397: Silicon Valley Bank, Crowdstrike Q4, Stack Fallacy, Google PaLM-E Robot, DuckDuckGo's Wikipedia LLM, Arena FPS, and Silo

"The fewer things you do, the more important taste becomes."

Do unto those downstream as you would have those upstream do unto you. —Wendell Berry

Portrait os Steve Jobs I made with StableDiffusion 💚 🥃

🛀💭 The fewer things you do, the more important taste becomes.

I’m thinking of “taste” in a fairly broad sense here. A vague gestalt that encompasses attention to detail (up to perfectionism, at times), discernment, giving a damn, a profound understanding of design, a holistic view of products/projects, knowing which problems to solve, selecting the right people for the job, etc.

Think of Apple in the years following Steve Jobs’ return 🍎, in the late 1990s and early 2000s when he saved the company from almost-certain bankruptcy, compared to Dell around the same time.

Part of Jobs’ strategy was to focus the product line on four products: two desktops and two laptops (one of each for consumers and professionals). That’s it. All your eggs are in four baskets, so it matters tremendously that you get it right for each.

Dell’s approach — which was also very successful — was to make everything for everyone, throw all kinds of stuff against the wall in case something stuck, allow endless customization options, and have lots of very similar products with very similar names (the GZD2000SA vs the GZD2001SB or whatever).

At the core of the strategies is relying on your own taste vs relying on the customer’s taste. One can lead to differentiation, while the other should aim for low-cost production because that way leads to commoditization.

Why is taste so hard to copy or acquire on purpose, though? A question for another day!

🕰️⏰🥱🥴 Can we all agree that changing the time twice a year is a bad idea. Why fight against biology? Let’s pick something and stick to it. Thanks for coming to my TED Talk.

Oh, and it’s 1000x worse with young kids, it takes them a week each time to get back to normal….

🤖🎮🧜🏻‍♀️🧟 I don’t know if anyone is working on this — probably, I don’t think it’s that original an idea — but here’s something I’d like to see in video games (particularly RPGs):

Using large language models (LLMs) to create non-playing characters (NPCs) that you can have real conversations with, ask questions other than 2-3 pre-written choices, and have them react realistically when you say or do something (especially something unusual — it’s always funny that in games, you can do the weirdest, most outrageous things right next to NPCs and they just look on with dead eyes, clearly not understanding what is going on).

One of the things that breaks immersion in open-world games is that NPCs tend to have fairly limited scripts, and they can quickly be made to talk in circles and repeat themselves like broken records.

For games in which you are on ‘rails’, game devs can predict exactly what will happen at any time, but if you have more freedom, it’s almost impossible to prepare dialogue for every situation.

It may soon be easier to have a LLM that has been trained on a certain amount of dialogue, story elements, and game lore, and just have characters in the world ‘improvise’ dialogue in real-time when something doesn’t quite fit their script.

With text-to-voice technology improving, we should soon be able to have them speak dialogue with realistic inflection and emotion.

Games will feel a lot less like Potemkin villages.

To do this, in theory, you shouldn’t need a giant LLM that contains billions and billions of parameters since you don’t ask NPC to code in python or rewrite Eminem songs in the style of Shakespeare. Much narrower models limited to one language and in-game knowledge could keep hardware requirements more manageable.

🚀⭐️ TIL: 'Ad Astra per Aspera'

It's Latin for: "to the stars through hardships"

I kinda like it ✨

(also: it's the motto of Kansas, among other places)

---

🏦 💰 Liberty Capital 💳 💴

‘Why Big Companies Keep Failing: The Stack Fallacy’ 🧗‍♂️