Thesis Week 13 (Spring)


Spent the first hour of the week reviewing all of the AI failures that Dr. Yampolskiy sent me. This is actually really helpful, seeing what kind of AI failures are happening out there. Most of them happen “in the lab” or on purely virtual spaces, although a few with a physical component happen too. AI that informs medical decision making seems to be the most dangerous at the moment, there was a case where the algorithm for determining severity for assigning organ transplants assigns lower severity to black patients than white patients, all other inputs kept constant. The reflection of racial bias is an issue with the technology only in a small way, the greater issue is that our data contains racial bias of our society and individual decisions. Algorithms that ingest this then spit it back out is dangerous because 1. feedback loops could greatly increase racial bias past our current baseline 2. software and datasets are likely to be used for some time, causing biases to be perpetuated longer than they would otherwise 3. culturally, we assume computers are bias-less and objective, so output from a computer is going to be treated as either fact or bug, but not with some nuanced failure like racial bias. (3) is quickly changing, though. The will remain problems but cultural acknowledgement of fallibility of AI and AIs ability to reflect human biases will mitigate the damage.

Somehow this information has made me even less certain of what to write. Oh, there was a Roomba-adjacent robot failure. It sucked up a woman’s hair while she was sleeping. Sleeping or laying on the floor on a mat or futon is common in South Korea and Japan, and so this failure could only have been predicted by someone who is aware of that.

I’m going to list all the links he sent me, to get an idea of what failures are happening or being discussed in them.

So this did take an hour and half to do, but I have come out of it with an insight - humans are incredibly good at what I’m calling “contextualization”, and AI are incredibly bad at it even when they get very good at other human traits, good enough that we can create a shared illusion that Alexa is a person.

Also, perfecting a technology that has a flawed premise doesn’t make for good outcomes. Both emotional state and gender are not captured completely on a person’s face. Both require more information that a single photo can express. More so, gender isn’t predictable - the “assuming someone’s gender” issue doesn’t exist because we have imperfect vision but because we’re at a point where gender is internal and personal, so the only way to know it is to ask.

I have a feeling that as human-like AI moves from novelty -> online-only -> recommendation systems -> robots, we’re going to see more and more of these issues. The physical presence of “smart speakers” has introduced a whole slew of novel failures.

I’m out of time for today - I’ve got to quit an hour early to go get dose 2 of my covid vaccine! It’s a good day. Although, it sucks that I’ll probably be sick the rest of the week while I’m trying to work.


Got the vaccine. I don’t really feel sick, maybe a little bit more tired than usual, and a bit of a headache. I was up early today, I thought about coming in early but instead I took a shower. Work time. What to do. I feel smarter today than yesterday, but maybe that’s just because I haven’t started yet. Safety recommendations are the next thing on the list.

Wow, work is going really well today. I’ll have this section done today, then all I need is another few case studies and I’ll be ready to start editing.

I can’t believe I’ve come this far. I read over my paper and I was actually really happy with it. I also noticed a TON of small things to be edited, so editing is going to be quite a task (but a good one). I’ll be ready for the writing center on Thursday.

Back from lunch. I took a quick nap (30 minutes). I think the tiredness from the vaccine is getting to me. I made a glass of orange juice, it always helps me feel energized.

Done with that section. I think it’s basically done. Now just to come up with some more case studies. I don’t think anyone will mind if these are haphazardly thrown together - the heart of my system is a bunch of rough guesses and estimates coming together to make a slightly better picture than you start with. Reading the paper doesn’t so much as forecast your risk as it does introduce you to the big ideas of Perrow and others so you can see that some humans have been trying to figure out how not to destroy the world for a long time and have some pretty sound advice.

Time is up for today. Tomorrow I’ll pick back up at editing the case studies to be more readable. They also all need citations for news sources, where I’m drawing the information from. Tomorrow I should have the case studies done. I think 6 or 7 is a good amount of them, I can make one in about 30 minutes, since they are relatively simple and don’t require serious research.


I’m supposed to send the writing center a draft today by 12, so I’m going to do all the editing I can now, then add some case studies in the afternoon.

Back from lunch. I went for a jog. I’ve been feeling really distracted all day and it didn’t really help that much. I didn’t get very far in editing before it was nearly 12 and I had to submit my draft. Oh well. I’m going to try to piece together some more case studies. Actually, I’ll keep proofreading. That deadline might have been midnight instead of noon, so I should get a better draft so we can spend more effort on actual important errors tomorrow.

Got a bit interrupted, mom called and I made a dental appointment. I’m going to work an extra shift today so no big deal missing half an hour here.


Proofreading is going much slower than I expected, this paper is LONG.

I’m only on page 30, this goes on forever. It’s quitting time already, and I need to be ready for my singing lesson. And I’m getting a headache for no reason. Ugh.


I started doing final formatting today. I have a few paragraphs to write, the abstract and intro need some work, I need to do at least one more case study, and I still need to finish proofreading the whole thing. I’m going to work all day today, I have that big crunch energy.

Short break for lunch, I’m only going to take like a 20 minute lunch break today.

Just kidding. The avoidant stress got the better of me so I took a +15 minute lunch break. Back to work!

Okay it’s 2:55 and I am satisfied with the formatting. Also I realized I need a conclusion! Oops lol. I’m going to read the rest of my thesis out loud to finish proffreading. Starting at the bottom of page 12 at 3pm, I’ll check in and extrapolate when I’m expected to be finished.

It’s 3:25 and I’ve made it 3 pages. I have 36 pages to go (but they get sparser), so at this reate it’ll take 6 hours. I can speed up a little bit, too. So 4 hours. Oof. It soooo long. Yike.

Oh shoot, I still haven’t written “robustness to distributional change”. It only needs to be a few sentences and is doing little more than summarizing the section of the same name in that other paper.

I’m on page 18 at 3:56, so that’s only 6 pages in an hour. I keep having to track down citations, and writing a few sentences (barely a few, really) took way longer than I thought. I will speed up, though.

I took a break from 5 to 6. At 5 I was at page 30, with 20 pages to go. Those pages are mostly lists and tables though so I think the rest will only take an hour!

It’s 7:15 and I’m done proofreading! Time to eat dinner, then I’ll come back and write the intro and conclusion.

8:30, done with dinner. Time to write that intro! (oof)

Huh, so the intro was actually basically fine.

Drat, I finished everything else and I just can’t think of any more case studies. Eh, I have enough.

9:30, checking out. Final work count checks in at 13,000. Pretty good. My blog posts for the thesis total over 20,000 words, which is even more astounding.

Next Week