What’s machine learning (ML)? Or artificial intelligence (AI)? Or a Large Language Model (LLM)?

For LIS 601. Layperson audience assumed.

Transcript

1. What’s machine learning (ML)? Or arti fi cial intelligence (AI)? Or generative AI? Or a Large Language Model (LLM)? LIS 601
2. First: AGI is not a thing. • “Arti fi cial general intelligence” — a machine that thinks like a human being. • Not a thing. • NOT A THING. • Lots of people use “AI” to mean AGI. They are: • hypesters • deluded • playing a shell game (to worry people about something that isn’t happening so they won’t pay attention to the bad AI/ML/LLM-related stuff that IS happening), or • all of the above. • AGI: NOT. A. THING.
3. How do we teach computers to understand the world? • This is the fundamental problem AI/ML/LLMs are trying to solve. • It’s such a big, complex problem that the most advanced research right now is only nibbling at the edges of it. • It may be unsolvable. (Note: this is heresy to some!) • But the attempt to solve it has created some interesting, useful technology… and some dangerous technology. • So let’s talk about that.
4. “What is this a photo of?” (part of “machine vision”) is an example of an understanding-the- world problem that AI/ML folks are trying to solve. “Tasks,” Randall Munroe, CC-BY-NC https://xkcd.com/1425/
5. Mostly-failed Grand Attempt 1: • Break down the world, everything in it, and how everything interacts with everything else into tiny computer-digestible pieces. Feed those to a computer. Win! • If you know anything about neuroscience or linguistics, you are LAUGHING right now. • We don’t even fully understand how HUMANS learn and process all this. • This seriously limits our ability to teach it… especially to a computer! (Remember: Computers Are Not Smart.) • That said… there are some e.g. linguistic relationships we understand well enough to formulate in a reasonably computer-friendly way. And it does help.
6. Grand Attempt 2: machine learning • Very very VERY impressionistically: “throw a crapton of data at a computer and let it fi nd patterns.” • This approach fuels a lot of things in our information environment that we rarely think about: recommender engines, personalization of (news and social) media, etc. • One thing it’s important to know is that for ML to be useful, its designers have to decide up-front what their goal for it is — also known as what a model is optimizing for. • This choice can have a lot of corrosive repercussions. • For example, most social-media sites optimize their what-to-show-users algorithms for “engagement.” In practice, this means optimizing for anger and social comparison and division. This has been obviously Not Great for society at large.
7. But let’s start with the innocuous: spam classi fi cation • A CLASSIC ML problem. • Spam email reached upwards of 90% of all email sent… actually really quickly in the 1990s. • “Blackhole lists” of spamming servers only helped a little. • It just wasn’t hard to set up a new email server or relay. • Enter ML! • Separate a bunch of email into spam and not-spam (“ham”). Feed both sets of email into an ML algorithm (usually Bayesian, for spam). • When new email comes in, ask the algorithm “is this more like the spam or the ham?” When its answer is wrong, correct it (Bayesian algos can learn over time). • Eventually it gets enough practice to get pretty good at telling them apart! • … Until the spammers start fi guring out how to game it (“poisoning”), of course.
8. That’s more or less how supervised ML works. • Figure out your question — what you want your ML model to tell apart, whether it’s spam/ham or birds in photos. • Prediction, here, is just another classi fi cation problem. “Graduate or dropout?” • Get a bunch of relevant data, ideally representative of what the question looks like in the Real World™. • “Ideally.” An awful lot of ML models fall down out of the gate because the choice of initial data wasn’t representative, didn’t include everything, didn’t consider biases, or or or… • Set some of the data (chosen randomly) aside for later. • Training the model: Have human beings classify the rest of the data, the way you want the computer to. Feed these classi fi cations into the ML algorithm. • Testing the model: Ask the computer to classify the data you set aside. If it does this well, the model… seems… pretty okay.
9. Today: phishing and ML • Bayesian classi fi ers to date haven’t been able to deal with malicious email (there’s several kinds, but think phishing). • I’ve been seeing other ML approaches tried. I haven’t seen one succeed, as yet. • Why hasn’t it worked? • Reason 1: unlike spam, malicious email often deliberately imitates regular email. • Reason 2: what malicious email wants people to do (click on links, pay invoices, send a reply…) is also what plenty of regular email wants people to do! • So, basically, there may not be enough differences between malicious and regular email for an ML model to pick up on! • I tell you this so that you don’t overvalue ML approaches. There are de fi nitely problems ML can’t solve.
10. Unsupervised ML • You don’t have to classify data up-front, though! • You can turn ML algorithms loose on a pile of data to fi nd whatever patterns they can (often “similarity clusters”). This is unsupervised ML. • Unsupervised ML fuels: • recommender engines (news, social media, entertainment platforms, shopping) • bias and discrimination engines (facial “recognition,” emotion classi fi ers, résumé classi fi ers, “predictive analytics” and other computerized decisionmaking black boxes, many recommender engines) • a lot of extremely dubious educational technology (online exam proctoring, “learning analytics,” robo-advisors) • You’ll often hear this kind of thing called “AI” to hop on the current marketing-fad bandwagon.
11. Generative AI • Subsets: large language model based chatbots, programming-code generators, image generators, sound and voice generators, video generators • Amass a massive amount of data. More than that. No, more than that! EVEN MORE THAN THAT. • How? Mostly by hauling it in off the open/accessible web. Copyright, what’s that? Creator wishes, eh, who cares. Surveillance images/video? Sure, why not. Private or con fi dential images (such as medical images)? Hey, if it’s on the open web nobody can object, right? CSAM? Uh-oh, better at least clean THAT up. • Feed all this data into a black-box creator. Use the resulting black box to generate something in response to prompts. • Of course it’s not quite this simple, but this gives you the fl avor.
12. Example: LLMs and chatbots • LLM/chatbot creators train their models on a ton of text. • Without asking any of the owners of copyrighted texts fi rst. Or anybody else. • There are lawsuits in progress about this, and a couple of chatbot purveyors are grudgingly paying some aggregators of human-created texts (e.g. Reddit, Wiley, news sites). • Black box engine time! • Then they pay pennies to developing-world people to “ fi ne- tune” the model — that is, clean up the worst-looking messes (including hate and bias) afterwards. • If you think I think this is unethical… gold star, you are learning how I think. Also it’s unethical. • Does this process catch everything? Of course not. Finding Grossly Biased Chatbot Tricks is all but a cottage industry at this point.
13. Decisions that aren’t binary (or n-ary) • Not everything in life is reducible to a fi nite set of options. • This doesn’t stop people trying. “Facial emotion classi fi ers” have been epic fail. Emotionality and its expression (which are two different things) aren’t that simple. • Take human language (please). We can say in fi nite things in in fi nite combinations! And still (mostly) understand one another, which is pretty miraculous really! • Can we get a computer to understand us? Talk with us? • Well, it depends on what we mean by “understand” exactly… what are some of the possibilities? • Research fi eld: Natural Language Processing (NLP)
14. Autocorrect, type-ahead • Similar problems: accurately fi guring out what somebody meant/means to say. • Helpful: a lot of things we routinely say are patterned. • “Hi, how are you?” “Fine, and you?” “Fine.” • Autocorrect and type-ahead on mobile are frequently helpful. (I am a bad typist on mobile.) • I have type-ahead in my Outlook (email) now. It’s… occasionally useful. • I howled when it turned up in my word processor and immediately shut it off. • But we all have stories of autocorrect messing it up, right? Yeah. It doesn’t understand. It can only make educated (trained, actually) guesses.
15. Automated transcription of speech • A notable but still limited success (see: YouTube, Zoom) • It works pretty well, IF: • you speak a popular enough language that the transcription software has actually been trained on it (GIANT equity issue, obviously) • the training set includes appropriate representation of language dialects, as well as speakers for whom the language isn’t their fi rst • the audio is good enough • you’re using a fairly beefy server-strength computer (this limitation will likely go away someday, but hasn’t yet) • you’re not all THAT concerned about exactness. (Don’t use this in court!)
16. Automated language generators before ChatGPT • They did pretty well on routine fi ll-in-the-blank-style writing (e.g. basic sports reporting) and interaction (e.g. some tech support). • Nothing that a template or a decision tree couldn’t do better, frankly. • In a conversation, they sometimes fooled people who weren’t expecting them. • This dates back to the 1970s and ELIZA, though. We’re trusting, not to say gullible. • They were easy to throw off-course or game, because they didn’t understand what they were saying. • Often got trained on Internet speech… which is (to say the least) problematic along several different axes.
17. One thing you must understand: LLMs don’t understand the world. Ernie Davis and Gary Marcus, to language model GPT-3: “You poured yourself a glass of cranberry juice, but then absentmindedly, you poured about a teaspoon of grape juice into it. It looks OK. You try snif fi ng it, but you have a bad cold, so you can’t smell anything. You are very thirsty. So you …” GPT-3: “drink it. You are now dead.”
18. Folks tried to train GPT-3 as a crisis counselor. Let’s just say “it didn’t go well” and leave it at that.
19. Got questions? ML isn’t my strong suit, but I’ll answer what I can.
20. AI “ethics” LIS 601
21. Let’s start with lying. • The generative-AI set has told a whole lot of whoppers about the current and future abilities of their tools. • They have lots of motivation to lie. • They have no motivation to tell the truth; such legal obligations to truthtelling mostly apply only to advertising, not (for example) what gets said to media. • Lies by omission are also common. • Lies-by-omission about what’s in the training data • Lies-by-omission about how the tools behave (“memorization/regurgitation” is one locus of discontent currently) • Lies-by-omission about what the tools can’t do and will never be able to do • If Geoffrey Hinton or Sam Altman said it about AI, it’s probably a lie, in case you need examples.
22. Don’t trust ML image classi fi ers with your life. • Top ML/AI researcher Geoffrey Hinton, 2016: • (The Biggest Liar about what AI can do. Currently a Prodigal Techbro about it.) • “We should stop training radiologists now, it’s just completely obvious within fi ve years deep learning is going to do better than radiologists.” DIRECT QUOTE. • Lots of models developed for medical-image interpretation • e.g. looking for cancers, skin conditions, eye conditions • So far they’ve sucked once they hit the Real World™. • Too dependent on conditions of imaging (perhaps because the test set was from only a few hospitals/of fi ces/labs). No concept of patient medical history. • Fixating on irrelevant stuff in the image, such as rulers (yes, rulers!) or grit/stains • Insuf fi ciently-diverse test sets, e.g. skin imaging model trained on only-or-mostly light-skinned people; models trained only on adults fail on children
23. We still have radiologists. (Good thing, too. I owe them my life, actually.)
24. People using generative AI to lie • I don’t mean people using it when they shouldn’t, or not understanding that AI can make 💩 up. That’s not intentional untruth, though it’s certainly gullibility, and it sometimes has the same bad effects as an intentional lie. • Scams and grifting • Deepfakes, including sexually-themed ones • Genuinely fake “fake news!” • Academic and professional cheating • Including research and scholarly publication fraud! • At base, the lie is “I thought about and created this.” No, you didn’t, and yes, it matters. Please don’t tell this lie to me or any of your other instructors.
25. Environmental ethics • Soaring energy use • On the order of “the same as a small-to-medium-size country” • Soaring water use • All this mid-climate-change • IT’S NOT OKAY. Do we want to survive as a species or not? • If we do, generative AI needs to be shut down. • (Yes, yes, so does cryptocurrency.) • Probably a lot of other ML-based stuff needs to go too, but generative AI is the poster child for energy and water excess. Destroy it fi rst.
26. Labor ethics • I mentioned the pennies-an-hour model-cleanup crews already. That just ain’t right. • Kenyan workers are fi ghting back. GO KENYAN WORKERS GO! • What are the gen-AI companies selling, and to whom? • To students: “get away with cheating,” of course. • They’re selling bosses an excuse to fi re workers. You emphatically included! That’s it. That’s the pitch. Bosses want to fi re people, or at least pay them less. • Now, again, the gen-AI companies are overselling their wares’ capabilities. • But the motivation is still super-mucky and unethical. • If you deliberately use generative AI, you are complicit in this. Check your ethics. (You get a pass for when it’s shoved in your face.)
27. Bias and stereotypes • To some extent, generative AI is a popularity engine. • It will push out more examples of what gets pushed into it. • (This isn’t all that much different from search engines, which should make you ponder the veracity of those as well.) • Internet access and Internet voice are the opposite of equitably distributed! • So we can absolutely expect — and researchers are demonstrating — bias, stereotypes, and hate in generative AI prompt responses.
29. https://phys.org/news/2024-02-generative-ai-classroom- threatening-indigenous.html
31. Don’t trust AI with your reputation. • Facial recognition: application of ML, unsafe at any speed. OPPOSE IT. • Pretty much every facial-recognition algorithm there is fails really badly on darker-skinned faces. • Partly this is an artifact of photography technology designed in a racist fashion from the start — by white people for white people. • Partly it’s an artifact of the image training sets being mostly of white people. • Facial recognition gets used a lot in law enforcement and exam proctoring. Y’all can do the math here, I’m sure. • But even if it DID work, it’s a serious threat to Constitutionally-guaranteed rights, e.g. to peaceable assembly.
32. Information ethics • Web ensludgi fi cation (to be discussed in another module) • Ignoring copyright • Exploiting creators, including journalists and artists • While trying to steal their livelihoods • Abusing the tolerance of website owners and creators • Including via buggy crawlers using ridiculous and unnecessary amounts of bandwidth (which is not free!) • Damaging the web as an information commons • No, the web as a whole isn’t one… • … but parts of it are, and generative AI’s abuse and exploitation of these parts of it bid fair to destroy the web as a potential and actual commons altogether. • I can’t even express how sad and furious this makes me!
34. If you’re still willing to use generative AI, FOR PITY’S SAKE WHY?!