On 12/21/22 7:50 AM, Barry Rowlingson wrote:
Next year one of my R programming assigments will read like this:

"Here is some R code written by a multi-million dollar AI system to
compute [something]. It doesn't work. Fix the bugs, then ask the AI to
write a letter to its creators apologising for how rubbish it is at
coding. Collect one million dollars."


You might want to be careful about such a promise. Kahneman, Sibony, and Sunstein (2021) Noise: A flaw in human judgment (Little, Brown and Company) claim that genuine expertise is acquired by learning from frequent, rapid, high-quality feedback on the quality of their decisions. Few people have access to such feedback. They call leaders in fields without such feedback "respect-experts", and note that respect-experts have only the illusion of competence.


1. They further say that most respect-experts can be beaten by simple heuristics developed by intelligent lay people.


2. Moreover, with a modest amount of data, ordinary least squares can beat most such heuristics.


3. And if lots of data are available, AI can beat the simple heuristics.


They provide substantial quantities of research to support those claims.


Regarding your million dollars, it should not be hard to write an R interface to existing AI code cited by Kahneman et al.


Do you really want one of your students initiating a legal procedure to try to collect your million dollars?


A quarter century ago, my friend Soren Bisgaard told me about a colleague who had raved about AI. Soren thought, "You prefer artificial intelligence to real intelligence?"


I perceive a role for AI in identifying subtle phenomena missed by more understandable modeling techniques. Let's use the best understandable model, and apply AI to the residuals from that. Then identify the variables that make the largest contributions to a useful AI model, and see if they can be added to the other model.


          Spencer Graves


On Sun, Dec 18, 2022 at 7:01 PM Boris Steipe <boris.ste...@utoronto.ca> wrote:

This email originated outside the University. Check before clicking links or 
attachments.

Technically not a help question. But crucial to be aware of, especially for 
those of us in academia, or otherwise teaching R. I am not aware of a suitable 
alternate forum. If this does not interest you, please simply ignore - I 
already know that this may be somewhat OT.

Thanks.
------------------------------------------------------

You very likely have heard of ChatGPT, the conversation interface on top of the 
GPT-3 large language model and that it can generate code. I thought it doesn't 
do R - I was wrong. Here is a little experiment:

MY PROMPT:
==========
Please write me some R code that will read a tsv file, remove all rows that have duplicates in the 
column "names", and store the median values of the remaining rows in a variable called 
"Medians". Of course, that calculation should only be done with columns that are numeric. 
Print the last variable in order from largest to smallest.


AI RESPONSE:
============

Here is some R code that should accomplish what you've described:

(Note: the code is formatted in a little terminal window, complete with a button to 
"copy code" at the top.)

----
# Read the tsv file
data <- read.table("file.tsv", header = TRUE, sep = "\t")

# Remove rows with duplicates in the "names" column
data <- data[!duplicated(data$names), ]

# Select only the numeric columns
num_columns <- sapply(data, is.numeric)
numeric_data <- data[, num_columns]

# Calculate the medians for each column
Medians <- apply(numeric_data, 1, median)

# Sort Medians in descending order
Medians <- sort(Medians, decreasing = TRUE)

# Print Medians
print(Medians)
----

This code assumes that the tsv file is called "file.tsv" and is stored in the 
current working directory. It also assumes that the first row of the file contains the 
column names.



MY PROMPT:
==========
Can you provide me with a bit of sample data that I can use to test this code?

AI RESPONSE:
============
Certainly! Here is some sample data that you can use to test the code I 
provided:

----
name    age     height  weight
Alice   23      5.6     120
Bob     32      6.2     180
Charlie 29      5.9     160
Alice   23      5.6     120
David   25      5.8     140
Eve     27      5.5     110
----

You can save this data as a tsv file and use it to test the code.


------------------------------------------------------

Notes:
1) The code, as written does not work with the test data. The reason is that the test data 
specifies a column "name", but the code assumes "names" (which I specified in 
the instructions). Once this is fixed in the test data, the code executes.

2) The instructions are not accurately translated: I had asked for row means. 
But changing the margin in the apply() statement from 2 to 1 does the correct 
thing.

3) Note that the test data contains both numeric and non-numeric columns. Also, the name 
"Alice" is duplicated in the test data, which is subtle, and the right thing to 
do.


------------------------------------------------------

A similar experiment, run on the beta-site of OpenAI's Codex code-completion 
tool gives the following code (complete with the comments.):

----
# Read in the data
data <- read.table("mydata.tsv", sep = "\t", header = TRUE)

# Remove all rows that have duplicates in the column "names"
unique_names <- unique(data$names)
data_unique <- data[data$names %in% unique_names,]

# Calculate medians
Medians <- sapply(data_unique[, sapply(data_unique, is.numeric)], median, na.rm 
= TRUE)

# Print last variable in order form largest to smallest
rev(sort(Medians[length(Medians)]))
----

Note that the strategy is quite different (e.g using %in%, not duplicated() ), the 
interpretation of "last variable" is technically correct but not what I had in 
mind (ChatGPT got that right though).


Changing my prompts slightly resulted it going for a dplyr solution instead, 
complete with %>% idioms etc ... again, syntactically correct but not giving me 
the fully correct results.

------------------------------------------------------

Bottom line: The AI's ability to translate natural language instructions into code is 
astounding. Errors the AI makes are subtle and probably not easy to fix if you don't 
already know what you are doing. But the way that this can be "confidently 
incorrect" and plausible makes it nearly impossible to detect unless you actually 
run the code (you may have noticed that when you read the code).

Will our students use it? Absolutely.

Will they successfully cheat with it? That depends on the assignment. We 
probably need to _encourage_ them to use it rather than sanction - but require 
them to attribute the AI, document prompts, and identify their own, additional 
contributions.

Will it help them learn? When you are aware of the issues, it may be quite 
useful. It may be especially useful to teach them to specify their code 
carefully and completely, and to ask questions in the right way. Test cases are 
crucial.

How will it affect what we do as instructors? I don't know. Really.

And the future? I am not pleased to extrapolate to a job market in which they 
compete with knowledge workers who work 24/7 without benefits, vacation pay, or 
even a salary. They'll need to rethink the value of their investment in an 
academic education. We'll need to rethink what we do to provide value above and 
beyond what AI's can do. (Nb. all of the arguments I hear about why humans will 
always be better etc. are easily debunked, but that's even more OT :-)

--------------------------------------------------------

If you have thoughts to share how your institution is thinking about academic 
integrity in this situation, or creative ideas how to integrate this into 
teaching, I'd love to hear from you.


All the best!
Boris


--
Boris Steipe MD, PhD
University of Toronto

______________________________________________
R-help@r-project.org mailing list -- To UNSUBSCRIBE and more, see
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.

______________________________________________
R-help@r-project.org mailing list -- To UNSUBSCRIBE and more, see
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.

______________________________________________
R-help@r-project.org mailing list -- To UNSUBSCRIBE and more, see
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.

Reply via email to