509: ASML's $700m <1nm Machine, Modest Proposal, EU Regulating Itself Out of the Race, OpenAI vs Microsoft, Apple, and Jupiter
"entering the angstrom-scale era"
You can’t value your own voice until people you respect question it.
—Dennis Lehane
🎥💰 I’m glad to see this 👆
I wonder if this film will noticeably move the needle in how an entire generation understands anxiety 🤔
This shouldn’t be underestimated: Many things get injected into the popular consciousness through culture and media (sometimes for the best, sometimes for the worst… the 1979 fiction film ‘The China Syndrome’ helped shape how a whole generation thought of nuclear power). This seems like it could be a good one.
🛀💭 Consumers love competition and choice in theory, but in practice, they usually want an obvious leader that stands out as clearly better than everything else ¯\_(ツ)_/¯
🧭🗺️📲 Pro-tip: before you travel, download offline maps for the area you’ll be in.
You can do it in both Google maps or Apple Maps. If you ever find yourself without a cellular connection or wifi, you’ll still be able to navigate using GPS.
🤔 Speaking of GPS navigation, both of my in-laws told me that they use Google Maps, but when I looked, neither of them had it installed on their phones.
They were using Apple Maps.
It was a good reminder that for many people, these are generic terms — like calling any tablet an ‘iPad’ — and defaults are extremely powerful
📬 We’ve got a bunch of follow-up today. First:
✉️ 😓😩 On anxiety, reader Kris made a good point in a comment:
One positive thing I think that can come from being anxious is vigilance. You will spend a lot of time researching and preparing for possible outcomes, which I think could be beneficial for financial related tasks.
The downside is the fatigue from being vigilant and also what you mainly focused on which is the possibility of making rash short term decisions based on irrational fears.
It’s a great point by Kris.
Anxiety is no doubt an adaptation — in the evolutionary psychology sense — and has many benefits.
Like many useful things, when pushed too far, things go wrong. But up to a point, I'm sure it can be a good source of motivation to plan and prepare.
✉️ 💊💉 Regarding the potential impact of Ozempic and similar drugs on motivation in areas of someone’s life other than food, friend-of-the-show Laurence Endersen wrote to recommend this book:
I haven’t read it yet, but it looks interesting.
💚 🥃 🚢⚓️ 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 💳 💴
🗣️ Interview: ‘Modest Proposal’ 🔥
Modest is great to follow on Twitter, and he’s great to listen to. This interview is self-recommending:
The discussion about under-supply leading to over-supply and how “Amazon built UPS in two years” (something I’ve discussed with you in past Editions) is a good reminder to keep the long view.
When we’re inside a cycle, it’s easy to see it as “the new normal”, but it rarely is. When prices get high, the market is very good at finding new supply.
🔬🐜 ASML’s Hyper-NA EUV Machines May Cost $700m Each 😯💰💰💰💰💰💰
The ASML machines necessary to make sub 1-nanometer chips — we’re getting close to entering the angstrom-scale era — around 2030 are expected to be rather pricey, to say the least:
ASML, the key lithography equipment supplier to the semiconductor industry, may charge up to NT$23 billion (US$700 million) for each of its next-generation Hyper NA EUV machines when they hit the market in 2030, media report, noting TSMC and Samsung face stiff bills for production line equipment.
Current ASML machines are already very expensive:
The report estimates ASML pricing:
EUV (current gen): US$181 million per machine
High-NA EUV (next): $362 million
Hyper-NA EUV (2030): $724 million
Think about how many of these foundries like TSMC and Samsung will need to produce the millions of chips the world needs 😬
Source (in Chinese), h/t Dan Nystedt
🇪🇺 Is the EU regulating itself out of the race? 🛑 🏎️ 🐢
The tech headlines out of the US and China focus on what is being built — bigger, better, faster things. The headlines out of Europe are all about more and more regulations and lawsuits.
Is Big Tech very powerful and deserving of close scrutiny and some clear regulation?
Absolutely.
However, is Europe’s approach effective? What kind of unintended consequences and second and third-order effects might it have? Is it likely to backfire over time?
EU bureaucrats have already ruined the web by forcing every website to have ineffective cookie popups like it’s 1998. This is annoying but may be benign compared to how they’re attacking the very foundations of some of the world’s most successful companies (ie. the vertical integration necessary for making certain kinds of innovation and product design).
Here are a few examples of what’s going on in Europe:
Apple may delay AI features in the EU because of its big tech law
Apple says upcoming features like its Apple Intelligence generative AI tools, iPhone mirroring, and SharePlay screen sharing may not be available in the European Union this year [...]
Why? The Digital Markets Act (DMA), says Apple [...]
According to Apple, the DMA requirements saying those gatekeepers must let third-party companies interoperate with their services could “force” compromises over privacy and security.
The DMA is written so broadly and ambiguously that it can easily be interpreted as meaning that Apple can’t add features to the iPhone that aren’t also available to Android and others on day one.
Product design by comittee of legislators doesn’t exactly have a great track record…
But it’s also not how innovation happens.
You can’t ask a company to have its most talented engineers and designers work on some features for years, spend millions in R&D and take big risks of things not working (think of all the dead ends that never see the light of day), etc… And then have a public API on day one that explains to competitors exactly how to rapidly copy your feature and implement it in such a way that there’s minimal product differentiation from all that investment.
It’s basically legislating that all platforms need to be beige and as similar as possible.
It’s removing incentives for innovation and risk-taking.
From a technical standpoint, it may seem easy to say “oh, just have an API”, but that’s also a big investment. It’s something that usually takes place many years into a development cycle once the code base and hardware have matured and are stable. For anything complex, asking for things to be opened-up to third parties from day one isn’t trivial at all, and will be expensive and slow down progress even after the first version ships. It can also compromise security because you can secure your own products and code base, but you have no idea if third parties will do a good job.
In this context, it makes sense for Apple to be cautious — especially since the way the DMA works in Europe, Apple can’t know if it is complying with the law until it starts selling something in Europe and then waits to see if they get a lawsuit. And if it is found to be non-compliant, according to the DMA it can be fined up to 10% of global revenue (😳) and can be increased to 20% for “repeated infringements” (😐).
For Apple, based on last-twelve months revenue, that would be around $38bn dollars (10%) and $76bn! (20%) 💰💰💰💰💰💰💰💰💰💰💰
The EU regulator feels very much like judge, jury, and executioner (⚖️), with a large incentive to impose huge fines because it throws that money into the EU coffers.
Maybe they keep raking in billions for a while, but over time, they may push it far enough that some of these companies simple exit the EU or sell a very limited selection of their products there.
EU citizens may root for regulators because it feels like they’re only going after the “big bad foreign companies”, but over time they may find themselves lagging behind in all kinds of metrics (investment, startups, innovation, productivity, wealth, etc), which isn’t just some theoretical thing — that means less wealth to fix real problems (healthcare, defense, education, infrastructure, etc).
Here’s another example, with Nvidia in France:
Nvidia set to face French antitrust charges, sources say
The French watchdog in a report issued last Friday on competition in generative AI cited the risk of abuse by chip providers.
It voiced concerns regarding the sector's dependence on Nvidia's CUDA chip programming software, the only system that is 100% compatible with the GPUs that have become essential for accelerated computing.
It also cited unease about Nvidia’s recent investments in AI-focused cloud service providers such as CoreWeave.
Companies risk fines of as much as 10% of their global annual turnover for breaching French antitrust rules, although they can also provide concessions to stave off penalties.
If they try to force Nvidia to open up CUDA to competitors, disallow them from having first-party only innovations, and/or try to fine them 10%/20% of revenue with more uncertainy about future fines because of the lack of clarity about what they can and can’t do, this will highly incentivize Nvidia to just get out of France.
This would hurt France more than it would hurt Nvidia, as its GPUs will still be used by the people and companies of France in the cloud.
Nvidia is similar to Apple in that it develops integrated products with software, hardware, and services working together. This vertical integration is at the heart of what the company *is*. If regulators try to attack the core of these companies, they will take it very seriously — as serious as a heart attack 🫀
One more, which came out yesterday: “Meta accused of breaching EU antitrust rules over ad-supported subscription service”.
In another case of design-by-regulator, they’re saying that Meta should offer an ad-supported version of their product that doesn’t do ad targeting, which would make it *way* less profitable, since the targeting is what makes the ads valuable.
In its statement Monday, the commission said that Meta’s ad-supported offering failed to comply with the DMA for two key reasons: one is that it doesn’t let users opt for a service that uses less personal data but is still equivalent to the “personalized ads”-based service.
Regulators said users should still be entitled to “get access to an equivalent service which uses less of their personal data, in this case for the personalization of advertising.”
The other reason cited by the European Union is that the Meta ad-supported service doesn’t allow users to exercise their right to “freely consent” for their personal data to be used to target them with online ads.
Here too, Meta could potentially be fined 10-20% of global revenue for not complying with the vague law (with no way to know in advance if they comply or not)…
🤖🤝💰 David vs Goliath: OpenAI's AI Model Revenue Surpasses Microsoft's (but that’s not the whole story…)
When Microsoft formed an alliance in 2019 with OpenAI to resell the startup’s artificial intelligence, it brought the reputation, connections and sales muscle to the table that seemed key for winning deals with large customers. Turns out OpenAI itself has gotten pretty good at cutting those deals. [...]
As of March, OpenAI was generating around $1 billion in annualized revenue from selling access to its models, according to someone who viewed internal figures related to the business. The annualized revenue rate refers to the prior month’s revenue multiplied by 12. In contrast, Microsoft’s comparable offering, Azure OpenAI Service, only recently hit $1 billion in ARR, OpenAI CEO Sam Altman told his staff this month.
Although it’s an impressive achievement for the much smaller and younger company, it should be kept in context: Microsoft is monetizing OpenAI’s models in many other ways than the Azure OpenAI Service (f.ex. Github Copilot and Office 365 integration).
There’s also the fact that Microsoft only started selling that API in January 2023 — I’d be interested in seeing the growth rates of both to see if maybe Microsoft is catching up and on a trajectory to pass OpenAI.
It’s kind of a strange dynamic that the partners are competing against each other in the market — pushing each other to lower prices and minimum commitments — while also working together to plan training super-clusters.
While this competition may mean leaving money on the table in the short term, perhaps it is making them harder to compete with in the long term by keeping them a bit meaner and leaner than they would be if they didn’t have this rivalry with the other company which offers the same models to customers 🤔
OpenAI also beefed up its sales operation, growing its team to more than 200 people from around 10 people a year ago, James Dyett, head of strategic accounts, said in a recent podcast interview. Dyett said his current priority is winning over a few large customers and keeping them happy.
Payments startup Stripe, for instance, previously bought OpenAI’s models both through Azure and directly from OpenAI, but more recently it has appeared to shift its spending to OpenAI, according to someone familiar with the situation. AI search startup Perplexity has made a similar shift, according to two people with knowledge of the change. (Source)
One advantage that OpenAI has over Microsoft is that their agreement has a loophole which allows them to offer near-final versions of their latest models to their customers before Microsoft gets access. This is because OpenAI only needs to provide Microsoft with completed models.
💪🍎🔎 Apple Poised to Get OpenAI Board Observer Seat
Another signal of who has the leverage in the relationship:
Apple Inc. will get an observer role on OpenAI’s board as part of a landmark agreement announced last month, further tightening ties between the once-unlikely partners.
Phil Schiller, the head of Apple’s App Store and its former marketing chief, was chosen for the position
How much did Microsoft have to invest to get an observer?
The board observer role will put Apple on par with Microsoft Corp., OpenAI’s biggest backer and its main AI technology provider. The job allows someone to attend board meetings without being able to vote or exercise other director powers.
🧪🔬 Liberty Labs 🧬 🔭
🛰️ Jupiter like you’ve never seen it 😯
This shot was taken by NASA’s Juno. Note that the image has been processed to better show the atmospheric features.
🇨🇳🇮🇳 China and India Lead Nuclear-Power Expansion ⚛️
Even Russia is doing better than the West…
🕵️♂️🤖 CriticGPT: Spotting GPT-4’s Mistakes with… GPT-4 🤔
It’s kind of cool that this works:
We've trained a model, based on GPT-4, called CriticGPT to catch errors in ChatGPT's code output. We found that when people get help from CriticGPT to review ChatGPT code they outperform those without help 60% of the time.
We are beginning the work to integrate CriticGPT-like models into our RLHF labeling pipeline, providing our trainers with explicit AI assistance. This is a step towards being able to evaluate outputs from advanced AI systems that can be difficult for people to rate without better tools.
I’m reminded (again) of M.C. Escher’s drawing:
Something like CriticGPT can help with one of the limitations of Reinforcement Learning from Human Feedback (RLHF): You need humans who are able to understand the output, find the mistakes, and give good feedback.
With very subtle mistakes or very complex output (like computer code), it’s very challenging.
And it doesn’t have to be just one or the other: Humans who are augmented by a model like CriticGPT perform better than humans alone or CriticGPT alone (though for how long?) ¯\_(ツ)_/¯
It also makes me wonder about how opaque a system that is largely being checked by AI becomes. These models are already mostly black boxes. Could this make them even harder to understand, or could such models potentially be leveraged by interpretability efforts..?
🎨 🎭 Liberty Studio 👩🎨 🎥
🎥 The Cinematography of ‘Dune: Part 2’ 🪱
Great video essay (🚨Spoilers
🚨 if you haven’t seen the film — go watch it tonight!).
That scene looks *so* good. I’m glad I got to see it twice on an IMAX screen.
Love that nuclear reactor Visual! 👍
Another great article. Thanks