Tag Archives: ChatGPT

Book review: Kevin Roose – The AGI Chronicles: The Inside Story of the Race to Create an Artificial Superintelligence

This book has been published quite recently and basically re-tells the last 10 years of developments with a focus on LLMs and the few dozen people, most located in the area of San Francisco, who are the main players in this “game”. The companies that are covered are mostly OpenAI, Google and Anthropic.

It does not really cover the 50 or 60 years of AI research before that.

The author, Kevin Roose is a relatively well known SF based tech journalist who seems to have pretty good access to many of the main players. He also hosts a very interesting podcast called “Machine Gods” together with Casey Newton

I found it quite an interesting read. A few things that stood out to me were the following:

  1. If you want to pinpoint one guy who has maybe contributed the most to develop ChatGPT then Ilya Sutskever would be my choice. He actually had the idea to combine the Transformer logic developed by Google with a neural network trained on speech. A rather well known artifact is that actually Google invented this but didn’t really understand its significance. The flip side is that Sutskever was not very successful in kicking out Sam Altman during the OpenAI palace coup.
  2. What I find the most perplexing aspect of this all is, that even the super experts like Sutskever don’t really understand how these models suddenly develop capabilities that they were not explicitly trained for. Andrej Karpathy once described it as “summoning a ghost from the machine”. I guess that also might be one of the reasons why some of the people involved developed quasi religious feelings about these models.
  3. However, what the experts understood early on is that more compute makes these models more capable. Later on, they also discovered that more data for training has the same effect. So the “more compute” mantra has been there since quite some time.
  4. Another really fascinating aspect of the book is, how small the group of people actually is who is behind this now world changing technology. Essentially, it is a few hundred people who mostly know each other, maybe even hated each other and often have lived in the same house for some time as “AI researcher houses” seem to have been a big thing some years ago,
  5. Equally interesting is how the author describes that Google almost lost the race. As many big companies there was a lot of infighting at Google. The main issue seems to have been that Google researchers are mostly working for themselves with no real coordination. So for quite some time, they lacked the computational resources to improve the LLMs because they could not pool enough internal capacity.  There was Google Brain and Deep Mind who both worked on certain aspects of AI, but didn’t fully understand the significance of Chatbots.

    Google management was also cautious because they feared for their cash cow, the  search business. According to the book, the return of founder Sergey Brin changed things dramatically. Although he officially only worked as a programmer, one phone call from him would suddenly make things happen. But I think Google was really close to becoming another Microsoft (which almost missed the internet and fully missed mobile)….
  6. The book also shines a light on the two main “faiths” in the AI race: Amodei & Co who are more on the Doomer side and campaigning for “pacing the frontier” and the “Accelerationists” who claim that this technology needs to be pushed forward as quickly as possible to create “heaven on earth” and 10% GDP growth p.a.. The accelerationist were headed (among others) by none other than Jensen Huang, the just fired CEO of Nvidia. With the results of the midterms and Jensen’s firing, one could expect that now the more cautious people prevail at least for some time.
  7. The book doesn’t try to predict the future, which I find very good. It is clearly a chronicle and doesn’t try to predict something that cannot be predicted.

Overall, I think the book allows a unique view into the ecosystem of AI experts, their beliefs and also their sometimes significant personal shortcomings.

Tf you want to read this book, now is probably the time , as already a lot of stuff is happening and history is moving fast these days.


The chapter of that history will be written most likely by the end of this week when Anthropic goes public.

What a time to be alive.

Visibility as a Creator/Writer on LLMs – A test and some thoughts

Visibility on LLMs

Management summary:

This post does not offer any actionable investment content. Rather I wanted to find out if my blog is visible on the various LLMs and if I want to be visible. I would be very interested in how fellow “creators” think about this and how they approach this topic.

Visibility of Value and Opportunity on different LLMs

Just out of interest, I asked several LLMs about the 5 best Investment blogs for European stocks. The results were quite interesting.

Google Gemini for instance distinguishes significantly in which language one asks and which model you use. A German language prompt gives a very different answer (mostly German language Blogs) than an English prompt and “fast” mode gives very different results from “thinking” mode.

Here are the 5 top blogs in Fast mode for the German prompt: (“Welches sind die 5 besten Investment Blogs für Europäische Aktien, insbes. Nebenwerte ? “):

And here the results for the same prompt in “thinking Modus”:

There is some overlap and I am on both of the lists, which is great, but still interesting.

A few days earlier I tried a slightly different prompt (“Nenne mir bitte die 5 besten Investment Blogs die sich mit Europäischen Aktien beschäftigen. “)

And I got very different results:

What is also interesting is that Gemini doesn’t look for Substacks when I ask for blogs. Asking specifically for Substacks, gives once again different top 5 for the fast and thinking model, but the V&O Substack does not appear when asking for Substacks.

When I ask Gemini in English for blogs, I get the following result for “fast” mode:

In Thinking mode, this is the output:

So I show up in both, but the other 4 are different.

Overall it is quite interesting that asking in German language automatically selects mostly German blogs and how much the results differ from fast to thinking mode.

Of course, different LLMs give different answers. The very same German prompt from above  gives this result overview in ChatGPT:

The English prompt gives the following result:

ChatGPT interestingly does not care too much in which language you ask, the overlap is higher than for Gemini. But it has remembered my 10 factor Scoring model and without asking has somehow mixed that into the decision.

Claude interestingly doesn’t seem to know my blog at all. I have to say I am disappointed 😉

The LLMs know Value and Opportunity

So after putting out content for 15 years, Gemini and ChatGPT LLMs clearly know about my blog, but it is really interesting how differently they answer to the very same questions. Also that language plays such a role for the results is kind of interesting for me.

Interestingly, if I use the normal Google search, my blog is not visible at all, at least not on the first 10 pages, irrespective of what kind of searches I do. This mirrors  a little bit the traffic statistics form my WordPress overview where Google as a source for traffic more or les disappeared a few years ago. Only when I ask for a certain analysis, for instance Eurokai specifically on the Value & Opportunity blog, I see my blog in the results. Otherwise no chance.

I have to admit that I have also become quite lazy to add a lot of Keywords etc but in general, Google search as such seems not to be “my friend” anymore. Some years ago, especially the more general articles received significant traffic, even years after I wrote them, but that has gone totally away.

How to optimize for LLM visibility  ?

I feel very lucky that I don’t have to optimize for traffic, otherwise I could imagine that trying to optimize LLMs is not so easy. I have briefly researched the topic and it seems that for now, LLMs seem to emphasize a longer track record and credibility.

One of the nice things is that one can ask the LLm to explain. Google Gemini’s answer is quite flattering I have to admit:

If I wanted to make more advertising for my work, I would basically copy& paste that answer.

Of course, I also wanted to know why I don’t appear on Claude’s list. This is what Claude tells me:

Typically for an LLM, it apologizes. What I find interesting is that Claude indeed seems to start looking in high traffic locations and then doesn’t go much further.

Do you actually want to be visible to LLMs ?

One question one has to ask is of course as a writer & creator: Do you want to be (fully) visible to LLMs or not ?

Despite my visibility on Gemini and ChatGPT, the LLMs do not refer a lot of traffic back to the site. I can see Gemini with a little traffic and ChatGPT with no referrals at all. So they know about the blog, but they don’t refer a lot of people to the blog. Maybe the answers are already good enough if my content gets shown. Outside my Email list, most traffic still comes from Google search and TwiX.

If you want to monetize your content directly, it is clearly not good when LLMs can read your stuff and summarize it perfectly. I was for instance quite astonished when a TwiX user asked Grok to summarize my Biontech post in TwiX and Grok did so with a pretty decent summary.

On the other hand it seems that at least for Gemini and ChatGPT, you need to show them your content in order to get recognized. I guess a good compromise could be to show some of the content so that the AI can learn about what one writes but then keep newer stuff behind a paywall or so.

Another strategy would be, not to share anything on the web in order to protect one’s “intellectual property”. As for now, the LLMs don’t give a lot of traffic back, so why should you be visible at all ?

In my case, I am lucky that (so far) I can monetize my content very indirectly.

For me, the main payoff comes through constructive feedback and, every now and then a nice email from a reader or even better, some personal contact and someone says “I read your blogs for x years and really like it”.

My other goal is also“make the world a little bit of a better place” by maybe teaching some people how to “invest” instead of just “gambling” blindly and help them to hopefully better secure their financial future. For this goal, getting my content “indirectly” distributed through LLMs is clearly helpful.

If someone asks if xyz-Shitco is a good investment and somehow in Gemini’s neural net it identifies a  “red flag” that it has maybe learned through my posts, this could be a very powerful “amplifier”. But this is clearly hard to measure.

Summary:

For now, I am quite flattered, that 2 out of 3 LLMs find my content good enough to put me into the Top 5 European Small Cap blogs. That is clearly niceclearly a nice feedback. 

Most of all, I feel very lucky that I don’t have to directly monetize my content. I think this will be less straightforwardstraight forward than in the “search machine age”. There will be some solutions for sure but I guess “cause and effect” might be less linear than in the old times.

I would be very interested in how fellow “creators” think about this and how they approach this topic.

Some thoughts on Vibe Coding, SaaS vs AI (10 Moats) and Guidewire Software

Executive Summary:

This post does not contain any actionable investment advice but rather some personal ramblings on Vibe coding and the attempt to analyze a specific Software company (Guidewire) according to a Template of 10 Moats for Software companies and their vulnerability to the AI threat.

Introduction:

My track record as a Software investor is to put it mildly, very poor. My best Software Investment so far is Chapters Group which I bought as a net-net before it even became a VMS Serial Acquirer. My blog and portfolio archive also tell me that I sold Microsoft in 2011 at ~25$ per share with a 4% gain because I thought that the Office products had no future. So please take everything I say about Software with a grain of salt or even better, just ignore it.

I do have a background in Software development. Although I would not call it Software development but “Code butchering”. It started as a teenager on a C64 with Basic and Assembler and ended in the late 1990s with Cobol/PLSQL working for a large US Consulting Company (yes, I was young and needed the money). Knowing the speed of financial institutions, I would not be surprised if some of my Spaghetti code would still be running somewhere….

Why am I saying this ? Because of course, Software stocks have been doing quite poorly over the past weeks/months. In addition, I also had the opportunity to play around with Claude Code first hand. 

Biontech: Checking in once again & using LLMs to value the Pipeline

Disclaimer: This is not Investment advice. PLEASE DO YOU OWN RESEARCH:

Management summary (Spoiler):

After selling Biontechs a few years ago, I reviewed Biontech once again by using LLMs to evaluate the development pipeline. Although this was an extremely interesting exercise, it is not an investment for me for the time being.

Background:

Biontech is a company I have owned in the past and written about. It became famous because they were the first to develop a MRNA based vaccine against Covid which they sold worldwide together with Pfizer.

My initial write-up can be found here, a follow up here.

Read more

Random ramblings on AI

You didn’t ask for it but you get it nevertheless: Some random thoughts on various aspects of Artificial Intelligence. Spoiler: No actionable insights (I think).

Gemini 3.0. vs. Nvidia

Google Gemini 3.0 seems to be a really good model. I am currently using it with my prompts and it seems a little bit better but not that much.  NotebookLM seems to have improved a lot.

However, according to various sources, the model was trained and runs exclusively on Google TPU chips. The Nvidia Bulls keep saying that Nvidia has such a large advantage including their software, that those ultrafat margins will persist for many years as there is no alternative. I am not so sure about this. 

This is the EBIT margin development of NVIDIA since 2002:

Read more

Was this the “Karpathy Moment” for the AI Industry ?

Who is Andrej Karpathy ? Born in Slovakia and despite only 38 years young/old, he is already an AI “Veteran” having initially studied under AI Godfather Geoff Hinton in Canada, did internships at Google Brain and DeepMind, Co-founded OpenAI, was leading AI at Tesla, went back to OpenAi and now is focusing on teaching AI to everyone who would listen.

Since I discovered his Youtube educational Videos, I am following him because when he speaks about something, there is always a lot to learn. 

Yesterday, he did a 2 hour interview with maybe the best current “AI Podcaster” Dwarkesh Patel. Those two hours are quite dense and I had to use Gemini in parallel to understand some of the stuff, but at least on my Twitter timeline, it raised quite a “storm in the teacup” among AI “experts.

Here are some of his main talking points (as far as I understood them):

Read more