We’re getting close to having a backyard sauna!

If you’ve been reading for a long time, you may remember past Editions where I talked about why saunas are so beneficial. A large part of the effect comes from mimicking cardio exercise and positively impacting sleep quality. Two of the very best things for your health are exercise and sleep, so that’s a 🎯

If you want to dive into that rabbit hole, there’s a good overview of the science on saunas here. 🐇🕳️

Anyway, my wife and I went to see some models last weekend. Initially, we had a different fave model, but after seeing a few in person, we’re now thinking of going with the one pictured above.

There are multiple configurations for benches, heaters, and accessories. We want an electric heater and L-shaped 2-level benches. The sauna is deep enough that my wife could lay down on the short part of the L and I could lay down on the long part, or we could have 4-6 people sitting,

🗣️🗣️🗣️🗣️🗓️📚🎟️ Quick reminder that if you’re anywhere near NYC on April 13th, you should come to the Future of Publishing Event organized by Interintellect and OSV.

I’ll be there along with all these interesting people (I aim to be the dumbest person in the room).

And just for you, you can buy your ticket here and get 20% off with the coupon “FOPE20”.

👶👩🏻‍🍼🎉 There’s a baby boom going on in the Liberty orbit! Another entry in Good News Corner, friend-of-the-show Andy aka Bizalamnac (💚 🥃) sent this press release on the newswire:

“Almanac Family is pleased to announce their second spin-off, who began trading 03/22/24

Chairman and Spokesperson, Biz, shared how impressed he was with CEO, Mrs. Almanac, as well as optimistic growth prospects for both children

No further guidance was provided"

💚 🥃 🚢⚓️ 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.

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 you make a big difference.

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

---

🏦 💰 Liberty Capital 💳 💴

Made with Midjourney V6 💚 🥃

🍎🤖📲 Apple’s Summer of AI: Let’s Think Through Apple’s AI Strategy 🤔

Rumors are circulating that Apple is talking to Google about licensing Gemini. No details, just “in active negotiations”.

This has sparked more rumors about potential partnerships with OpenAI, Anthropic, etc.

Let’s think this through:

No company can be the best at everything.

Life is trade-offs.

What makes you good at one thing may make you worse at something else.

Apple is a product company. They are organized internally to maximize the odds of shipping high-quality products (vertical integration of hardware-software-services, functional org, a culture that grew around products, etc).

Despite the services narrative, everything revolves around products. 📱💻🖥️⌚️

Service revenues (App Store, Google Search deal, iCloud, Apple Music, etc) are strong because they sell a lot of devices and have a large installed user base. But we shouldn’t forget which is the dog and which is the tail (as an aside, a lot of Apple’s recent problems come from them letting the tail wag the dog, but that’s a different discussion…).

In the tradition of commoditizing complements, I expect Apple to view AI as something that makes a product better, not the main course. Their priority is to sell more iPhones, not monthly subscriptions (those will come if they sell products).

They should frame it something like this:

Others are great at making foundational models. We want to focus on is the *experience* of using these models. So we’ve created Apple Air Control™️, an AI that sits on top of other models, abstracts them away, and picks the best tool for the job, all with a beautiful and intuitive interface!

Do most users care which model is under the hood as long as it is really good and useful?

If you doubt that UX can be critical, remember that OpenAI’s breakthrough was taking GPT-3 and creating the ChatGPT UX (partly design, partly tuning the model so it’s more conversational — but it was mostly a product breakthrough, not a technical one).

If I were Apple, I’d go with a constellation of models. Fully modular.

I’d start with small models that can run on-device to take advantage of the very powerful silicon in iPhones and ARM Macs. They have fast GPUs and Neural Engines that are great at accelerating generative AI and their unified memory architecture means that more RAM is available for genAI than for most of the competition.

If simple tasks are run locally, you get lower latency and save on cloud inference costs.

Anything complex can be sent to larger models in the cloud. That’s where an Apple traffic cop model 👮🏻‍♂️ would do triage: some tasks would go to open-source models like Mistral or LLaMA and others to OpenAI, Anthropic, or Gemini. 🤖🤖🤖🤖

By using multiple models, they could probably get better pricing and potentially better performance by selecting the model that is best at any one moment or for a specific task. Things change so fast that today’s leader may be tomorrow’s laggard.

This would also keep the spotlight on Apple’s brand, not on the models (ie. from a user’s perspective, you’d think of going to Siri AI, not ChatGPT or Gemini).

Having multiple models would also help distribute the load since Apple’s installed user base is incredibly large and active. This is why it makes sense that they’re talking to Google — they have the infrastructure to handle such a large load (plus they already have a very lucrative search deal with Google, so this could be seen as an expansion of that).

But while Google has lots of infrastructure, Google + others = even more infrastructure.

It would make a lot of sense for Meta to be one of the partners if the two companies could ever bury the hatchet. 🪓

Having a non-exclusive relationship with model vendors would make Apple less dependent on a single partner as they keep working on their in-house capabilities. This is similar to Apple’s deals with the hyperscalers for iCloud storage and compute.

Whatever happens, we’re likely to know this summer. Apple’s developer conference (WWDC) seems like the perfect moment to unveil a bunch of AI features.

🤖🎥 OpenAI Showcases New Sora Videos, Goes to Hollywood 🎬

Check out this page showcasing a bunch of videos made by various artists using Sora, along with their thoughts about the process. If you read between the lines (so to speak), you can see the limitations of the model by what the artists choose to do and what they don’t do.

But as a vector rather than a point, this is all very 🤯

Remember, it wasn’t that long ago that we were impressed that genAI could even make images at all. Do you remember the avocado chair?

Think about where this is going, not where it is today…

High-quality genAI images are common (I’ve used a few in this Edition). Video is the next frontier:

OpenAI wants to break into the movie business.

The artificial intelligence startup has scheduled meetings in Los Angeles next week with Hollywood studios, media executives and talent agencies to form partnerships in the entertainment industry and encourage filmmakers to integrate its new AI video generator into their work, according to people familiar with the matter. [...]

In late February, OpenAI scheduled introductory conversations in Hollywood led by Chief Operating Officer Brad Lightcap.

Lightcap demonstrated the capabilities of Sora, an unreleased new service that can generate realistic-looking videos up to about a minute in length based on text prompts from users. Days later, OpenAI Chief Executive Officer Sam Altman attended parties in Los Angeles during the weekend of the Academy Awards.

That’s going to generate a lot of heat. There’s a storm brewing over this.

While many in the industry are excited to use the new capabilities to better realize their vision or take over the low-value, repetitive aspects of visual effects so that humans can focus more on the high-value, more creative aspects, many in Hollywood are very strongly anti-AI, seeing it as a threat to their livelihoods (and for many, it probably is).

The big wild card is regulators and the court system. Who knows how lawsuits over copyright will shake out?

I think the sane way to manage it is that copyright should be enforced on outputs, not inputs.

Human artists learn from mountains of copyrighted materials, and that’s desirable! But if an artist recreates something that is copyrighted, they shouldn’t be able to profit from it (if you try to make a video game full of Pixar and Marvel characters, you’ll quickly hear from a Disney lawyer).

Same for AI models.

They should be able to *learn* from public information even if copyrighted — extract patterns and statistical relationships from them, similarly to how humans learn from things by abstracting out the patterns *not* by having a perfect copy of everything they’ve ever seen or heard in their brains — and as long as what genAI produces is transformative/new, I don’t see a copyright problem.

🛒🇨🇳 Temu Spending Big Bucks on North American Ads 💸

Meta Platforms’ top advertiser by revenue in 2023 was the e-commerce company Temu, an upstart discounter founded in China [...]

PDD Holdings, the parent company of Temu, spent nearly $2 billion on advertisements last year at Meta… Temu also became one of Google’s top five advertisers by spending last year, according to people familiar with the business.

Not only are they spending big on ads, but they’re trying to improve their logistics:

Temu, fast-fashion giant Shein and other online shopping platforms with Chinese roots are spending aggressively to reach American consumers, pushing up digital advertising prices, poaching logistics employees and delivering so many products that they have become a boon to the shipping industry.

It remains to be seen whether Temu and Shein can create profitable businesses in North America or if once they drop the heavy promotion they have trouble turning a profit. (they spent tens of millions just on the Superbowl!)

Another potential snag would be if the relationship between the US and China keeps deteriorating and politicians decide that TikTok isn’t the only Chinese company that they want to block.

---

🧪🔬 Liberty Labs 🧬 🔭

🪨 Burning Coal in the UK vs China 🏭

Note the Y axis. At its peak, the UK still burned less coal than China in 1960. Of course, populations are very different…

🏎️ ⚛️ China is Building Nuclear Power Plants Quickly 🇨🇳🏗️

One of the things likely to bend the curve on the graph above is that China is good at building nuclear reactors:

Nearly every Chinese nuclear project that has entered service since 2010 has achieved construction in 7 years or less. This real-world trend flies in the face of the tiresome and longstanding claims that nuclear energy technologies inherently exhibit a negative learning curve, and that nuclear reactors require a decade or more to build.

This graph shows how long each plant took:

Note the orange cluster. Some models take longer, but that’s still under 10 years.

Every single conventional commercial-scale reactor project in Chinese history has achieved completion in under a decade [...]

annual nuclear electricity production has actually kept an eyebrow-raising pace of its own, increasing by 343 TWh [since 2010] (75 TWh to 418 TWh). [...]

Since the start of 2022, China has completed an additional five domestic reactor builds, with their completion times ranging from just under five years to just over 7 years. This continues the consistent completion record of Chinese projects even despite potential disruptions from the intervening COVID-19 pandemic.

Why?

the large-scale growth of industrial and civil infrastructure countrywide in past decades has cultivated considerable megaproject management experience and construction capacity. In particular, private and public-sector projects have learned to target construction economies of scale by planning and co-locating multiple identical units or manufacturing lines at the same site, organized in successive phases of site development and expansion (Figure 2). This base of heavy industrial capacity in turn aids nuclear project development thanks to the wider availability of supply chain assets like heavy forges for reactor pressure vessel construction.

More details here.

☠️ Deliberately Poisoning AI Training Data 🏴‍☠️

That sounds worse than it is, but it’s still a theoretical way to damage foundational models:

While data poisoning is a concern with all types of machine-learning algorithms, some researchers say generative AI models could be particularly vulnerable because they must ingest vast amounts of text, imagery and other data from the public internet to gain the knowledge they need to create something on their own.

This reliance on a vast number of data sources from the open web—rather than curated, locked-down data sets, which are harder for hackers to penetrate—could make it difficult to spot and eliminate poisoned data, only a small amount of which is needed to affect AI’s outputs.

If you can figure out what is being scraped and when, you may be able to find vectors to inject bad data in the training set.

The team looked at Wikipedia, which is used to train many large language models. Wikipedia doesn’t allow companies and researchers to individually scrape the site for information; rather, it provides a complete-site snapshot periodically. This is a regularly scheduled event, so if attackers know the articles likely to be included in an AI model’s training data set, they could edit those articles to include false or misleading information right before the snapshot is created, according to the researchers.

Even if the bad entries are fixed quickly, the poisoned snapshot would remain, and any AI models training on the snapshot would digest poisoned information, says Tramèr. He estimates that about 5% of Wikipedia articles could be manipulated in this manner.

Something similar could be done by purchasing expired domains and replacing the original content with an altered version. It would represent a small % of the training data, but if you target it to a specific topic that has few other sources, a small % may be all that is required.

The paper is here if you want to dig in.

🍳 Cooking is chemistry & physics… So we may as well understand the science of splatter! 👩🏻‍🍳

This is very useful info to keep your kitchen clean and make cooking more enjoyable.

(you may have to click to see the video, I think they disabled embedding)

---

🎨 🎭 Liberty Studio 👩‍🎨 🎥

🎥 Video Essay on IMAX & Aspect Ratios 🎞️

Great explanation of the various aspect ratios that filmmakers have to deal with, what’s different about IMAX, and the trade-offs involved.

It’s also a great reminder that we’re living in such a golden age compared to the VHS pan-and-scan era that I grew up in…