Dan Shipper, the
CEO of Every and an AI podcaster, notes that AI is being used in various ways.
He distinguishes the way computers and science see the world, both of which are
trying to:
“…reduce the world into a set of really clean
universal laws that apply in any situation. If X is true, then Y will happen.”
He says that large language
models (LLMs) see something different a:
“…dense web of causal relationships between different parts
of the world that all come together in unique, very context specific ways to
produce what comes next.”
He says LLMs think, or rather
operate, much like human intuition in that they are trained by many hours of
direct experience.
First, he talks about
rationalism, its journey from Socrates to LLMs, and its limitations. He posits
rationalism as the way we see the world. He calls Socrates the father of
rationalism because he was the first to “officially” study truth and what makes
things true due to his emphasis on inquiry. He points to one of Plato’s
dialogues where Socrates was debating with Protagoras whether excellence,
sometimes translated as virtue, can be taught. Here, Socrates stresses the need
to define what excellence really is before determining whether or not it can be
taught. Defining terms and ideas is one way we know things, expressing ideas
through words. Thus, in a sense, to define is to
know.
Shipper says that the
rationalist worldview came to the fore in the Age of Enlightenment due to
thinkers like Descartes, Newton, and Galileo, who used it to explain the world.
With them came the affirmation that one needed to be able to rationally
describe ideas, preferably mathematically.
Shipper sees the social
sciences as following the same framework or structure as the physical sciences.
However, he acknowledges that there are limitations to doing that since the
“soft” or social sciences have subjective aspects that do not lend themselves
well to objective analysis. He mentions the current experimental replication
crisis in psychology as a kind of proof that the social sciences are less
amenable to understanding in the same way as the physical sciences are.
AI is similarly difficult to
rationalize, he suggests. Here, he notes that there is no universally
agreed-upon definition for AI. AI began in the 1950s with a method known as
symbolic AI:
“The idea that you could embody thinking in, essentially
logic, logical symbols, and transformations between logical symbols, which is,
it’s very similar to just basic philosophy.”
Early AI, he says, could only
solve simple problems, not the complex ones that permeate the world. AI follows
rules, but with rules, there are exceptions, and those exceptions must be
defined, and some may need continuous training in as the exceptions can be
dynamic and constantly changing.
Interestingly, the idea of
neural networks was around when AI was young, but wasn’t taken seriously until
the 80s and 90s. Neural networks are inspired and informed by the way the human
brain works.
“What you can do with a neural network is you can get it
to recognize patterns by giving it lots of examples. For example, if you want
it to recognize whether an email is important, what you can do is you can give
it an example, say here’s an email from a coworker, and have it guess the
answer. And if the answer is wrong, what we’ve done is we’ve created a way to
train the network to correct its wrong answer.”
“Language models are a particular neural network that
operates by finding complex patterns inside of language and using that to
produce what comes next in a sequence.”
The LLMs can utilize the
whole internet to accurately predict what is next in a sequence. He points out
that the rules for neural networks are non-explicit, or rather, intuitive, and
often, only some rules can even be found. Here, he gets back to the
similarities to human intuition:
“And what’s really interesting about neural networks is
the way that they think, or the way that they operate it looks a lot like human
intuition. Human intuition is also trained by thousands of hours of direct
experience.”
We have long-standing
metaphors of the human mind being “like” a computer. Shipper sees this as
problematic. Rather, he sees rationality as emerging out of intuition. He then
goes on to suggest that intuition is beyond the reach of rationalism and,
rather than being overshadowed by it, still retains its importance in human
knowledge and understanding. He goes back to the dialogue where Protagoras
argues with the aid of myths and metaphors, instead of rational definitions. He
also says that neural networks are the first thing we have invented that works
like human intuition. It is now widely acknowledged that intuition can be
another way of knowing in addition to rationalism. He also notes that many of
those who use ChatGPT often have noted that they can intuit what it will be
good at or not, and when it is hallucinating, much like we can intuit the
feelings of a close friend.
“The interesting difference between how a language model
sees the world and how a traditional computer sees the world is this: a
traditional computer tries to reduce everything into a set of clean, universal
laws that apply in any situation — essentially, “if X is true, then Y will
happen.” It relies on clear, context-free chains of cause and effect.”
“And what language models see instead is a dense web of
causal relationships between different parts of the world that all come
together in unique very context-specific ways to produce what comes next. I
think language models do something really, really unique, which is that they
can give you the best of what humanity knows, at the right place, at the right
time in your particular context, for you specifically.”
Thus, he says, neural
networks and language models are contextual. They are more based on pattern
matching and a kind of fuzzy logic, and they use previous experience to predict
the future.
“…the way that a more intuitive relational fuzzy pattern
matching type experiential, contextual type way of knowing about the world has
to be underneath the rational stuff for the rational stuff to work at all. It’s
really about recognizing the more intuitive ways of knowing about the world as
being the original parent and partner of rationality, and appreciating that for
what it is.”
The contextual knowledge of
LLMs and neural networks can enable hyper-personalized knowledge that can solve
specific problems. They don’t need explicit definitions or rational models to
be successful. These networks and models are trained to detect and predict. He
says they change a science problem into an engineering problem. He thinks that
better and more thorough data access and sharing, especially by the Big Tech
companies, will lead to more thorough AI training. Shipper thinks we can
utilize our intuition and basically put it into a machine that we can pass
around.
Shipper thinks AI will
seriously enrich our understanding of ourselves. He sees AI as a mirror and as
a metaphor. However, limiting AI to only what is provable limits it.
“The thing about it that makes it powerful is that it
works on probability, it works on thousands of correlations coming together to
figure out what the appropriate response is in this one very unique, very rich
context. And allowing it to say only things that are provable, obviously begs
the question: what is true and how do we know?”
AI can be messy, but the key
thing to know is that with adequate training, it works, though we may not be
able to entirely explain how it works.
“Something to remember is each model builds on the
models that came before it. They actually have a dense, rich idea of what it is
to be good from all the data that they get. They also have a dense, rich idea
of what it is to be bad. But in a lot of ways, the training that we’re doing
makes them less likely to do any of that stuff.”
“There’s something very practical and pragmatic about,
we have a machine, we don’t know fully how it works, but we’re just going to
teach it, and we’re going to iterate with it over, and over again until we
basically get it to work.”
Shipper sees AI as analogous
to a gardener rather than as a sculptor who creates something from nothing. A
gardener gives his plants the conditions to succeed without forcing that
success directly. Gardening is much like a model, he suggests.
“AI does a lot of the more repetitive specialized tasks,
and it will allow individuals to be more generalistic in the work that they do.
And I think that would be a very good thing.”
In the last section, he
mentions the idea that we’re moving from a knowledge economy to an allocation
economy. He also says that an allocation economy will require more of a certain
kind of worker with specific skills. Those specific skills will often be the
skills of human managers, which include knowing one’s human assets and
capabilities, such as “knowing what any given person on your team can do,
what are they good at, what are they not good at.”
“In a knowledge economy, you are compensated based on
what you know. In an allocation economy, you’re compensated based on how well
you allocate the resources of intelligence. There’s a particular set of skills
that are useful today but are not particularly widely distributed that will
become some of the main skills in this new economy, in this new allocation
economy. And that is the skills of managers, those are the skills of human
managers, which make up a very small percentage of the economy right now. I think
it’s like 7% of the economy is a human manager.”
Some of the ongoing problems
that must be solved include how to manage the other humans working on the
problem. He also suggests that managing humans and managing a model are
similar. Managing a model, like managing humans, requires some intuitive
abilities. Shipper mentions an interesting idea that intelligence is like
compression in that it compresses a range of possibilities as answers into a
small space, or rather, it can find the right answers quickly. He thinks that is
partly how brains and consciousness work as well. He suggests that, in a sense,
these models may have consciousness, or at least we can benefit from
understanding them in that way. Just in case, he says:
“I always say please and thank you to ChatGPT because
you never know when the machine apocalypse is going to come.”
Well, now I understand neural networks and LLMs better than I did, and I’m grateful, so a worthwhile article. It was actually a transcribed podcast that I worked from. There is a video as well at the link in the references.
References:
How
Large Language Models View Our World. Big Think. August 27, 2025. How
large language models view our world - Big Think
No comments:
Post a Comment