Value
Investing and AI
“…a test-and-learn
mentality is essential
for translating…data from
raw material into
currency.”
Harvard Business
Review, July 2023
From
a general vantage point, the large language models of generative AI are
becoming increasingly useful and accessible: In the 11/24/23 issue of Foreign
Affairs, Manyika and Pence, the latter a former dean of the Stanford
Business School, wrote about generative AI, large language models (LLMs).
“Generative
AI has several features that suggest its potential economic impact could be
unusually large. One is exceptional versatility. LLMs now have the capacity to
response to prompts in different domains and to respond to prompts in many
different domains, from poetry to law, and to detect different domains and
shift from one to another, without needing explicit instructions Moreover, LLMs
can work not only with words but also with software code, audio, images, video,
and other kinds of inputs, as well as generated outputs – what is often
referred to as “multimodality.” Their ability to operate flexibly among
multiple disciplines and modes means that these models can provide a broad
platform on which to build applications for almost any specific use….LLMs are also noteworthy for their accessibility. Because
these are designed to respond to ordinary language and other ubiquitous inputs,
LLMs can be readily used by nonspecialists who lack technical skills. All that
is needed is a little practice in creating prompts that elicit effective
responses. These two features (multimodality and ease of use) give LLMs far
more extensive potential uses than previous digital technologies, even those
involving AI.”
But
the article then goes on to warn, “At least in the near term, such exuberant
projections will likely outstrip reality. Numerous technological,
process-related, and organizational hurdles, as well as industry dynamics stand
in the way of an AI-driven global economy. But just because the transformation
may not be immediate does not mean the eventual effect will be small.”
For
the academic year 2023-2024 the Stanford Bulletin, Explore Courses,
lists 149 courses involving artificial intelligence ranging from “Contemporary
Art in the Age of Artificial Intelligence” to “Computational Imaging.” Silicon
Valley has been pouring billions into the development of LLM capabilities,
smaller custom language models for focused applications, around 427 new
ventures in the general technology in 2023 (WSJ 4/29/24; “Investors are
Showering AI Startups with Cash. One Problem They Don’t Have Much of a
Business.”), and energy infrastructure.Very
clearly, generative AI isn’t going away. The following examines the technology
of AI, its data requirements, and suggests both its present shortcomings and
its possible future role in the financial markets.
By
its bottom-up nature, LLMs can address new contexts and media multimodalities.
LLMs partially use vector based cosine similarities at
the attention, transformer level, and can then synthesize new datapoints,
based on existing datapoints, in various databases-essentially creating a new
viewpoint for the combined data. There is no magic in the mathematics. The
ability to first place data in the context of the query and then second to
create a new context for the combined data is why we think LLMs are relevant to
new endeavors in the arts as well as the sciences.
There
are, however, quite a few hurdles this technology must cross before it is
suitable for wide-spread use:
· Hallucinations.
Due to the bottom-up nature of LLMs, they can just make things up, unconnected
with reality. They can’t be totally trusted.
· Lack of
specificity and sustained reasoning in various situations. A LLM has not taken
a course in physics 101, a top-down discipline, and therefore cannot design, in
the sensible world, a hi-fi amplifier or a bridge. The model, however, can refer to URLs where
real people, properly invoking the appropriate laws of nature, have done just
that. We found lots of these sources on the internet.
· Biased data. This
is really an interesting question. A biased dataset contains non-representative
data. A good dataset contains representative data. This means, that in
reference to a theory or to a practical business situation, you have identified
(substantially) independent variables. Things then depend upon your
goals, about which more later.
· Extensive human
intervention for the “right” answer. According to a Harvard Business Review
article dated 2017, “A hybrid of human and AI is remaking retail, marketing,
and customer service (appropriate for a service economy). It turns out that AI,
just like humans, struggles to make tough decisions about what content should
or should not be included, depending on what criteria or (n.b.) values we want
to impose.”
We
took a Zoom continuing education course in AI at Stanford taught by Professor Vasseur, the head of business analytics at the business
school at Berkeley. Rather than groping around in obscurity, we learned a lot
about AI and therefore its applicability to future economic productivity.
The
Data
The
foundation for effective AI is always the data. For data to be usable in a
business decision, where costs and benefits are calculated, it must be FAIR:
· FINDABLE. It may
be all over the place, in spreadsheets and notes.
· ACCESSIBLE. In
disparate computer systems.
· INTERPRETABLE. It
may not exist in a common data format.
· REUSABLE. It can’t
get lost.
The People
Only
humans have the insights that can be validated or optimized by one of the (sic)
thousands of AI models and their architectures. Only humans can form the
organizations that implement these models, that lie above four or five massive
foundation models such as ChatGPT and Gemini. In other words, YOU - with your
goals - are the person that can improve, reflect, collaborate, or innovate.
What
LLMs do is to make, through natural language, the many and sophisticated models
of AI much more usable and accessible. Through LLMs it is possible with Python
to reorganize a data base of 2 million customers in a second. It is also
possible to plot in several dimensions and colors model outputs in 4 minutes.
These tools, if adopted, can increase productivity. But they are far from an
oracle; you have to make them work appropriately. To
remain current, AI models don’t exist in a static universe; they have to be run often as things change.
Experts
train these models for more focused applications. Experts have a significant
top-down component to their thinking, for which LLMs are the opposite. Gaining
their trust takes time.
We
talked with several AI researchers, who have hands-on acquaintances with the
developing technology. The profitable application of business AI, with smaller
models in specific industries, still requires lots of work.
R&D
Creativity
is the ability to see things anew, often from existing building blocks.
Generative AI has the ability to form, subject to some
constraint, the building blocks of new drugs, products and services. We just
note that it takes several years to develop new drugs and materials.
The
sum total of what we are saying is that it will take
several years for the billions invested in generative AI to turn a profit. The
present stock market assumption, which we detailed in the 1/29/24 posting, is
either that the economy has become non-cyclical, present earnings will grow
forever; or that the economy remains cyclical, but
grows in productivity at twice (4%) the historical rate. Given PROBLEMS in
the market environment, we don’t think these are at all possible.
· Climate
Change
· Large
U.S. Deficits
· Deglobalization Call this parent
class: PROBLEMS
· Demographic
Changes
· Threatened
Political Change
As we noted in an earlier post, stock investors have to be optimistic about the future. AI, though now a
fad, lends some hope for this. But, as the following graph indicates, present
economic productivity is below our long-term equity assumption.
As value investors, we also look to the past, within
reason. & Who are those people across the seas?