499: Microsoft vs OpenAI, Google UX, Anthropic Hires Instagram Founder, Musk + Oracle, Chinese EVs, Peak Coffee, Gene Therapy, $20bn Fabs, and Hans Zimmer

"Moral judgment is a stranger process than we realize"

Clarity is forging your imagination into a pebble that, when tossed, will ripple through other minds.

—Eliot Peper

🛀💭 What if Hermann Goering had looked like Fred Astaire?

What if Lavrentiy Beria had looked like Clint Eastwood?

What if Pol Pot had looked like Bruce Lee?

What if Kim Jung Un looked like George Clooney?

Would history perceive them differently?

I don’t mean a complete reversal, from villains to heroes. But if they currently stand at ‘-10’, would they be viewed more positively if they were very handsome? And if so, by how much? From ‘-10’ to ‘-9.5’ or something more dramatic like ‘-7’? 🤔

And is it not kind of insane since their looks have nothing to do with the atrocities they committed? Moral judgment is a stranger process than we realize.

But it’s a fact that, on average, we can hate ugly people more easily… It’s the other side of the coin of the halo effect. That’s why propaganda and dehumanization campaigns always make the other side look bad.

🏢🥵🔄🥶 As much as it is valued, competence is still undervalued. We need to start a cult of competence. We need to foster a culture that values competence for its own sake and teaches the intrinsic satisfaction of a job well done.

For YEARS my wife has been telling me about the HVAC SNAFUs in the office building where she works. Sometimes it’s like a sauna and other times it’s like the tundra.

Yesterday, she had to come home because it was too hot to work. She brought a digital thermometer. It read 29 Celsius/85 Fahrenheit — if you live in the tropics that may not seem hot to you, but it’s pretty hot for Canada.

Clearly, someone somewhere isn’t doing their job!

I don’t know if it’s the technicians, building managers, or owners, but someone in the chain dgaf.

Things break and sometimes it takes time to fix them, that’s understandable. But after YEARS without improvement, something else is going on. If I were the owner of that building, I wouldn’t be able to sleep at night knowing that hundreds of people are annoyed daily because of me.

I’d hire HVAC experts, I’d install sensors, I’d do whatever it takes to figure it out and fix it. Most buildings don’t have rollercoaster temps, it’s a solved problem with known techniques — there’s no need for a big R&D effort.

But sadly, giving a crap and doing a competent job is not a given ¯\_(ツ)_/¯

💚 🥃 🚢⚓️ 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 and the private Discord.

If you only get one good idea per year, it’ll more than pay for itself!

---

🏦 💰 Liberty Capital 💳 💴

📩📬 Google’s Product Culture, Gmail Edition

It's surprising that after 20 years, the only way to display unread emails in Gmail on the desktop is by typing "is:unread" in the search box. This says a lot about Google's approach to product design and user experience.

Don’t get me wrong, it’s great that power users can do all kinds of cool filtering by using command line parameters in the search bar. But it shouldn’t be the only way.

But there should be an obvious button somewhere near the top for the 99% of users who will never figure out “is:unread”. Even power-users don’t want to type that every time they want to see their unread email.

Is anyone still working on improving Gmail UX or is this ˜2 billion user product in maintenance mode? What are Google’s thousands of designers and UX/UI experts doing all day? Is at least one of them working on improving GMail?

Microsoft is Training MAI-1, a Foundational Model to Compete with OpenAI, Google, and Anthropic 🤖💰💰💰

While Microsoft has ostensibly chosen OpenAI as its horse for the AI race, committing vast amounts of capex dollars and Azure compute credits to it, they have also made moves in the background to become less dependent on it.

Part of their contingency plan involved absorbing Inflection (technically not an acquisition, but they got the team and access to most of the IP & data…), and we can now see why:

The new model, internally referred to as MAI-1, is being overseen by Mustafa Suleyman, the ex-Google AI leader who most recently served as CEO of the AI startup Inflection before Microsoft hired the majority of the startup’s staff and paid $650 million for the rights to its intellectual property in March. But this is a Microsoft model, not one carried over from Inflection, although it may build on training data and other tech from the startup. It is separate from the models that Inflection previously released, according to two Microsoft employees with knowledge of the effort.

While Microsoft had already trained a bunch of smaller models (in Edition #493 I wrote about Phi-3 and how they were using synthetic data), this one will be pretty big:

MAI-1 will be far larger than any of the smaller, open source models that Microsoft has previously trained, meaning it will require more computing power and training data and will therefore be more expensive, according to the people. MAI-1 will have roughly 500 billion parameters, or settings that can be adjusted to determine what models learn during training. By comparison, OpenAI’s GPT-4 has more than 1 trillion parameters, while smaller open source models released by firms like Meta Platforms and Mistral have 70 billion parameters.

Like the MHz race between CPUs a few decades ago, not all parameters are created equal, so you can’t judge a model’s quality and capabilities purely based on that single number. Data quality matters a lot and multiple architectural considerations play a role (ie. being natively multi-modal, dense vs MoE, etc), along with RLHF and other training and fine-tuning steps.

But creating a 500bn parameters model would certainly put Microsoft in rarefied air along with a handful of companies (f.ex. Meta is training a version of Llama 3 with 400bn parameters).

👨‍🎨 Anthropic hires Mike Krieger, co-founder of Instagram, as Head of Product 🤖

As CPO, Krieger will oversee Anthropic’s product engineering, management and design efforts, Anthropic says, as the company works to expand its suite of AI apps and bring Claude, its generative AI technology, to a wider audience.

In addition to consumer-focused projects like the recently released Claude app for mobile, Krieger also will be responsible for Anthropic’s enterprise services, subscriptions and software.

Who knows if this will make a big difference to the direction of Anthropic’s products, but I’ve generally been a fan of Krieger’s work (even when it didn’t ultimately work out, like with his most recent startup, Artifact), so it’ll be interesting to keep an eye on.

Anthropic has a strong model with Claude 3 Opus. It was my default model for a few months because I thought it generally wrote better than GPT-4 Turbo. I only switched back when GPT-4o came out, both to test it out and because it was so much faster than Claude Opus.

Anthropic’s go-to-market is weak compared to Google, who has incredible distribution because of its existing products, and OpenAI, through their own products and Microsoft’s distribution pipes.

Now that AWS has a new leader and is expected to do more in the AI space, perhaps Anthropic can better leverage their partnership with them to improve distribution in the enterprise while they create a better standalone app to try to some more share away from ChatGPT on the consumer side? Maybe they should take a more opinionated product direction to stand out (like Perplexity)? 🤔

Founder Energy 💳

This type of founder customer service used to only happen on Twitter (the screenshot below is Substack Notes):

This reminds me of how Amazon executives — including Bezos in the past — answered customer service calls to be closer to customers and their problems. (I don’t know if they still do it, but I know they did in the past)

🤖💸 Musk’s xAI in talks with Oracle for $10bn GPU deal

Software used to be capital light. Not these days!

The numbers thrown around the AI space these days are head-spinning:

Elon Musk’s 1-year-old artificial intelligence startup, xAI, has been talking to Oracle executives about spending $10 billion to rent cloud servers from the company over a period of years [...]

It comes as xAI is finalizing a $6 billion equity funding round to help pay for cloud costs. The proposed cloud deal implies xAI would need to raise considerably more capital down the line.

Musk’s goal is to get to to Big Tech scale:

On a recent video call with potential and existing xAI investors, Musk said the company would spend a large portion of its capital on renting AI chips so it can improve the model behind Grok, an investor said. Musk implied that he was one of only a few executives who could raise enough capital to compete with the leading AI companies [...]

Last month, Musk said in a live audio chat on X that xAI will need 100,000 Nvidia H100 GPUs to train Grok 3.0. He said the company is currently training Grok 2.0 on around 20,000 H100 chips (Source)

While I wouldn’t bet against Musk’s ability to raise a lot of money or get a small team to match or do better than bigger rivals (ie. more bureaucratic), the next logical question is: What is xAI’s business model to even hope to get a return on all these billions?

Musk said he could sell access to Grok through his other companies, including Twitter/X and Tesla, but it may be hard to get a decent ROI if being competitive with Google and OpenAI means spending tens of billions.

🇨🇳 Chinese EVs & US Tariffs ✋🇺🇸

ChinaTalk has a good piece looking at the Chinese EV market and the recent US tariffs imposed on them:

• 📄 The Electric Curtain: Tariffs and the Roadblock to China’s EV Dominance

I’m very interested in EVs, but while I know a fair bit about the models available in North America, I know relatively little about the Chinese EV market.

If that’s also the case for you, let’s learn together:

The actual capabilities of Chinese EVs have started to converge. Models over ~$20-25k have similar features — long range, advanced infotainment and speaker systems, robust ADAS (advanced driver assistance systems) — in other words, things the everyday consumer cares about. [...]

What I found especially fascinating is how Chinese designers approached vehicle-design tradeoffs, especially in domestic Chinese markets: there’s much less emphasis on absolute performance (e.g. 0-60 mph times) for competitively priced models. This approach makes sense: most EV drivers live in urban environments, and many areas in China have tons of speed cameras (capping highway speeds at 62-75 mph/100-120 kmph). [...]

It’s important to note that Chinese brands didn’t do it alone. NIO, for example, hired designers from prominent Western auto giants to accelerate their own vehicle development. Even so, at this point, Chinese automakers are ahead, especially given that US and European automakers are delaying their EV efforts.

Since Chinese EV manufacturers “poach” Western designers, many people might assume that only the Chinese are stealing product ideas and intellectual property. But in the world of EVs, we’re actually starting to see the Chinese innovate (out of competitive necessity) and Western EV manufacturers starting to “borrow” stale Chinese designs. [...]

Longer-term support is where “legacy” Western, Japanese, and Korean automotive brands have the edge. Due to the competitive intensity of the Chinese EV industry, many brands are not expected to survive. [...] Bankruptcies or suspended operations of several Chinese EV brands have already occurred

Maybe instead of tariffs, other countries should tell Chinese companies that they can sell their EVs in their markets if they want, but to do so they have to establish a local manufacturing presence, do a joint-venture with a local company, and agree to IP transfers. Ha!

How do you like them apples now? 🍎

☕️ Peak Coffee in the U.S. was WWII 🥤🫖

Not what I would have guessed, but I suppose people used to make a lot of it at home and drink a ton at work and it didn’t matter that there wasn’t a Starbucks on every corner.

And during the war, well, gotta stay awake, victory depends on it!

But mostly, it looks like soft drinks have displaced coffee as an alternative caffeine & sugar delivery mechanism…

Meanwhile, tea is just honey-badgering through everything and minding its own business ¯\_(ツ)_/¯

h/t friend-of-the-show Ben Gilbert (🎙️🎧)

---

🧪🔬 Liberty Labs 🧬 🔭

Note the logarithmic scale!

🐜🐜🐜 ‘How to Build a $20bn Semiconductor Fab’ 💰💰💰

Great post by Brian Potter:

A large fab will have hundreds of thousands of square feet of cleanroom, and the facility might be spread over hundreds of acres. Building it requires tens of thousands of tons of structural steel, and hundreds of thousands of yards of concrete. Intel boasts that its fabs use twice the concrete as the Burj Khalifa, and five times the metal used in the Eiffel Tower.

Putting this material in place at the necessary level of precision requires thousands of specially trained construction workers. Intel’s new fab in Magdeburg is expected to require over 9,300 workers at its peak, and TSMC’s new fab being built in Arizona is using 12,000.

While the buildings are expensive, the tools are where most of the tomatoes go:

the fab exists to provide the necessary environment for the thousands of process tools to function, and it's these tools that are by far the most expensive part of building a new fab. Roughly 70-80% of the cost of a new fab will be the process tools that go in it.

I encourage you to read the whole thing. It’s almost a short book and acts as a primer on the semiconductor industry as a whole.

🧬 ➡👂🏻 Gene Therapy makes two deaf children able to hear at “normal levels”

Science ftw!

Regeneron Pharmaceuticals today announced that the investigational gene therapy DB-OTO improved hearing to normal levels in one child (dosed at 11 months of age) within 24 weeks, and initial hearing improvements were observed in a second child (dosed at 4 years of age) at a 6-week assessment. Both children were born with profound genetic deafness due to variants of the otoferlin gene, and the child dosed at 11 months of age is one of the youngest in the world to receive a gene therapy for genetic deafness.

The results are from an ongoing Phase 1/2 CHORD trial, so this hasn’t yet finished making its way through the regulatory snake, but hopefully no showstoppers or terrible side effects are discovered and this can help more victims of this genetic disease.

In the trial, both children received a single intracochlear injection of DB-OTO in one ear. [...]

At baseline, both participants had no behavioral (PTA) or electrophysiological (ABR) responses at maximum sound levels (≥100 dB). Following treatment with DB-OTO, both children showed auditory responses at the first efficacy assessment of 4 weeks.

Even the experts are surprised:

“The opportunity of providing the full complexity and spectrum of sound in children born with profound genetic deafness is a phenomenon I did not expect to see in my lifetime,” said Lawrence R. Lustig, M.D., Chair of Columbia University’s Department of Otolaryngology - Head & Neck Surgery and a clinical trial investigator.

Source.

🔌⚛️ ‘Our mission is to achieve 3 cents/kWh with nuclear’

Friend-of-the-show Matt Loszak talks about what Aalo Atomics is building and why.

🔌⚛️ 🔧 Constellation Energy is Adding a Phantom Nuclear Plant… by Upgrading Existing Reactors

I wish Constellation Energy would build new reactors instead of just milking the existing fleet that was built decades ago (they canceled earlier projects due to temporarily cheap natural gas… and now the US wishes it had more LNG to export to its allies).

But they seem like good operators who know how to squeeze more out of the existing plants:

The company is already spending $800 million to add new equipment that will boost output by 135 megawatts at two Illinois nuclear sites, and Chief Executive Officer Joe Dominguez is planning to pursue similar projects at other facilities. This so-called uprate process will eventually add about 1,000 megawatts of capacity at the company’s 14 nuclear plants, he said Thursday.

1 gigawatt would be the equivalent of one big reactor.

Electricity demand is expected to increase thanks to EVs and data centers, so I’m sure this ‘phantom reactor’ won’t have trouble finding buyers for its clean electricity.

---

🎨 🎭 Liberty Studio 👩‍🎨 🎥

🎥🪱🎶 Hans Zimmer on creating the music for Dune: Part 2

The more I hear Zimmer talk about music, the more I like him. He’s having fun and loving the creative process. We should all be more like that.