CSE 390HA - 122 Honors Seminar

• the history of segregated schools and pools is relevant to computer science education (CS Ed) the 2000s
• metaphorically, being "stuck in the shallow end" happens when you don't have access to the right skills and resources. This is just like CS Ed: it's not a lack of interest, but a lack of resources.
• similar to swimming, insecurity is a huge barrier to entry for CS. These insecurities stem from historical elements of society.
• expectations are built from existing societal structures - similar to the floating line that separates the shallow and deep ends of a pool, students may feel like an invisible expectation prevents them from doing CS.
• in both cases, we often don't solve the true, structural problem - often ignoring it or applying a bandaid fix.

• the book explicitly discusses the stereotype that computer scientists are white or Asian (or Indian); this pressure feels real!
• similar identity issues around being a "math-oriented", "STEM-oriented", or "writing-oriented" person.
• computer scientists are perceived as technical, concise speakers, and good at math - even though CS is much more than that!
• a relationship between "these groups are good at math" to "these groups are good at computing"
• core underlying issue: representation. Not seeing someone who looks like you in computer science impacts if you'd engage yourself.
• as a class, we then tried to name famous swimmers (Michael Phelps, Katie Ledecky, and Ryan Lochte) - who were all white.
• as a class, we then tried to name famous computer scientists (Bill Gates, Steve Jobs, Tim Cook, Paul Allen, Ada Lovelace, Alan Turing, Richard Feynman, "the people from Hidden Figures"). Matt commented that typically, people are unable to name women, and that none of these people are asian - even though that's a heavy stereotype of software engineers.
• "Asian" isn't a monolithic categorization, and depends on what you count as Asia (e.g. is Russia in Asia? the Middle East?). Participation and inclusion is very different for East Asians, Southeast Asians, and further breakdowns.

Note from Matt: going to mostly omit the specific points here since they're personally identifying. Instead, these were just the broad-strokes themes - not a play-by-play of the conversation.

• large difference in expected graduation rates (~50% versus always 100%)
• in a different state, had single-digit Latino/a representation - "almost too few people to discriminate against".
• connection to generational expectations in the book. Some students grew up in engineering centers where everybody had at least one parent as an engineer: so, for them and their peers, tech was highly valued, and going to college was highly valued - the expectation was that you'd do it!
• stark differences for many students who moved during their childhood, with different emphasis placed on technological education, "traditional" education, and sports. [note from Matt: this was a big chunk of the conversation, but it's hard to do this without leaking info about each student!]
• in IB schools, taking computer science is awkward with the rest of the curriculum.
• in affluent areas in Washington, some schools had many, many CS classes (including cybersecurity and robotics)! Perhaps influenced by the demographics of the parents, more of which work in tech. Yet, some of our students still didn't take them!

• book mentions "technologically rich, educationally poor". The government has funded programs, but have relied on (new) organizations like Code.org to build out courses and pedagogy.
• Scratch is much more popular and is frequently used to teach kids programming!
• the field of computer science has skyrocketed (especially with social media), with a bigger emphasis on it being a great career, having prestige, and being put on a pedestal
• everything is now digital: menus, QR Codes, even Disneyland! But, access to technology has not improved proportionally, which widens the technological gap.
• not having computer skills puts a harsher limit on your socioeconomic status.
• separate from access, teachers are now more overworked, burned out, and (sometimes) hate their jobs! Giving them more technology will not solve the problem by itself - broader reform is needed.
• accessibility has two different meanings in computer science, but both are related. Connection to CSE 121 reflection TED talk on "CS Unplugged": how to teach CS on pen and paper.

• argument: Word is interacting with software, but CS is more about writing something out to accomplish a function/goal
• counterpoint: Excel does this too! You can write formulas that do most things Java methods can do (including 121 P3). Does this mean that Excel is the most used programming language in the world?
• in "Stuck in the Shallow End", there is a specific difference between computer literacy (using computers) and computer science (thinking about computers).
• everybody should learn computer literacy, it should be a mandatory part of K-12.
• but, computational thinking - and computer science - is not so different from other sciences (e.g. learning Java is like designing a lab, analyzing results - just with different tools). So, can think of this as adding another "science" (similar to biology or physics).
• computer science is not just programming - there's also an element of how computers work. For example, in AP CS Principles (AP CSP), you learn how wifi works!
• depending on how deep you go into Excel, it could be computer science? But, literacy is more about font sizes and solving "simple" problems. CSE 121's problems take hours to dive in to.
• CSE classes are about demystifying computers. Learning to use Excel is like learning to label pictures, while CS classes are like describing them in your own words and analyzing their meaning. CS is learning how these computing tools work and how you apply them in your lives.

• if we do, we need to improve access to technology (via education reform) - you need a system that guarantees access to computers for all students.
• personally regret not taking CS, even though the high school had it. Avoided it because of the association that "only smart kids take it", and if it was required, could have dispelled that notion.
• completely agree: same experience in high school, and only realized after doing 121. While this doesn't "fix" the perception/insecurity issue fully, it helps quite a bit!
• had Scratch in middle school - this felt like a good balance, sicne it wasn't a heavy requirement but still left time for exploration.
• if we add another requirement, could delay graduation more. Some students already struggle to graduate because they can't pass their required classes - this would make it worse!
• earlier exposure in middle school (with simpler content and lower stakes) may be helpful.
• also, many bad experiences with high school CS teachers - perhaps mandatory requirements induce more demand for good teachers?
• or, we'd run out of good CS teachers - and we'd have a big problem!!
• requiring CS feels idealistic - would be great if we had infinite resources, but we don't.
• if the requirement is "just have a CS class", that's too broad! Especially since school curricula are often very, very specific items with learning objectives. What would these be?
• one model: don't require it, but let it fulfill a "bundle" like math or a science (instead of physics or biology), and make sure that it hits the same learning objectives.
• going back to "Stuck in the Shallow End": students didn't take CS then because they didn't know what it was. We have the same problem now, and requiring it could help fix this!
• if CS is optional, could result in less emphasis on making sure that all students can access a computer.
• feels related to "test-optional" applications, which slightly backfired (by exacerbating existing equity gaps).

• the Xbox Adaptive Controller: did not grow up with video game consoles, and didn't think about the implications this had on disabled gamers.
• had not heard of the difference between subtitles and captions before the reading, but definitely makes sense in retrospect.
• shocked to hear that people deliberately give others seizures via flashing images, and sad that research is needed to defend against this.
• reading ability was a surprising but important one. It was jarring to see how out-of-date school standards can be.
• the adaptive controller had many different use-cases depending on the relevant disability (e.g. pressing the buttons with many different body parts), and that seems related to broader ideas when designing for accessibility.
• was familiar with the idea of switches, but had not realized that breath-controlled devices were switches (and is curious to learn more)!

• short, concise text can be very helpful to those who have ADHD or other attention deficit disabilities.
• outside of disability, it's very helpful to those who don't speak English as a first language - e.g. in helping grandparents understand documents.
• important for things like taxes and bank statements to be readable by others in simple language. But, this is different from academia, where more nomenclature might be needed.
• similar to other ADA standards (e.g. all buildings need to be built to specific codes), it would be great if all government documents, bank statements, etc. were readable.
• also similar to conversations around absurdly long terms & conditions for apps (you shouldn't need a law degree to understand the terms & conditions)!
• the argument of "this is just learning English" isn't effective - there's a difference between literature analysis and finishing day-to-day tasks.

• a product that designers make that try to solve a problem that may not actually exist, not correctly solve the problem, or overlook elements of that disability.
• when it's created, it often doesn't take into account the experiences of those who face those issues.
• learning about user research in INFO 200 right now! However, one common pitfall is when designers come in with preconceptions on how users would use their work, and only really ask questions based on the designer's own experience (rather than the user) - similar to this!
• feels like those creating disability dongles are trying to compensate for guilt/pity towards disabled people - instead of properly putting themselves in other people's shoes, and not getting the input from actual users.
• favorite example was the stair-climbing wheelchair: why not install a ramp (instead of charging people thousands of dollars)!
• instead of making something accessible in the first place, disability dongles are almost an "add-on" (that you have to pay for), which is not useful!
• real dongles are converters which exist when hardware doesn't have the necessary plug. Instead, we should try to standardize things so that a dongle isn't necessary.
• as an aside, nobody enjoys using dongles!!

• when solving a problem, figure out if there is a systematic solution or not
• one example: "smart" glasses for deaf people that display captions on the lenses sound interesting. But, this puts the burden on deaf people. Instead, systems like closed captions or sign translators fix these systemic issues, rather than forcing deaf people to help themselves.
• currently taking HCDE 315 (inclusive design) which focuses on disability as a mismatch with how the world is designed. It's the world that we should change!
• too easy to slip into "let's build technology" and build something cool, rather than actually helping people.

• AI feels too much like a buzzword - e.g. AI rice cookers?
• but, it's certainly helpful - e.g. with automatic captioning. However, it's still unreliable/inaccurate and may not solve the problem.
• in a way, feels like a dongle or a bandaid fix.
• going back to the plain text readability project: you could have a crowdsourced database that converts complicated text into plaintext. AI could do some of this, but it can miss a lot of nuance. Useful as a tool, but not the solution.
• what are we letting AI learn off of? Machine learning could be learning data from exclusionary sources or replicate biases.
• and, generative AI tools are still inaccurate!

• reasonable as an only option - better than nothing. But, not a permanent solution.
• feels like a constant "imperfect solution with issues" versus "is this perpetuating the problem" debate.
• AI is not at a place where we can trust it to perform important jobs, but AI is in a "beta" phase - and we need to collect feedback!
• throwing AI at things without monitoring its use "feels weird", e.g. image-generating AIs could be trained on other images generated by AI, creating nightmare fuel
• can't discount how big of an impact this could be: imagine if someone could select and copy-paste a paragraph into ChatGPT, and ask "summarize this concisely" or "translate this into another language" - that's a huge deal!
• need to be careful of utilitarian thinking - can be easy to fall down the rabbit hole of "if it doesn't work for everybody, we can't use it". Okay to use things even if there are some downsides.
• we should help develop what people are using - so if disabled people are using ChatGPT, we should work to make that better!

In-person, we discussed:

• implementing some legislation forcing websites to be accessible (e.g. requiring all images to have captions)
• but, people could just write "image", or add some well-intentioned (but bad) alt text.
• what would you do about legacy websites?
• creating software or languages that make it easier to be accessible (e.g. describing images by default, making it easily navigable)
• related solution: can we feed pre-existing structure from things like markdown into a website?
• teaching accessibility somewhere. But, where? Last week, we saw there are many tough questions (what age, what program, should it be required)?
• should be similar to how we teach civil engineers or architects ADA guidelines - through accreditation. Why not add it as a checkmark for a CS degree?
• for K-12, we should teach it somewhere, but maybe not the technical details (are students building websites)?
• K-12 doesn't immediately solve the problem, since you need to convince many CEOs and VPs. How long would it take for kids now to get to these positions of power?
• need some sort of workforce education; could be training programs or through representation.
• when onboarding as a research assistant, had to do mandatory modules about Title IX, OSHA, etc. - could do something similar.
• but, many people don't pay attention to mandatory trainings, and those people are the most important people to teach! You need to teach people to care!
• as a resident assistant, have to go through many accessibility trainings (over a week). But, they're quite effective - in part because they weren't just tacked on.
• in addition, there were tangible outcomes - e.g. if a poster wasn't accessible, it would be denied.
• could require government agencies to have accessible websites through law, and incentivize private companies to make accessibility a priority through tax breaks.

I also asked us all to answer this question on paper as an "exit ticket". Here are a few of the answers that touch on ideas we didn't talk about in-person.

• on a personal level, taking more initiative to add alt-text and other screen-reader friendly items for individual creations.
• teach about accessibility from a very early age - so that thinking about how to make the world accessible is something we all think about, in everything that we do. Almost making it the default!
• avoid reinventing the wheel - instead, first ask if the design is needed!
• Wordplay!

• The video uses a very simple example (3 dimensions instead of 120000), so it's not clear if computers are actually subtracting gender to assume that things are feminine or masculine.
• Using gender could lead to the wrong context: for example, "king - man = queen" ignores that queen doesn't always refer to royalty.
• This adding/subtracting directions feels a bit binary.
• You could have issues where historical biases creep in: e.g. maybe "president" is assigned a male connotation since the U.S. has not had a female president, so "subtracting" female could give you vice-president or first lady?
• In some languages (e.g. French or Spanish), nouns themselves have genders (and this is a core part of the grammar). How does this work?

• Algorithms & humans are actually quite similar: these algorithms are human-made. So, if the person who makes it has a subconscious bias, than the algorithm will too.
• Throughout U.S. history, policing has always been racist - so we can't just "fix the data" when all of the data is the problem. Even if we remove explicit categories (e.g. "race"), there are many proxies (e.g. see redlining and low-income housing).
• Algorithms are controlled, tailored, and have lots of nuance; they can be more explainable than humans (since you can't pry open a human's brain). So, a more realistic solution could be fixing our current algorithms.
• Change in human behaviour is very hard, especially since biases can be integrated into your environment (e.g. parents, school, where you're from). Also noticed that the locations discussed in the ProPublica article are more likely to be historically racist.
• The questions that COMPAS asks are inherently biased - is it even possible to tweak questions like "have your parents gone to jail?" to not have racial undertones?
• You could try to weigh algorithsm differently, but it's a delicate balance. For example, Google's Gemini tool went way too far in the opposite direction. And swings in either direction have big impacts on people's lives.
• Fundamental problem is that humans are prejudiced and we need to fix that within humans. But, since governments and companies are going to use these tools, we need to fix them anyways.
• In order to debias algorithms, you need completely unbiased data and people. But that's not possible - this is all systemic!
• Slight disagreement with previous point - do you need unbiased people to recognize and fix bias in data? We're all biased, and yet can see issues with current systems.
• Would like to see AI with a human in the loop - AI cannot have historical context or understand why data has different outcomes, but a human could!
• Related to our education discussion: colleges have essays because we don't think just an SAT/ACT score fully tells someone's story. Algorithms can't truly understand language, systemic oppression, or context!

• In everday life, we use lots of things we don't understand (like our brains or our phones). This should be no different! (but of course, like the brain, we should still try to understand it.)
• What's most important is if they help us - and in this case, these do, so we should use them.
• Tying back to computing education - those who understand computer science (and LLMs) have much more power and can make these decisions on behalf of many other people. So, we need to think about who has this knowledge (and what biases that reflects).
• If you understand that ChatGPT is a black box and can make mistakes (and is not an infallible god), it would be good to use. However, not sure if this is how people really view it.
• There is some accountability - e.g. the recent "Lying AI Chatbot" case with Air Canada.

• missing from the article: the physical cost of making the computers, like raw minerals and labour
• many of the raw minerals come from conflict zones (e.g. cobalt in the DRC)
• land use! (to host the data centers and servers)
• people who do not use these technologies still have to pay for them - e.g. people in the DRC may not even be using ChatGPT, and are being disprivileged of their own land and resources.
• climate change is regressive (i.e. marginalized communites get affected more). For example, unhoused people are disproportionately affected by climate change.
• also, climate change affects everyone! literally, everyone!

These datasets are gigantic, within a few orders of magnitude of "the entire internet". For proprietary models (like GPT-4), it's hard to get an exact number (since OpenAI has chosen not to disclose this). For GPT-3 (its precursor), OpenAI's 2020 paper says roughly 500 billion tokens from the Common Crawl dataset, various public domain books, and Wikipedia. OLMo, an open-source LLM developed in part by people at UW, uses the 3 trillion token dataset called Dolma.

Many of these datasets combine archives of written books and content produced on the internet (often gathered through "web scraping"). The internet is famously toxic, and many AI models can replicate this toxicity - Microsoft's Tay was an infamous example that was shut down within 24 hours of being released. One of the most famous papers studying this phenomena (RealToxicityPrompts) is from CSE professors at UW!

Data sources certainly create biases in machine learning models! "Bias" can mean all sorts of things - from racism, sexism, and ableism to working better for certain human languages or opinions the model may repeat. It's hard to summarize the history of bias in ML in one paragraph (or even one article), but Joy Buolamwini's TED Talk is a common entrypoint.

Avoiding bias is very challenging, and broadly speaking is an open problem in machine learning (and in computer science). To summarize the lay of the land: "just get unbiased data" is typically not feasible, and may not itself resolve the problem. Many researchers work in a subfield of AI called fair machine learning (and related ideas, such as "responsible AI"). Some approaches are purely technical (can we quantify "fairness"? can we then optimize for this metric), while others focus on transparency (e.g. "Model Cards"), representation, or regulation. Timnit Gebru (from the rabbit hole readings) is one of the leaders in this field!

Long story, short, "their best"! Like the video discusses, many of these models are based on probabilistic reasoning. Typically, models will still pick the most likely option (e.g. most likely token) rather than explicitly fail, and then continue chuggling along. Among other things, this explains how language models can output gibberish (if you give it inputs of things that don't exist)!

Great question! Generally speaking, there are two broad steps in machine learning (that involve different physical structures).

The first is "training", where the model tries to figure out the best weights (the linear algebra mentioned in the 3Blue1Brown video). Practically speaking, training is a bunch of math (in particular: matrix multiplication and some calculus + probability calculations). This is done on specialized computer chips that are really, really good at doing math; the most common example is a "Graphics Processing Unit" (GPU), which is particularly good at doing matrix multiplication (the fundamental operation for much of computer graphics and games). The exact data is not available for proprietary models (though Sam Altman has claimed that training GPT-4 was at least $100 million). Using OLMo as a proxy again, they trained their model twice: once on 1024 AMD MI250Xs (10k per chip), and once on 216 Nvidia A100s (very hard to actually buy, but also ~ 10k). Competition for buying graphics cards is fierce - this is one of the biggest "moats" established ML players have.

OLMo also publishes power consumption and carbon footprint estimates - they say that training used 239 MWh of energy, which is about the amount of energy generated by all hydroelectric dams in the Northwest United States (thanks Nathan Brunelle for finding this link!). They estimate training used about 70 tonnes of CO2, which is about 150000 miles of driving.

The second step is deploying the model - or in other words, letting people use it. Generally, this is done with servers across the country (and the world!) - every time a website is visited, a complicated set of algorithms figure out how to direct that query to a specific computer, that then tokenizes the input, does the matrix math, and then sends it back to the user. Cloud providers (such as Amazon Web Services, or AWS) own hundreds of datacenters that serve this exact purpose (and make boatloads of money). One of the most famous AWS regional datacenters is us-west-2, which is right beside us in Oregon! There is no explicitly public data on how many of these servers AWS owns, but almost every estimate puts it into the millions.

In short, not really (but it depends)! These models are so complicated (with billions of parameters), so it's not currently possible to truly "explain" what each of the individual pieces of the model are doing. This subfield of AI is broadly called "explainable" and "interpretable" AI, and is one of the most active fields of research.

However, researchers (and members of the public) can make broad explanations of how some parts of the model work - such as the explainer video that we just watched!

In short, this is really challenging! Very briefly (and reductively): people used to use hard rules to split tokens up into nouns, verbs, adjectives, and other "rules" of grammar. However, these tended to be very manual, error-prone, and not scalable (especially to other languages). Modern approaches often blend some of this domain-specific knowledge with statistical tools that try to "guess" what tokens are based on a set of data (in other words ... more machine learning). A famous (very technical) tutorial that gets quite close to the bleeding edge is Andrej Karpathy's Let's build the GPT Tokenizer.

In their reflection, a student recommended InfiniteCraft, which is a fun game that shows some of this in action. Neal also has a ton of other fun visualizations on his website!

I will mostly skip this question since it's not super relevant to our discussion, but long story short, it takes in a set of numbers, and "normalizes" them to all be between 0 and 1 (but keeping their relative "size" to each other). Why? In math, we typically define probabilities as being between 0 and 1 - so softmax lets us "shrink" large sets of numbers to be valid probabilities.

Long story short, models do lose context, overfit, and often cannot fit on cheap computers. So, lots of smart people work on this (and invent cool tricks to make it work). One of the most relevant ones (that you'll explore yourself in CSE 123) is "compression" - basically, treating the input from the user, weights, and other "big" things in the language model as things that we can put into a zip file and make smaller. If you stick around in 123, you'll learn about one of the classic ways of doing this, Huffman coding; modern linear algebra techniques like sparse matrix computation are extremely helpful.

Generally speaking, it isn't. Some existing laws that apply to technology (like intellectual property and copyright) may apply, but courts are in the process of deciding that just right now! The closest thing we have to American legislation is an executive order from the Biden administration from October 2023, but it's not yet clear how this will be enforced.

This was a great question! Unfortunately, after some digging, I couldn't find a reliable first-party source (many articles claim to know the answer, but looking through the citations left me unsatisfied).

Some things I could verify were:

• as of April 2024, OpenAI charges $30 for 1M input tokens and $60 for 1M output tokens via its GPT-4 API. The token:word ratio seems to be 4:3, so if you estimate that a query is about 50 words and a response is 250, you get a napkin math pricing of 1.4 cents per query.
• BLOOM, a comparable LLM, is estimated to generate 1.5g of CO2 per query (per a paper written by members of the BLOOM team
• estimations are quite a bit more complicated because it's hard to factor in upfront costs from purchasing hardware and needing to repair it. There are many more estimates with "amortized" costs; the controversial Nature paper The carbon emissions of writing and illustrating are lower for AI than for humans
may be of interest!