AI minus the BS (Part 1)
Why investors should take this seriously
Unless you’ve been living under a rock (which is totally possible, no judgements here), you’ve probably heard of this thing called ChatGPT, and perhaps even text-to-image tools such as Midjourney.
You might even be thinking:
“Yes, ValuePunks. I’m sick of hearing about this thing. Why are you jumping on this bandwagon too?”
Well, give us a chance.
So far, we’ve seen three main types of reactions.
Hype: ‘Computers are sentient! Look at all the things it can do! This is going to change the world! We are going to cure every disease and invent calorie-free pizza!’
Fear: ‘Computers are sentient! Look at all the things it can do! It’s over for humans. We are going to have 20% unemployment by next year!’
Skepticism: ‘Eh, all this hype feels just like all the other things that were supposed to change the world. It’s just a toy.’ Most value investors are probably in this bucket!
These reactions are to be expected when something new and shiny comes along. At the same time, they are not particularly helpful in understanding what is happening, and translating the work into making sound investment decisions.
We’ve spent some time trying to take a measured look at this breakthrough, without hype, fear and dismissiveness. We understand that any technology that comes with the claim “this will change the world” has a low base rate of doing so. Yet, we believe that this is a topic that is still worth your time and attention.
What follows is a snapshot of our progress, in three parts.
Part 1 will explain why this deserves your attention and explore ChatGPT’s capabilities.
Part 2 will explore the basics of how GPT models work and their key limitations.
Part 3 will explore the implications of advancements in AI. We will attempt to develop a framework for assessing these changes.
We hope that this series of articles will spark many conversations.
Let’s dive in.
Why should ChatGPT be taken seriously?
On November 30th, 2022, OpenAI released ChatGPT1 to the world, which is a chatbot powered by the GPT-3.5 large language model. ChatGPT featured capabilities that had never been seen before.
But wait - isn’t this just another Silicon Valley-hype-cycle-led-toy that promises to change everything but in practice, doesn’t really do much?
Despite the hype, we think that ChatGPT and developments in the field of AI need to be taken seriously for five reasons:
ChatGPT represents a culmination of multiple genuine technology breakthroughs. The combination of improvements in hardware capabilities, and architectural improvements have led to computers gaining abilities that were previously thought impossible:
It can process natural languages. It can now “read and write” in multiple languages.
It can “understand” your questions, and give you relevant answers. It can also stay on topic when carrying out a series of queries.
It can generalize, answering questions that it was not specifically trained to answer.2
The pace of development is fast - both at the platform level and in the application of the technology.
Platform level: ChatGPT was based on GPT-3.5. The newest model, GPT-4, was released on March 14th and came less than four months after the release of 3.5, with significant improvements.3
Application: ChatGPT is being given new abilities. The original public version was walled off, with dated information. The latest version can now browse the web, execute code that it wrote, “see” images, as well as interact with different sites via plugins.
The pace of adoption is also likely to be fast. Unlike the internet or the smartphone, ChatGPT deployment requires no infrastructure build-out4 or hardware purchase. It is also being built into existing tools as an upgrade. As a result, the uptake of this breakthrough technology is likely to be quick.
For instance, many software engineers have already started to utilize generative AI tools such as Github Co-Pilot X to help write code faster.
“The barriers of adoption from technology point of view have been significantly reduced. In fact, a lot of the code that is required to integrate the ChatGPT with your applications can be generated by ChatGPT itself.”
-Former Director, Automation Anywhere (March 2023, Stream Transcript)
The scope of adoption is broad. GPT models are general-purpose technology. They are not point solutions that will impact a narrow set of small industries. Its impact is likely to span multiple industries.
Besides a few notable gainers like Nvidia and Microsoft, equities have yet to price in the economic outcomes of an AI revolution - if one is coming.
We think that a genuine technological breakthrough, with broad impact radius, continued rapid improvements and a very low adoption barrier is definitely something to pay close attention to. As investors, it is our job to be proactive and analyze potential changes and their implications, instead of reacting after the price moves.
“There are predictions that 90% of the content on the internet, be it software, music, images, texts, 90% of all of that content on the internet will be produced by generative AI by the year 2030. That number was less than 1% in 2021. That means that everything that is out there in terms of software will get disrupted by this generative AI field…Every field will probably start using components of generative AI in the next couple of years”
-Former Product Manager, Google AI (January 2023, Stream Transcript)
In addition, what ChatGPT and other AI tools are capable of producing should force us to confront our own skill sets in an honest way. Given what it can do today, and how quickly it is improving, we should be thinking about what we are good at, what is replaceable, and what won’t be. At the very least, it should have us ask ourselves “how do I incorporate this into my research process so that I can become a better investor?”
In our experience, the best way to gain clarity in fast moving, high uncertainty situation is to develop a rudimentary framework, make specific predictions, and compare predictions with how events unfold, updating the framework as necessary. This series of articles will culminate with the discussion of a framework which allows investors to assess the implications of this emerging tech on various businesses and industries.
But first, we must cover some basics - so let’s start with what this thing can do.
What can ChatGPT do?
At a base level, ChatGPT has the ability to:
Take prompts and transform them into answers that are accurate enough that to the user, they will appear to have come from an intelligent being. It can do this in multiple languages.
Stay on a topic through multiple prompts, thereby appearing to carry on a conversation.
Most importantly - do this in subjects that it wasn’t specifically trained for. In contrast, large language models such as GPTs are trained over a huge amounts of general text and can answer questions in multiple domains.5
The combination of this base layer ability means it can do a lot of things that we don’t commonly associate with computers or software:
It can provide you with correct answers on a wide variety of technical subjects. It has passed standardized tests in a wide variety of domains.
It can manipulate text in most ways that humans can. It can “summarize”, “expand”, and generate new content based on a few lines of text.
It can “translate” - not only between languages but between image and text, audio and text and most importantly, computer code and text. Using this capability, it can transcribe, “understand” what’s it an image, “draw”, “code”, and “de-bug” the code.
It can combine multiple concepts that are not related, to create something new. In this way, it displays some level of “creativity.”
It can interact with external programs and sites, through plugins that are built in English. For example, it can “browse” the web, come up with a trip plan, then book that trip on Expedia.
It displays a surprising level of “understanding” of the world and how it works. For example, it can answer questions that require spatial sense and it can reason about the emotional state of others in complex situations. It also has enough context to “understand” nuanced jokes.
You may have noticed that a lot of words above are in quotation marks. It is important to note that as a generative model, ChatGPT isn’t technically doing any of those activities. It just provides us with an accurate enough answer, enough of the time that it appears to be doing so. This leads to some important limitations (though they are being worked on), which we will cover in the next part.
Here are a few examples of what ChatGPT can do. Please keep the above paragraph in mind as you review them. Some of these examples come from a recent research paper from Microsoft Research which we highly recommend you give it a read. Finally, if you haven’t yet tried ChatGPT, there is no better way to get a feel than to try it out yourself!6
Example of ChatGPT (version 3.5) “writing a story.” While the story is very basic, the prompt is also extremely broad. It also has been cut off for length.
Example of ChatGPT (version 4), “drawing” (Note that GPT has been trained only on text data - the fact that it can handle images is intriguing!)
Example of ChatGPT (accessed through Bing), “reading” search results and providing a summary.
Example of ChatGPT (version 4), correctly explaining emotions and the reasons behind them.
Example of ChatGPT (version 4) displaying spatial awareness.
Example of ChatGPT (version 4) decoding a jumbled sentence. H/T:@Liv_Boeree
And finally - while this article is focused on ChatGPT, we couldn’t help but throwing in one Midjourney / image example. Note that both photos below are completely AI generated.
We hope that this article and the above examples are enough to convince you - there’s something special happening here.
Coming up: Part 2: How do GPTs work and what are their limitations?
P.S. That “living under a rock” joke that we started with actually came from ChatGPT. We all have a shot at being funny now. Yes - even Value Punks.
While the article’s focus will be on ChatGPT for simplicity, there are tons of other AI tools worth exploring. For example, Midjourney is a model that can produce photorealistic images based on just a few lines of text.
Note - there is a big difference between - it can answer questions in areas it wasn’t specifically trained on, and - it can provide accurate answers in areas it has no data on (It can’t). This will be explored in Part 2.
To be fair, GPT-4 probably has been in development for much longer than just a few months. But new versions of GPT has been coming out at 1-2 year interval, since 2018.
I mean, technically, there is the data center build-out that will be necessary, but it won’t be anything close to what was required to lay cable across the world.
This ability to generalize is not perfect, of course. While GPTs can answer some out-of-domain questions, the less well-represented a domain is in the training data, the less likely it is able to answer it correctly.
ChatGPT’s capabilities are evolving quickly, as new functionalities are being added on weekly basis. Note that as of March 28th, 2023, the free version is still on version 3.5 (vs. the much improved 4 which is available to ChatGPT Plus subscribers which costs $20 per month). But even the paid version lacks access to web browsing and plugins.
If you are looking for an expert network to get up to speed on industries and companies, then we highly recommend Stream by Alphasense.
Stream by Alphasense is an expert interview transcript library that has been integral to our research process. They are a fast growing expert network with over 25,000 transcripts on a wide variety of industries (TMT, consumers, industrials, real estate and more). We recommend Stream for its high quality transcript library (70% of experts are found exclusively on Stream) and easy-to-use interface. You can sign up for a free trial by clicking here.