a foundational approach to artificial hallucinations in reinforcement learning

José Manuel Rodríguez Caballero josephcmac at gmail.com
Wed Mar 8 18:15:57 EST 2023

Yu Li wrote about the conversation with:

> I would like to share my conversations with ChatGPT about Gödel's
> incompleteness theorem.


> Overall, ChatGPT is surprisingly performing!

Firstly, if systems based on reinforcement learning, such as ChatGPT, are
to perform mathematics in real-life scenarios, such as controlling an
industry automatically, then the foundations of this new method of
mathematical problem-solving should be incorporated into the field of
foundations of mathematics. Proof assistants, which are more stable than
reinforcement learning systems, are already a subject of discussion in the
foundations of mathematics community. Each category of machine learning
method that performs mathematical tasks should have its own foundations, as
it is unlikely that there exists a single theory to encompass them all.
Limiting the foundations of mathematics to human-made mathematics is a bias
known as an anthropocentric worldview [2]. This bias can prevent us from
recognizing forms of non-human intelligence, such as animal cognition [3].

Mikhail Gromov said that only mathematicians and schizophrenics may trust a
chain of ten consecutive arguments [1]. In life, one or two arguments are
sufficient because the chain tends to break down. It is easy to estimate
the length of the chain of consecutive arguments that a given reinforcement
learning system at a given time can handle when performing mathematical
tasks. Establish a hierarchy of elementary mathematical problems that
involve increasingly more non-trivial steps in the solutions. I emphasize
that these steps should be non-trivial because the neural network easily
handles trivial steps.

As an experiment, which can be replicated by anyone, I challenged ChatGPT
to solve the following problem known as “Cows in the Meadow” [4]: The grass
in a meadow grew equally thick and fast. It was known that 70 cows could
eat it up in 24 days, while 30 cows could do it in 60 days. How many cows
would crop the grass of the whole meadow in 96 days? The solution is that
20 cows would have eaten up all the grass in 96 days, but the reasoning
involves several non-trivial steps and many people get confused by that.
After writing equations like a student on the blackboard, ChatGPT concluded:

Therefore, 35 cows would be needed to eat all the grass in 96 days.

I asked it:

Can you find an error in the above reasoning?

and it answered “Yes” and the explanation of its own mistake included:

Equating the cow-days required by 70 cows and 30 cows, we get:
> 70 x 24 = 30 x 60
> 1680 = 1800

This is a contradiction, which means that there is no solution for the
> number of cows needed to eat the grass in 96 days with the given
> information.

The task for foundations of mathematics concerning reinforcement learning
systems is not to increase their creativity, which is the data scientist's
job. Foundations of mathematics may be interested in developing a method to
prevent nonsensical solutions, also known as artificial hallucinations. In
humans, this issue was solved to some extent by the axiomatic method.
Therefore, the problem could be to develop an axiomatic system that will
resonate with the way reinforcement learning works. This is a particular
case of the subaxiomatic foundations' goal: to develop an efficient
foundation of mathematics that is more computationally efficient than the
human-centered axiomatic foundations.

>From a statistical mechanics point of view, using reinforcement learning is
like cooling the system, and the user interacting with the system is like
heating it. When the temperature is too high, there will be a phase
transition, that is perceived by the user as an artificial hallucination.
The main theoretical problem is to find a device that will shoot down the
system before it becomes too hot. Maybe proof assistants are that device.

Kind regards,
Jose M.

[1] Mikhail Gromov - 1/2 Probability by Homology, URL:

[2] Boslaugh, S. E. (2016, January 11). anthropocentrism. Encyclopedia
Britannica. https://www.britannica.com/topic/anthropocentrism

[3] Pepperberg, Irene M. “Grey parrot numerical competence: a review.”
Animal cognition 9 (2006): 377-391.

[4] Yakov Perelman, “Algebra Can Be Fun”, MIR. 1979.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: </pipermail/fom/attachments/20230308/08fb9ceb/attachment.html>

More information about the FOM mailing list