Reflection on my first "real" 100%+er

Last week, it was at $349.00!

INTERNAL CULTURE

I've heard that Google has one of the best internal systems for innovation. Not sure, but I think this was credited to Brin and Page? These have held up to the test of time, and the OpEx spent on random SWEs building this internal tooling has paid off. Google (especially GDM) has a unique position to enable genuine research effort -- credit, of course, goes to Demis on this. Hiring research scientists for $400k/yr for years helped them.

WAYMO

I haven't entirely looked into it, but there is no way Uber and Lyft survive. People can say whatever distribution argument they want -- they're just wrong. Google has better distribution... it's called Google Maps. No one talks about it, but Lyft is probably paying Google an insane amount for the integration. The day Google pulls Lyft out and begins aggressively pumping Waymo, Lyft's done. Then, you have this margin compression pressure from Tesla which also hurts Uber/Lyft. Everyone I know prefers Waymo over human drivers. It's an easy, simple change for most. Sundar or whoever just helped Waymo add a +$16B into their bank account, which is also great for them.

OK - so why are people unsure about Waymo??? It's really only a perception of Elon/Tesla fanboys. The idea: predict next token (angle + speed) using ViT + some custom archs. Why? It's cheaper than LIDAR + rule-based efforts. I agree, but is Waymo somehow behind on this? Like, Waymo is publicly saying they are also working on this, lol. It's a natural progression of the system. Google can subsidize Waymo for as long as necessary to either get Lidar tech down to $0, or in the case that Tesla bulls are right and you REALLY don't need them, Waymo flips a switch and now they are running the same models as Tesla.

Also, as an aside, why aren't they using flow matching? Seems like a great use-case of continuous prediction of the vehicle as a variable to then use in RL / sims. Or stochastic differential equation model stuff. Or nODEs, but those are slow/probably not feasible.

LIFE SCIENCES

This is becoming increasingly clear to the markets and folks who weren't paying attention to GDM before AlphaFold2. However, most still don't know about what Iso + Calico + Verily are up to. Overall, it's a bet that they'll be able to generate some value once AI can reliably create value in bio, sciences, and health. This could take anywhere from 3 to 15 years, depending on who you ask. Do we need a virtual cell? Or is lead optimization enough of a cost saver? etc.

Iso - owned wholly by Alphabet but did raise $600m from Thrive et al. Building maybe the first legit platform techbio that will actually succeed? Sell to every pharma. Main question is the moat? Academia seems to catch up 1-2 years later and then pharmas can clone the models internally and FT on internal data. Talent alone isn't enough either b/c of Chai, Boltz, etc. Iso is basically just a — "Hey, let's put our money into the one company who is most likely to succeed on AI drug dev tasks and generate value."

Calico - majority owned by Alphabet. Aging is going to be big in the future. Many in biotech are most excited about aging as the next frontier.

Verily - majority owned by Alphabet. I feel like Verily could've been a great CRO. I think folks overestimated how hard becoming deeply embedded in pharmas is -- might've been better if these co's instead became "AI-native" and "vertically integrated" and then competed, and eventually grew into an IQVIA like player.

GV + CAPITAL G

These speak for themselves. But a lot of people I talk to don't know how good they really are. Anthropic??? SpaceX??? Stripe??? Oh, and my favorite, Blue Bottle Coffee!

BIG FISH: GDM

It was clear to Sam, Greg, Ilya, and Musk that Deepmind was going to get to AGI first. If you don't know the story, read Walter Isaac's description of it in Elon Musk. I do not think this has changed. Let's talk through a couple scenarios.

LLMs will get commoditized! OK, but what about a super intelligent LLM? If it has such a capability, is it in the interest of any frontier lab (incl. Chinese) to release this? No! You would either use it yourself if you have enough chips (via something like Thrive Holdings x OpenAI), or you sell it for $100/hr, to near-infinite demand. There will be such high demand, that it literally cannot become a race to the bottom. I think one of the best data points to point at here is OpenAI's IMO model. They could make a lot of money on this model by selling it. Yet, they choose not to. Regardless, GDM, OpenAI, etc. will all make money. Race to the bottom pricing is not possible because opportunity cost of selling a token to a user is higher than using this token to improve the model.

We don't get AGI because LLMs are parrots! Well, then we plateau at PhD-level intelligence in particular domains (anyone who disagrees with this has probably never used something like Opus 4.5). Classic RL on verifiable tasks = coding solved, accounting solved, law solved etc. If GDM can automate 80% of white-collar verifiable tasks, GCP surpasses Azure and AWS easily (and Google adds $1T to market cap). This is basically the bear scenario here -- AI leads to like a +25% for $GOOGL and that's it. But, even if LLMs are parrots, how long until the next paradigm? World models are coming. You don't like world models? OK, give it another 10y. They won't stop, and there will be a continued influx of AI talent (both living in Mountainview/London and in data centers :))

OK - you get the point. AGI is coming and, even if it doesn't, Google wins. But it will.

I'm friends with some hardware folks. These people revere GDM and the TPU teams. They describe it as if no one is close. They describe it like impending doom for Nvidia. Thus, the best way to hedge against your neoclouds x Nvidia is just buying Google! I was talking about this before it was hot. Just imagine a TPU v9... at scale. This thing is going to be a beast and I just absolutely cannot wait for Gemini 5. And quantized Gemini 5 Flash being available for free to the public -- so they finally realize how powerful these models are.

I'm honestly not an authority on hardware, and do not *really* understand what the ecosystem is like. Of course I know the talking points--custom ASICs, CUDA moat, etc. However, I have high faith that TPUs are superior and will lead to recursive self-improvement on the models -> chips, which will propel Google far ahead.

GOLDEN GOOSE: SEARCH

As much as I love to base my investments on technology and product theses, you cannot be serious unless you talk about search. So, let's be rigorous here and discuss search. For the last three years, the loudest bear case on Wall Street was that ChatGPT and Perplexity (hah) were going to eat search. "The margins will collapse due to LLM compute costs!

GOOGLE GLASS?

So, back to product and technology. I'm a firm believer that the phone is an extension of the mind. Taxi drivers have larger posterior hippocampi (hippocampuses? idk) because of how much they need to remember locations. Axons get myelinated, and circuitry grows. Google Maps comes along, and now these posterior hippocampi shrink. So yes, the phone is literally a piece of your brain. There's a X second delay to open it, and it doesn't "feel" like it's part of your mind, but it is.

You've probably heard of BCIs already. This is the next step in human evolution. Real-time interface to the internet. There is absolutely nothing wrong with "shrinking" your X part of your brain, so long as you are replacing it. You may not think this is true, but what are you doing when you look at Google Maps? Well, you're following instructions, and developing better reactions to screen stimuli. What's called your caudate nucleus actually GROWS.

However, real-time BCIs with internet are sci-fi right now. I can say that as someone who studies neuroscience. We need universal decoder + stimulation and, of course, the billion electrical engineering + data compression innovations that go with that. We are many, many years away. Glasses are the obvious, best form factor. Apple, Google, Meta, and the brilliant product folks at startups know this. It's the next iPhone (although Apple might try to convince you you need both). I ultimately don't think Google will win here, however, even a Pixel-like outcome is great, and it is awesome they have been innovating in hardware. Back to the company culture paragraph!

YOUTUBE! CPMs!

The world is getting sloppified from AI and addictive short-form content. I don't think this is going to stop, unfortunately. This is genuinely detrimental to society, but no one cares about it. BUT! Long-form is still growing as folks shift from TV, movies, shows, etc. onto YouTube. YouTube is the best platform that exists, and somehow is winning short-form in many countries. Especially developing countries. As these countries develop, their CPMs shoot up.

Have you ever seen YouTube channels like T-Series, or random gamers or comedians from India who regularly pull 10M+ views on each video? There is massive traffic, and currently it is not being monetized the same way that content American or Australian audiences watch. But, the CPMs will rise, and with it, YouTube will print far more revenue than Netflix. People currently call and value YouTube as a Netflix within Google. But Netflix will eventually become a mini-YouTube, and then and remnant of the past.

MUSIC, MEDIA, CREATIVES