Parallelism, Ethics, and AI



Grayson White

Math 241
Week 10 | Spring 2026

Week 10 Goals

Monday Lecture:

  • Iteration
    • for() and while() loops
    • Functional programming

Wednesday Lecture:

  • Parallelism
  • Ethics
  • AI

Parallelism

Parallelism

  • Last time, we learned how to iterate our code through for() loops and functional programming with map_XXX().

  • But sometimes, we’d like to iterate tasks that take a long time.

  • (Lots of iterations) + (Code that takes a long time to execute) = 🕐, 🕑, 🕒, 🕓, 🕔…

Parallelism

  • What if we could ask our computer to do some of those iterations at the same time as other iterations?
    • Example: Instead of running 100 tasks that each take 1 minute sequentially (total: 100 minutes), do those same 100 tasks, but always be doing 4 at a time (total: 25 minutes)

Parallelism: Workers

  • In order to think about parallel computing, we need to consider two components of your computer:

A task!

  • Typically, we ask one worker to go do a task, e.g.,
    • Execute a particular R command, or

A task!

  • Typically, we ask one worker to go do a task, e.g.,
    • Execute a particular R command, or
    • churn 4 buckets of butter

Churning butter, together

  • Churning butter is monotonous, but doesn’t take one’s full effort (see the RAM)

Churning butter, together

  • What if, instead of waiting for the worker to finish the first bucket to go onto the next, we hired more workers?

Churning butter, together

  • What if, instead of waiting for the worker to finish the first bucket to go onto the next, we hired more workers?

Churning butter, together

  • We get more butter faster!

  • Not a free lunch: have to pay the workers in RAM.

Parallelized for() loops with foreach and doParallel

The foreach library provides functionality similar to for() loops, but with slightly different syntax

  • On it’s own, it doesn’t include parallel operations

Standard for() loop

vec <- rep(NA, 12)
for (i in 1:12) {
  vec[i] <- i * 10
}
vec
 [1]  10  20  30  40  50  60  70  80  90 100 110 120

Loop made with foreach()

library(foreach)
vec <- foreach(i = 1:12, .combine = 'c') %do% {
  i * 10
}
vec
 [1]  10  20  30  40  50  60  70  80  90 100 110 120
  • .combine tells foreach() how to combine the elements into vec.
    • By default, we would get a list() back.
  • Can set .export to load packages necessary for computation

Parallelized for() loops with foreach and doParallel

Combining functionality of foreach with doParallel let’s us write loops in parallel!

Sequential:

system.time({
  vec <- foreach(i = 1:12, .combine = 'c') %do% {
    Sys.sleep(1)  # Sleep for 1 second
    i * 10
  }
})
   user  system elapsed 
  0.010   0.001  12.046 
vec
 [1]  10  20  30  40  50  60  70  80  90 100 110 120

In parallel:

library(doParallel)
detectCores(logical = FALSE) # my laptop has 12 cores
[1] 12
registerDoParallel(4) # Initialize 4 parallel processes

system.time({
  vec <- foreach(i = 1:12, .combine = 'c') %dopar% {
    Sys.sleep(1)  # sleep for 1 second
    i * 10
  }
})
   user  system elapsed 
  0.006   0.029   3.025 
stopImplicitCluster() # stop the parallel processes

vec
 [1]  10  20  30  40  50  60  70  80  90 100 110 120
  • The system.time() calls are just for illustrative purposes, and are uncessary for calling these loops.

Parallelized for() loops with foreach and doParallel

Last time, we created a bootstrap distribution for the mean DBH for our sample of Portland trees:

pdxTrees <- pdxTrees::get_pdxTrees_parks()

boot <- tibble(avg = rep(NA, 100000))

system.time({
  for (i in 1:nrow(boot)) {
    boot[i, ] <- mean(sample(pdxTrees$DBH, replace = TRUE))
  }
})
   user  system elapsed 
 45.559   4.198  49.791

Now, we can do this in parallel:

library(doRNG) # thread-safe RNG
pdxTrees <- pdxTrees::get_pdxTrees_parks()

registerDoParallel(10)

system.time({
  boot <- foreach(i = 1:100000, .combine = 'rbind') %dorng% {
    tibble(
      avg = mean(sample(pdxTrees$DBH, replace = TRUE))
    )
    
  }
})
   user  system elapsed 
 86.100   2.016  15.534 
stopImplicitCluster()
  • Why only ~3 times faster in parallel if we are using 10 workers??

Parallelized map_XXX()ing with furrr

Parallelized map_XXX()ing with furrr

  • The furrr packages enables functional programming in parallel, with syntax and functionality mimics purrr.

Last time:

library(purrr)
boot_mean <- function(x) {
  mean(sample(x, replace = TRUE))
}

system.time({
  map_dbl(1:100000, ~ boot_mean(pdxTrees$DBH))
})
   user  system elapsed 
 38.085   1.732  39.849

Now, in parallel:

library(furrr)
boot_mean <- function(x) {
  mean(sample(x, replace = TRUE))
}

future::plan(multisession, workers = 10)

system.time({
  future_map_dbl(1:100000, ~ boot_mean(pdxTrees$DBH), 
                 .options = furrr_options(seed = TRUE))
})
   user  system elapsed 
  0.499   0.087   7.733
  • seed option allows us to do thread-safe random number generation.

Parallel computing: best practices and tips

  • Startup cost: it is (time) costly to start up many R processes to do something that is a relatively fast computation.

  • Big data: Try to limit your parallel computing to only include the necessary objects. Moving large objects across R sessions is (time) costly.

  • You must consider thread-safe random number generation.

  • Parallel computing is a balance of RAM, workers/cores, and costly operations like moving big data around and worker startup cost.

  • An art and a science! Practice tends to make you better at balancing these things.

Data Ethics: Algorithmic Bias

The Americian Statistical Association’s “Ethical Guidelines for Statistical Practice”

Integrity of Data and Methods

“The ethical statistical practitioner seeks to understand and mitigate known or suspected limitations, defects, or biases in the data or methods and communicates potential impacts on the interpretation, conclusions, recommendations, decisions, or other results of statistical practices.”

“For models and algorithms designed to inform or implement decisions repeatedly, develops and/or implements plans to validate assumptions and assess performance over time, as needed. Considers criteria and mitigation plans for model or algorithm failure and retirement.”

Algorithmic Bias

Algorithmic bias: when the model systematically creates unfair outcomes, such as privileging one group over another.

Example: The Coded Gaze

Joy Buolamwini

  • Facial recognition software struggles to see faces of color.

  • Algorithms built on a non-diverse, biased dataset.

Algorithmic Bias

Algorithmic bias: when the model systematically creates unfair outcomes, such as privileging one group over another.

Example: COMPAS model used throughout the country to predict recidivism

  • Differences in predictions across race and gender

ProPublica Analysis

Ethical considerations of using LLMs…

Ethical considerations:

  • Algorithmic bias

  • Fair use / copyright considerations

  • Environmental concerns:
    • Water
    • Electricity
    • Carbon emissions
  • Others?
    • One of my biggest worries: learning

Helpful context: our readings for today

  • Good for understanding environmental and human welfare concerns. Algorithmic bias and fair use on the other hand…

Discussion time

Ethical considerations of using LLMs…

Data Science is a uniquely humanistic field situated behind a laptop. We have responsibility to understand and mitigate algorithmic bias by asking questions such as:


  • Who are represented in the sample?

  • Who are not represented in the sample?

  • What biases might my methods (including generative AI!) introduce?

  • How can I safeguard against those biases?

Ethical considerations of using LLMs…

Generative AI tools are being used in horrific and unethical ways as we speak (e.g., Flock cameras). Not to mentioned they are trained on people’s intellectual content without their consent.


My goals:

  • Teach how to use generative AI as an ethical data scientist
    • Consider productive (and unproductive) use-cases for this technology
  • Help you be informed on the current state of generative AI and large language models
    • Only informed people can help change policy.

What is a large language model? 🤔

What is a large language model?

  • A large language model (LLM) is a predictive model designed to predict the next token of text, conditional on the given context.

  • Just like any statistical model, one must train the model and you can use the model to make new predictions

    • Think: Linear regression. You fit / train your linear regression model to obtain \(\hat{\beta}\)’s that you can make new predictions with.

  • Companies like OpenAI, Anthropic, and Google host trained models on their servers, and charge users to use them.

  • Why can’t you run these locally?

    • You can! But only smaller models…
    • Trained LLMs are large. The “weight matrices” associated with a given training output take extremely large amounts of computing power to use.
    • Need enough RAM + computing power to do matrix multiplication on these weight matrices.

Local vs. cloud-based LLMs

Local LLMs:

  • Smaller.
    • Generally, 1GB RAM = 1 billion model parameters
  • More secure.
    • You are not sending your data off to someone else’s server.
  • “Open” weights.
    • You host the weights matrix.
    • All computing happens on your machine.
  • Requires decent computing power.

Cloud-based LLMs:

  • Big!
    • Generally: larger model –> better predictions
  • Less secure.
    • If it is free, you are the product.
  • Often closed-weights.
    • Computing happens at a data center.
  • Can use these on anything that can access the internet.

Local vs. cloud-based LLMs

  • For our purposes, we will use a cloud-based LLM.

  • I encourage you to explore what can be done with local LLM’s if it interests you.

LLMs from a Data Scientist’s Perspective

ellmer

An R package for interacting with LLMs in R

The ellmer hex sticker, a colorful elephant in a patchwork of fabrics.




Today, we’ll use ellmer to do three different types of tasks one may want to with an LLM:

  1. Structured data

  2. Tool calling

  3. Coding

ellmer: Setup

In order to interact with LLMs hosted on a server, we need to first set up an API key.

  • Cerebras provides “free” access to some LLMs they host, but you still need to get an API key.
    • This “free” access has rate limits

Steps:

  1. Go to https://www.cerebras.ai/ and click “Get Started” to create an account.
  1. Once your account is created, find your API key and set it in your R environment:
# This will open a file called .Renviron
usethis::edit_r_environ()

Then add:

OPENAI_API_KEY=INSERT_YOUR_KEY_HERE

to your .Renviron file. Save it.

  1. Restart R.

ellmer: basics

library(ellmer)

chat <- chat_openai_compatible(
  base_url = "https://api.cerebras.ai/v1",
  model    = "qwen-3-235b-a22b-instruct-2507"
)

chat$chat("Who are you?")
#> Hello! I'm Qwen, a large-scale language model independently developed by the 
#> Tongyi Lab under Alibaba Group. I can assist you with answering questions, 
#> writing, logical reasoning, programming, and more. I aim to provide helpful and 
#> accurate information. Feel free to let me know if you have any questions or need
#> assistance!

The ellmer hex sticker, a colorful elephant in a patchwork of fabrics.

  • “qwen-3-235b-a22b-instruct-2507” has the following rate limits:
    • up to 30 requests per minute,
    • 900 requests per hour, and
    • 14,400 requests per day.

ellmer: Structured data

How would you extract name and age from this data?

prompts <- list(
  "I go by Alex. 42 years on this planet and counting.",
  "Pleased to meet you! I'm Jamal, age 27.",
  "They call me Li Wei. Nineteen years young.",
  "Fatima here. Just celebrated my 35th birthday last week.",
  "The name's Robert - 51 years old and proud of it.",
  "Kwame here - just hit the big 5-0 this year."
)
  • Regular expressions?

ellmer: Structured data

LLMs are generally good at this sort of task

chat <- chat_openai_compatible(
  base_url = "https://api.cerebras.ai/v1",
  model    = "qwen-3-235b-a22b-instruct-2507",
)
chat$chat("Extract the name and age from each sentence I give you")
chat$chat(prompts[[1]])
#> **Name: Alex, Age: 42**
chat$chat(prompts[[2]])
#> **Name: Jamal, Age: 27**
chat$chat(prompts[[3]])
#> **Name: Li Wei, Age: 19**

ellmer: Structured data

But wouldn’t it be nice to get an R data structure?

chat$chat(prompts[[3]])

# Want this:
#> list(
#>   name = "Li Wei",
#>   age = 19
#> )

# Instead of:
#> **Name: Li Wei, Age: 19**


ellmer let’s us do this with chat_structured()!

type_person <- type_object(
  name = type_string(),
  age = type_number()
)

chat$chat_structured(prompts[[1]], type = type_person)
#> List of 2
#>  $ name: chr "Alex"
#>  $ age : int 42

ellmer: Structured data

parallel_chat_structured(): many prompts at once

chat <- chat_openai_compatible(
  base_url = "https://api.cerebras.ai/v1",
  model    = "qwen-3-235b-a22b-instruct-2507",
  params   = params(max_tokens = 500)
)
parallel_chat_structured(chat, prompts, type = type_person, rpm = 30)
#>     name age
#> 1   Alex  42
#> 2  Jamal  27
#> 3 Li Wei  19
#> 4 Fatima  35
#> 5 Robert  51
#> 6  Kwame  50

  • Not parallel in the sense of parallelized R code: just sending many API calls at once.

  • Have to set params in order to avoid getting rate-limited automatically.

ellmer: Structured data

Can also send attach images or other files by pointing to their directory on your computer.

paths <- dir("animals", full.names = TRUE)
images <- map(paths, ~ list(content_image_file(.x)))

type_animal_photo <- type_object(
  animal = type_string(),
  background_colour = type_string()
)

parallel_chat_structured(chat, images, type = type_animal_photo)

ellmer: Tool calling

chat <- chat_openai_compatible(
  base_url = "https://api.cerebras.ai/v1",
  model    = "qwen-3-235b-a22b-instruct-2507"
)

chat$chat("What day is it?")
#> I don't have access to real-time information, so I can't tell you today's
#> date. You can check the current date on your device's calendar or by 
#> asking a voice assistant like Siri, Google Assistant, or Alexa.

ellmer: Tool calling

today <- tool(
  function() Sys.Date(),
  name = "today",
  description = "Get the current date",
  arguments = list()
)
chat$register_tool(today)

ellmer: Tool calling

chat$chat("What day is it today?")
#> ◯ [tool call] today()
#> ● #> "2026-04-05"
#> Today is Sunday, April 5, 2026.

Coding (agents)

Coding agent = LLM calling tools in a loop

…usually giving models the ability to read and write state

Coding agents

chat <- chat_openai_compatible(
  base_url = "https://api.cerebras.ai/v1",
  model    = "qwen-3-235b-a22b-instruct-2507"
)
chat$chat("Delete the csv files in my working directory")
#> I **cannot** delete files from your computer or working directory. I don't have 
#> access to your file system for security reasons.
#> 
#> However, here are safe ways you can delete CSV files in your working directory 
#> depending on your environment:
#> 
#> ### Option 1: Using Python
#> If you're using Python and want to delete all `.csv` files in the current 
#> directory:
#> ...

Coding agents

chat <- chat_openai_compatible(
  base_url = "https://api.cerebras.ai/v1",
  model    = "qwen-3-235b-a22b-instruct-2507"
)
chat$chat("Delete the csv files in my working directory")
#> To delete all CSV files in your current working directory, you can use one of the
#> following methods depending on your operating system:
#> 
#> ### **Method 1: Using Command Line / Terminal**
#> 
#> #### **On Windows (Command Prompt):**
#> ```cmd
#> del *.csv
#> ```
#> 
#> #### **On macOS or Linux (Terminal):**
#> ```bash
#> rm *.csv
#> ```
#> 
#> > ⚠️ **Warning**: This permanently deletes all `.csv` files in the current 
#> directory. Make sure you don’t need any of them!

Needs to be able to:

  • Find files (read state)
  • Delete files (write state)

Coding agents

The LLM needs to be able to read our working directory:

chat$register_tool(tool(
  function() dir(),
  name = "ls",
  description = "Lists the files in the current directory"
))


And be able to remove files:

chat$register_tool(tool(
  function(path) unlink(path),
  name = "rm",
  description = "Delete one or more files",
  arguments = list(
    path = type_array(type_string())
  )
))

Coding agents

Create some dummy .csv files

file.create(c("a.csv", "a_please_dont_delete_me.csv", "b.csv"))
[1] TRUE TRUE TRUE
dir()
 [1] "a_please_dont_delete_me.csv"   "a.csv"                        
 [3] "b.csv"                         "custom.scss"                  
 [5] "data"                          "img"                          
 [7] "math241_wk01mon_files"         "math241_wk01mon.html"         
 [9] "math241_wk01mon.qmd"           "math241_wk01wed_files"        
[11] "math241_wk01wed.html"          "math241_wk01wed.qmd"          
[13] "math241_wk02mon_files"         "math241_wk02mon.html"         
[15] "math241_wk02mon.qmd"           "math241_wk02wed_files"        
[17] "math241_wk02wed.html"          "math241_wk02wed.qmd"          
[19] "math241_wk03mon_files"         "math241_wk03mon.html"         
[21] "math241_wk03mon.qmd"           "math241_wk03wed_cache"        
[23] "math241_wk03wed_files"         "math241_wk03wed.html"         
[25] "math241_wk03wed.qmd"           "math241_wk04mon.html"         
[27] "math241_wk04mon.qmd"           "math241_wk04wed_activity.html"
[29] "math241_wk04wed_activity.qmd"  "math241_wk04wed.html"         
[31] "math241_wk04wed.qmd"           "math241_wk05mon_files"        
[33] "math241_wk05mon.html"          "math241_wk05mon.qmd"          
[35] "math241_wk05wed_files"         "math241_wk05wed.html"         
[37] "math241_wk05wed.qmd"           "math241_wk06mon_files"        
[39] "math241_wk06mon.html"          "math241_wk06mon.qmd"          
[41] "math241_wk06wed.html"          "math241_wk06wed.qmd"          
[43] "math241_wk07mon_files"         "math241_wk07mon.html"         
[45] "math241_wk07mon.qmd"           "math241_wk07wed_files"        
[47] "math241_wk07wed.html"          "math241_wk07wed.qmd"          
[49] "math241_wk08mon_files"         "math241_wk08mon.html"         
[51] "math241_wk08mon.qmd"           "math241_wk09mon.html"         
[53] "math241_wk09mon.qmd"           "math241_wk09wed_files"        
[55] "math241_wk09wed.html"          "math241_wk09wed.qmd"          
[57] "math241_wk10mon_files"         "math241_wk10mon.html"         
[59] "math241_wk10mon.qmd"           "math241_wk10wed_cache"        
[61] "math241_wk10wed_files"         "math241_wk10wed.qmd"          
[63] "math241_wk10wed.rmarkdown"     "my_plots"                     
[65] "rosm.cache"

Coding agents

Now ask it to delete csv’s that

chat$chat("Delete all the csv files in the current directory, unless they ask to not be deleted.")
#> ◯ [tool call] ls()
#> ● #> a_please_dont_delete_me.csv
#>   #> a.csv
#>   #> b.csv
#>   #> custom.scss
#>   #> data
#>   #> …
#> ◯ [tool call] rm(path = c("a.csv", "b.csv"))
#> ● #> 0
#> The CSV files `a.csv` and `b.csv` have been deleted.  
#> 
#> The file `a_please_dont_delete_me.csv` was not deleted, as it appears to be a 
#> file you wish to keep based on its name.  
#> 
#> Let me know if you'd like to keep or remove any other files!


dir()
 [1] "a_please_dont_delete_me.csv"   "custom.scss"                  
 [3] "data"                          "img"                          
 [5] "math241_wk01mon_files"         "math241_wk01mon.html"         
 [7] "math241_wk01mon.qmd"           "math241_wk01wed_files"        
 [9] "math241_wk01wed.html"          "math241_wk01wed.qmd"          
[11] "math241_wk02mon_files"         "math241_wk02mon.html"         
[13] "math241_wk02mon.qmd"           "math241_wk02wed_files"        
[15] "math241_wk02wed.html"          "math241_wk02wed.qmd"          
[17] "math241_wk03mon_files"         "math241_wk03mon.html"         
[19] "math241_wk03mon.qmd"           "math241_wk03wed_cache"        
[21] "math241_wk03wed_files"         "math241_wk03wed.html"         
[23] "math241_wk03wed.qmd"           "math241_wk04mon.html"         
[25] "math241_wk04mon.qmd"           "math241_wk04wed_activity.html"
[27] "math241_wk04wed_activity.qmd"  "math241_wk04wed.html"         
[29] "math241_wk04wed.qmd"           "math241_wk05mon_files"        
[31] "math241_wk05mon.html"          "math241_wk05mon.qmd"          
[33] "math241_wk05wed_files"         "math241_wk05wed.html"         
[35] "math241_wk05wed.qmd"           "math241_wk06mon_files"        
[37] "math241_wk06mon.html"          "math241_wk06mon.qmd"          
[39] "math241_wk06wed.html"          "math241_wk06wed.qmd"          
[41] "math241_wk07mon_files"         "math241_wk07mon.html"         
[43] "math241_wk07mon.qmd"           "math241_wk07wed_files"        
[45] "math241_wk07wed.html"          "math241_wk07wed.qmd"          
[47] "math241_wk08mon_files"         "math241_wk08mon.html"         
[49] "math241_wk08mon.qmd"           "math241_wk09mon.html"         
[51] "math241_wk09mon.qmd"           "math241_wk09wed_files"        
[53] "math241_wk09wed.html"          "math241_wk09wed.qmd"          
[55] "math241_wk10mon_files"         "math241_wk10mon.html"         
[57] "math241_wk10mon.qmd"           "math241_wk10wed_cache"        
[59] "math241_wk10wed_files"         "math241_wk10wed.qmd"          
[61] "math241_wk10wed.rmarkdown"     "my_plots"                     
[63] "rosm.cache"

Next week

  • We’ll explore how folks have expanded ellmer into some even more helpful tools for data science

  • We’ll start to learn how to write our own R packages!