Some perspective on AI

Investment Analysis Clubs Saul’s Investing Discussions
1

I think this is a good perspective to keep in mind as we evaluate our investments in and around AI technology:

Chamath Palihapitiya:
“We are in the first inning of what probably should be an enormous tectonic shift in technology. I think whoever wins in the first inning usually isn’t the one that’s winning by the ninth inning. … The future is unknown and the more disruptive the technology is the more entropy there is, which means there’s going to be more changes, not less. I would just look at Search as an example, I would look at Social Networking as an example. When you look 20 years later, the people who captured all the value were not the ones at the beginning who everybody thought was going to win.”

That was said in relation to OpenAI possibly going IPO.

I think this applies to the software side of AI more than the hardware side. Yeah, the hardware side will eventually slow down, but the uncertainty Chamath is talking about is on the software side - Apps, Platforms, Business models, etc. That really is still an unknown and the lessons on early search (Yahoo, Alta Vista, Ask Jeeves, etc.) are as strong as for early social networking (MySpace, Friendster, AOL, etc.).

21 Likes
2

A more important distinction is between AI suppliers and AI users. Compare Google (chip user) vs. Intel (chip maker), for example:

GOOG vs. INTC from GOOG’s IPO

goog
goog720×510 25.5 KB

but Intel had a great run early on
INTC vs. GOOG from INTC’s IPO

intc
intc720×510 33.1 KB

One can clearly see the hype cycle playing out

  • Innovation Trigger
  • Peak of Inflated Expectations
  • Trough of Disillusionment
  • Slope of Enlightenment
  • Plateau of Productivity

Nvidia might be at the Peak of Inflated Expectations but has the advantage over intel in that it does not have the capital intensive Fabulous Fabs.

My bet on AI would be on Tesla, monetizing AI via FSD, Optimus robot, and RoboTaxis.

Denny Schlesinger

29 Likes
3

Your example fits my superior characterization of software versus hardware. The category of “AI users” is too broad to be useful, encompassing not just Google as a “chip user,” but also on-premise servers, cloud providers, software platforms, software applications, and then, of course, application users.

As we saw with the internet broadly, and later with the rise of SaaS, different technologies and business models will become profitable and mature on different timelines. It doesn’t help me at all to say there are chip makers and chip users. Is SMCI a “chip user” the same way that OpenAI with ChatGPT is a “chip user,” the same way that someone at work using MS Co-Pilot is a “chip user?” Not at all. Saying both Apple and Instagram are both “chip users” wouldn’t have been helpful either. Instagram needed a mobile computing platform for its software tech to be successful.

Understanding the progression of these different technologies and business models will be key to making profits off them in the stock market.

Fabs were a key advantage for Intel back in the day when they led in Fab technology. Fabs only became a disadvantage for Intel when they stopped adopting the latest and greatest Fab tech. TSMC, which is almost solely a Fab company, invested heavily in EUV and made the technology work, giving them years of headstart over Intel, and so they did (and are doing) quite well. That’s two big waves (low power chips for mobile, and EUV tech for AI chips) that Intel missed.

Back to my original point, outside of Nvidia chips, we don’t yet know whether on-premise will eventually be a smaller market than cloud for AI processing. We don’t yet know what software platforms will become dominant, what applications will be written to run on those platforms, nor what the big growth markets for AI usage will be - at least on a timeline upon which we can profit. And if AI is anything like what we’ve seen from tech in the last few decades, the early first-movers may not be the eventual big winners.

23 Likes
4

It ISN’T! Not that I’m an AI expert but I started out as a computer programmer. My best source for AI is Andrej Karpathy, the former head of AI at Tesla. There is a paradigm shift between the heuristic programming methodology I was brought up on and how neural networks generate AI. Heuristics is basically boolean thinking while neural networks are brute force, the more data and the bigger, more powerful computer will generate the best results. Andrej Karpathy said that there is a direct correlation between data and computer power and quality of output. What that means is that AI productivity is highly correlated with capital availability. That gives the rich players and the big data collectors the advantage. Google has tons of data. Tesla has tons of driving data from millions of cars on the road. Tesla has FSD, RoboTaxis, and the Optimus robot to monetize the AI.

I think the key will be who can best monetize the AI. Nvidia is monetizing AI as a chip supplier while Tesla is/will monetize it as an AI user. How can companies monetize Large Language Models?

https://karpathy.ai

Denny Schlesinger

My musings on the brain and on AI…

23 Likes
5

I’m talking business models and profits, not tech. Of course the tech is different, but as well all know, better tech does not always (or even mostly) equate to better profits.

It’s neither here nor there, but I’ll just note that Tesla’s biggest advances in FSD were achieved well after Karpathy left Tesla.

Again, chip supplier/user is not a useful characterization. SMCI is not a chip supplier, yet it’s making money on hardware. Microsoft is making money not just as an AI user, but as an AI service provider. Amazon is providing AI as a service. OpenAI may go IPO. Snowflake and Databricks are battling out what AI means for databases. You can’t say everyone except Nvidia/AMD/Intel/SMCI/Dell/HPE are “AI Users” and have that materially inform your investing decisions.

AI tech is different than the internet or cloud booms, but the same product infrastructure and business model characteristics will apply. Who will make money on hardware and infrastructure? Who will provide computing services, and what will the mix of cloud/on-premise be? Who will establish dominant positions in software development and run-time (inference) platforms? And who will release the most compelling AI applications and products?

Not a company I recommend, but Adobe has brought AI features into its products, using cloud-based AI calculations to remove people or things from a photo and create a seamless background fill-in, and, of course, generating completely new images from images you supply with descriptive of text of what you’d like it to generate.

Microsoft has AI products like CoPilot that do everything from help programmers write code to perform functions in Excel, like “Create a bar graph showing the sales growth between Q2 and Q3” or “Add a new column showing the percentage difference between column A and column C.”

What’s the value of AI applications that help companies manage supply costs, assist in making expansion decisions, or suggesting companies for investors to invest in? The list is endless, and I suspect there will be whole new businesses, business models, and industries that sprout up around AI tech.

29 Likes
6

The most BG2 podcast delves into AI Costs vs Revenue:

(starts at 22:07 in).

It’s interesting to me that they are both of two minds. That AI is the coming thing and so investments in it are important, but also that perhaps the AI hype is too great and these early investments are premature at best. That some of the claims are being scaled back (coding benefit down to 20% from 50% claimed, etc.), and that the timing of new releases keeps getting delayed. Maybe benefits will be cost reduction rather than new revenue. Some discussion around how it’s fine for big companies like Meta and Microsoft to make big investments (and maybe training clusters get repurposed as inference), but harder to justify for venture-backed businesses (that have raised billions).

No conclusions to be found, but the discussion is good as it brings up a number of potential issues, backed with some basic data.

13 Likes
7

Off subject but valuable – personal experience with calcium scans mentioned in Smorgasord’s link.

Twenty three years ago, I read about calcium scans. My cardiologist did not have me on any medication for blood pressure or cholesterol control meds. I suggested I get a scan. He said it was a waste of money, I do not recommend this scan for you. I had the scan anyway. It turned out I had more plaque than 90% of males my age. My cardiologist said, I do not know what possessed you to get a scan, but the results mandate we lower your blood pressure and get you on a cholesterol lowing medication immediately. I had not had a heart attack prior and still have not had one.

Graydrake

15 Likes
8

Things that did not exist cannot get cheaper! Growth happens because new, large, and highly profitable markets are created.

Denny Schlesinger

6 Likes
9

Another data point on AI:

Larry Ellison essentially saying the demand for AI training is only going to grow from here. Cites things like specialized AI for biopsy interpretation - a specialized training on millions of biopsy slides as indicative of “a lot of very, very specialized models.”

He goes on to talk about an “ongoing battle for technical supremacy” that’s going to last for 5 years at least, probably more like 10.

“There’s no slowdown or shift coming.”

40 Likes
10

What is happening is AI is complicated. It takes lots of organization and data and computing power. Not easy things to be done overnight. Much less to be put into mainstream use. Putting together a website is a breeze compared to mainstreaming AI. So it will take time. The hyperscalers and software companies know the possibilities and they are spending more than $100 billion to meet initial demand as it scales.

In three years get back to me.

14 Likes
11

Three years is long. How about Nvidia’s Blackwell is going to explode the market again. Demand will be off the charts again and it will enable even more AI training

18 Likes