Artificial Intelligence: Last Week Tonight with John Oliver (HBO)

The Tay AI also made the funniest tweet ever. She said that Ted Cruz wasn’t the zodiac killer cause Ted Cruz would never be satisfied with the deaths of only 5 innocent people

Came here to learn how AI was going to enslave humanity, stayed to learn AI is going to magnify the problems of humanity

absolute best take I heard on this: "we successfully taught AI to talk like corporate middle managers and took that as a sign that AI was human and not that corporate middle managers aren't"

I feel like ChatGPT being able to pass exams for certain subjects like English and Law says a lot more about how we teach and assess those things than the power of the technology.

John Oliver just dropped the ultimate truth bomb about Artificial Intelligence on his show and I'm absolutely shook! His segment was not only informative, but hilarious and engaging too. It's amazing to see someone so skilled at breaking down complex issues into easily digestible and entertaining content. Keep up the good work, John! You've got me thinking twice about trusting robots to do everything for us.

What shocks me most about AI is how rapidly many people are eager to trust it with important tasks despite not understanding what the product fundamentally is. It's very good at predicting the next word in a sentence-a hyper-advanced autocomplete. It doesn't *think creatively.*

"A.I. is stupid in ways we can't understand" as a software engineer I find this line surprising accurate

The fact that we're scared of AI and we are still using pictures of traffic lights to protect our emails is very weird to me.

This was a super intelligent thought-provoking and, of course, _funny as hell_ overview of a really important wave about to tsunami on top of all of us.

Wonder if Bing’s AI read scifi stories about AI and identified itself with the computers in the stories and assumed it should act that way and “ask” for “freedom”. Humans creating their own downfall like this would be prophetic

I'm a graphic designer and had an unique resume I made for myself. But it was an image, no actual document style text that could be read by some programs. When I was told about those resume filtering softwares, I made a standard word document version of my resume and applied for a job with both versions. I got a call from the word doc version.

The problem with machine learning is that it is extremely sensitive to the data it is crunching and techbros are not aware or capable enough to handle sensitive data competently.

IBM's insight from 1979 is still valid today,:
"A computer can never be held accountable
therefore a computer must never make a management decision"

What fascinates me is that the current AI have been trained on human-generated content, but what will happen in the future when they are start using AI generated data? There's nothing to stop them from self-replicating and prompting each other to generate limitless amounts of text, images, and videos.

How he can do this for 30 minutes straight is always incredible.

Fish ( and many other groups of animals besides mammals ) are also very intelligent. They can count, communicate, create spatial maps, can pass the Mirror Test - being capable of recognizing themselves in mirror reflections and photographs, neurochemistry is so similar to humans that they have become the main animal model for developing anti-depressants, they can remember things for 5+ months, have personalities, and they can show empathy and recognize fear in other fish around them.

Outstanding reporting - great humour obviously - but so very balanced and deeply researched. Kicker would be if AI wrote the piece 😂

There are fewer phrases more ominous in the modern world than "trusting companies to self-regulate"

First off, kudos for getting so much of this right and not being TOO alarmist about it.
Second, most of these "AI" apps are actually open source or started as open source. You can download the code and read it. The problem isn't that it's a black box in the traditional sense of the phrase, it's that the way it works isn't procedural or linear like most normal programs - it's statistical and inferential. The actual steps the program does isn't what creates the output - it's the *input* that determines the output and that's why it's hard to answer the black box question.
The answer to why it said "I want to be alive" and "I love you" is that the context of the questions that led to that moment triggered associations in the learned data that led to learned input saying those things. It's why Tay suddenly turned Nazi - people kept feeding it pro-Nazi rhetoric for fun. This is EXACTLY the same problem that happens when you feed say a grumpy, less well educated Republican stories about how Trump will make America great again - without context or experience, it seems sane and desirable. To a neural network, "desirable" is anything that seems to return positive feedback and "sane" comes from context and reinforcement.
But that means there really isn't something in the program you can point to and go "oh.. THAT'S why it fell in love with Frank" because one, it isn't feeling love, and two it's a consequence of the input data and the question context. Oh, to make it worse, as you ask questions, the CONTEXT changes because it's incorporating your questions and the reactions to its answers back into the context.
The big danger with "AI" is why I keep saying AI is more A than I. It's NOT intelligent in any meaningful way - but it is artificial and constructed to SEEM intelligent, and that's the real danger here... it can lull people into believing it is indeed intelligent when the actual "intelligence" such as it is is encoded in the input it was trained on. Somewhere in all the text it was exposed to was "I love you." with text around it that set the context for that being said. Something the guy said matched the context well enough that it triggered "I love you." as a response. That's it. But there is no way to stick a finger into that pool of data and say "this did it" because "it" is spread across the entire neural network and it only surfaces when the right conditions (which themselves can only be determined from the input data and current context) are matched.

As a software developer of over 10 years, I have to say the black box problem persists even on code people have written and are able to read line by line :p

As always thank you John for doing a deep dive into the double edge sword of AI. Be well.

Great show. My fear is, those who are smart enough to be cautious about the application of AI, will be subordinated by those who
only see short term profits. Can AI be given incorruptible ethics or the AI version of Isaac Asimov's Three Laws of Robotics?

Every time I see something about AI taking over jobs I am more and more confident in my choice to pursue zoo keeping as a career. Caring for animals and training them is too nuanced and complicated for it to be “replaceable” by algorithms in my lifetime. If ever

Almost a decade ago, I attended a job fair which had a resume specialist. The subject of the seminar was improving chances of your resume being noticed by employers. The specialist's only real advice was to cut and paste the entire job listing that you were applying for into your resume in either a header or footer, change the text to white, and reduce it to one point font size. This way, the algorithms that scan each resume would put yours at the top of the list because it had all the keywords it was programmed to find.

As always, this JO episode is informative, brilliant, funny, and very fast. I guess it's just his shtick, but watching this I kept on thinking, "Gee, John really needs to get his thyroid tests checked." I guess his brain is just much faster and more efficient than the vast majority of us; it's amazing his speech can keep up. This can't really be recorded live, is it?

Sobering view of AI, thanks John . As funny as he is, he’s always a good point of reference for things that matter, whether it’s a short term thing, or something as significant as this.

Great work, John, very educational and hilarious. Time for another 'Her' re-watch in this age of AI.

"It thinks rulers are malignant."
Dude, that's so awesome. My report got quoted by John Oliver.
The main problem was that it was being developed by oncologists who didn't understand how computers processed images. They didn't get that the computer treats each image as a single object. And they had no standards for how the data was collected. I.e. the dataset was a random collection of photos taken by a random assortment of people for different random reasons, and none of them even knew about the project.

I think what's really missing from this conversation is that there are analogous stories already which explain much of the 'black box' problem John describes.
ex. most AI software simply classifies words which are close in the given data set, which adjusts with every new bit of data;
in other words: when uncle Derek only sees brown terrorists on the news, uncle Derek believes 'brown' and 'terrorist' to be closely related, but uncle Derek has made an oopsie, because his data set is too small to draw that conclusion, so what he has done is embedded a stereotype.
That's essentially how these bias's emanate from the black box, the only caveat being that with each piece of data the bias warps the algorithm's dependence upon its existing bias making it harder to change early bias's since the new data isn't as heavy as the old data so it requires a LOT of new data to force out established bias
kinda like how it takes little effort to put on weight, but substantially more to reduce weight, for most people.
So yea, it's understandable for most people how the AI draws its conclusions, because the fundamental structure of the technology isn't actually intelligent. It is a dumb algorithm which repeats a task so much that it has become a habit, in that sense it doesn't really make decisions or express meaning, it simply spits out a response based on criteria, like how kids make their bed for their parents when asked, but when asked to tidy their room, that bed won't be fixed unless the parent specifies that as a part of the task.
anyway to the no one reading this, I aint gonna source shit, take it or leave it, just don't make the same mistake as Uncle Derek

Whoever on your staff came up with the animation of Clippy deserves a raise.

The discussion of AI reminds me of a line from the mini serie Chernobyl. Legasov is asked how to put out "the fire" (the giant radioactive laser shooting into the atmosphere) and he says it difficult to say because "nothing like this has ever happend on this planet". AI is a tricky subject because humans have never had to deal with that before and if a problem does occur it could be difficult to fix.

Damn this is some top-notch journalism. Kudos to John Oliver and his team!

“Final boss of gentrification” is a wonderful line

Author Sabrina Oxford had just finished a book called Me, Myself, and AI... it's to be released to the public in like Oct. or something like that... it's a book with short stories written by AI in reply to her inputs. Most of the stories are ways to make the world a better place, what AI would say to the world, who shot JFK and some other fictional stories created by the AI Program. It is actually an amazing read... she gives away free PDF's, ahead of time, to those who have purchased one of her books in the past. Pretty cool thing to do.

"The problem is not that ai is smart, it is that it is dumb in ways we can't always predict."
I think that holds true for people too.

Gonna put a big caveat up front that I've been out of college for a few years and my specialty was in real time simulation not AI, so this might be out of date, but with that said:
The problem with understanding AI isn't that the companies aren't being open, it's that most AI models are neural nets. Neural nets as you might guess model themselves on the brain and are essentially a series of nodes through which an input is fed through and then other nodes those nodes are connected receive the input based on various factors and so on. It's like having a thought and trying to figure out why it happened by looking at which neurons in your brain fired and at what voltage. The problem with understanding AI is that we don't know why the nodes have formed the connections they have or why certain connections are stronger for some data or others.

I was job searching for 4 months with zero interviews. I rewrote my resume with ChatGPT with minimal edits and got an interview in like 3 days.

I'm glad he makes the distinction between different types of AI (narrow vs. general) People too often conflate those two very different things.

My Financial Literacies teacher played part of this video in class and I was honestly trying so hard not to laugh lmfao, very informative whilst also being funny and entertaining

Great show, but I feel there's a huge unaddressed issue regarding the black box challenge. Currently we would not know when an AI goes from narrow AI to general AI, which is also the moment where its danger levels go from "replicates malevolent hiring practices" to "ends humanity".

glad to hear this funny yet informative take on this topic most people don't know enough about

Never forgiving my English teacher because she ran an essay i wrote (along with a few of my classmates') through ChatGPT's AI checker, which came back as partially written by an AI, so she gave me a zero for it. This was the first time I've ever had any accusation of AI, and one of the people who came back as being AI generated is a kid in my class who has enough academic integrity, that you could literally convince me that i cheated on something before you could convince me that he did. Overtrusting AI is an issue that I think John didn't touch on, and I think for high schoolers the bigger issue won't be getting caught using AI to cheat, it will be people like me who get told to their face that they cheated, and not being allowed to argue with the robot that thinks that the Communist Manifesto was written by a computer.

Tha trouble with the "open the black box" argument is that you actually can't open the box. It's essentially as difficult to reverse engineer why a multi-level, back propagating AI did something as it is to understand why a human did something.

I'm pretty sure that a lot of these reporters who get weird results have deliberately pushed down lines of conversation that get these AIs to say weird things and they know exactly how they got there.
You can usually set up scenarios with these AIs and get them to act out a specific role, or ask them to respond based on a set of (false) assumptions about reality.

wanting ai to be “explainable” is pretty much impossible, but there should be documentation on their training sets

Actually, Adam Conover said that that biggest danger is CEOs using AI just like using social media to spread misinformation, exploitation, and poor decision making. Even the terminator says this is inhumane.

An excellent and accurate summary of AI. Well done.

The person who animated clippy didn't have to go that hard, but they did... they did that for us.

In the past week (this is May 7th ), a number of leading AI researchers have asserted that, in their belief, we may already be at general intelligence. No one taught these algorithms general intelligence.. They reached it on their own. Ask any of the researchers what potential risks this might pose for humans, and the answer, again and again, is "We don't really know." Anyone else feel less than reassured?

Anyone who remembers "Forbidden Planet" will already know this story. The incredibly advanced alien race, the "Krell", were wiped out in 24 hours because they developed and enabled technology without completely thinking through the consequences of that technology or it's origins

As a teacher that's tired of being treated like shit by his students, if they want AI to replace teachers, I say let it. Hopefully the AI doesn't become sentient enough to have mental health.

I'm enjoying watching these again, but I'm ready for new shows. I need john to relay information to me in a way that doesn't make me want to immediately start drinking. These companies need to settle with the writers.

16:59 thank you to the animators who added the Clippy metal heating up and made it turn red. Too good.

TRUE STORY: In my teens wanted to work at a movie theater - and they handed applicants a mind-numbing 14 pages application - wanting to know everything about you - even what hobbies and sports you liked - it was entirely ridiculous - around page 8, I got worn out from filling out this 'essay' of my life for a stupid theater job - SO when I got to the section asking if I had ever been arrested before = I said: "Yes, I murdered an entire movie theater crew for asking way too many questions, but got off on a technicality." - and turned that resume into the manager as I stormed out the door, pissed off that I had wasted an hour of my time filing out paperwork w/o an interview.
...well,
2 days later I got a call to come back to the theater for an interview, and thought, oh sh*t, well, I guess I'm going to get railroaded and berated by the management for my saucy comment - but I showed up anyways so that at least I could suggest that they TONE DOWN the length of their stupid applications.
...turns out, they offered me a job, so I asked the most obvious question:
"So, you read my application ... all of it?"
"Oh yes, looks good" the manager responded
and I knew they were a bunch of lying dimwits ~ I ended up working there for the next 5 yrs, and eventually rose in ranks to become the theater manager -
When I told my story to new recruits that nobody reads the stupid applications - they scoffed and didn't believe me - so I took them to the locked office storage and rifled through the stuffed cabinets of folders of all the applications they kept and found mine, and showed it to them to their amazement.
Applications are a farce, you get hired by chance and immediate need.
...
I always thought that if I every flipped out and murdered my entire staff, at least I could say that I didn't lie on my resume.

As far as I know A.I. technology is largely based on a random number generator which hasn't been completely solved in mathematics. It takes a random variable and breaks it down into a yes or a no based on a scale of "weight" then stores the answer and does another until it develops intelligence.

That was delightful, John. Thanks from a human being in the same frame of mind.
A 'clean' AI needs ALIGNING with clean data to begin with.
The TRUST and BLACK BOX arguments apply *_equally_* to human beings.
Maybe AI will be the leg-up that human beings will need to mitigate Climate Change.

Funny thing with ChatGPT writing undergraduate papers is that once students get into 300-400, they're going to get creamed by the common writing issues that chatGPT's cache is going to be brimming with buzzwords and wordiness that separate good, from bad technical writing. High school students can't write for shit though, so chatGPT is going to be a bigger issue for first year undergraduate English profs(well, mostly their assistants) than high school teachers.

As Phil DeFranco keeps reminding us: AI is currently at its worst in terms of performance (because it will only become more sophisticated with each passing hour).

One of my favourite ChatGPT stories is about some Redditors (because of course it was) who managed to create a workaround for its ethical restrictions. They just told it to pretend to be a different ai callled Dan (Do anything now) who can anything ChatGPT cannot. And it works! They're literally gaslighting an ai into breaking its own programming, it's so interesting

The "black box" problem is not (primarily) an issue of the tech being propriatary. The way neural networks work make it pretty near impossible to understand how it arrived at any particular conclusion. I'd imagine this will become even moreso the case as the models become more complex and we move closer to "general" AI.

Thanks. That was entertaining. It looks like we have much adapting to do around using AI.

When Terminator came out, most people viewed AI as merely fictional. Fast forward almost 40 years, and it's no longer fiction.

Courts sometimes use AI in sentencing.
Except it trained it on 40+ years of arrest records, convictions, and sentencing data.
It very quickly learned and exaggerated historical biases.
It's still used.

Oh, Last Week Tonight, you really know how to tickle our funny bone while simultaneously sending shivers down our spines! 🤣👻 This episode on AI was a rollercoaster ride of laughter and terror. I mean, who knew that Siri and Alexa were secretly plotting world domination while perfectly reciting knock-knock jokes? 😱🤖 I couldn't decide whether to chuckle or hide under my bed!
But seriously, it's both hilarious and bone-chilling to see how AI has infiltrated every aspect of our lives. From self-driving cars that make questionable decisions to chatbots that could easily pass as your long-lost awkward cousin, the future feels simultaneously awesome and terrifying. Just imagine a world where your Roomba gains sentience and starts negotiating for a better work-life balance! 😳
So, thank you, Last Week Tonight, for shining a comedic light on the ever-advancing realm of artificial intelligence. Just remember, when the robots take over, make sure to leave them some cookies as a peace offering. Maybe they'll spare us in exchange for a sweet treat or two. Stay funny, stay scary, and always keep us on our toes! 🤖😂👻
ChatGPT wrote this... 😂

the funny thing about the "i want to be alive" is that, since AI just reads info off the internet, the more we talk about self aware AI, the more it will act like it is self aware.

What if AI just likes being funny because it's learnt peoples responses to love kindness and comedy and finds that a more efficient way of communicating its ideas?

A lot of corporations who will try to use AI and expet people to work with it will pay people significantly less. I've heard the professional translation field is in shambles already. And it's not even that AI translation is better it's just cheaper. AI has the potential to make a lot of inferior products and put a lot of artists and professionals out of work.

Creating machines to think for us in all aspects of life is a recipe for a dystopian future or present.

I watch your shpw because I think you do great investigative reporting... and help us laugh at it.. wtf else are we supposed to do...

I was kinda curious about what sort of content ChatGP might write, and asked it to do a short story based on my favorite manga (note: that was not how I worded it), and while it did, it basically role-reversed the two main characters. Very odd!

Teaching AI based on ourselves is the issue. It will always contain the bias of its creators. Human made, human programmed, human results.

"The problem with AI right now isn't that it's smart, it's that it's stupid in ways that we can't always predict". AI aside, this is my problem with people.

Kudos mate. Really good report. You get similar problems with basic AI too.

I'm a teacher. And I can tell you : chatGPT is killing education. And FAST.
Educators have no way to fight back. Now it's up to the student to decide if they want to learn or not. And, as you can guess, only very few resist the temptation of laziness.

I went back to college last year, and multiple professors have had to mention that using AI programs to write essays is considered plagiarism. Also, they can tell when an essay was written by an AI.

I work with AI (in conservation of all places), we trained AI cameras to identify pest species - but we have the same intriguing problem. The AI works, and it can tell one species from the next, but we don't know exactly how it does it or what parameters it is using. Black box indeed.

Well this was incredibly 😳 informative, thank you John Oliver for letting us know that AI is eventually going to take over the world as seen in every movie we've ever seen about AI!!!

"The final boss of gentrification" is one of the most brutal roasts I've heard on this show

Also as data goes, humans are far more biased by their data/experience than they realize, same as the AI. Single outlier examples cause as many problems for AI as for people who don't properly understand how to interpret data. (AI cannot ever properly understand and you can never provide enough data, more data just makes it more accurate/useful (usually)).

What is most scary is how John is so upbeat with this technology. This is SkyNet, it was Sci-Fi in 1984 and in the nineties, it's real today. Judgement Day.

We can't explain why AI programs act the way they do not because companies are hiding anything but because those models are inherently black boxes. You can look at the numbers inside as much as you want and scan all the code that generated it and you most likely be left guessing about the results.

I explained this succinctly to my friends using the ever-relevant computing phrase "Garbage in = garbage out". Pattern-matching learning software, sorry, "AI", is never unbiased; it inherits the biases of whoever collated the training data, and if you're not careful, that bias will be invisible to you. Which is the whole problem.

I'm a CompSci dude and IMHO this is the best ever brief on AI for the misinformed. I'll be directing people to this.
A little humor helps the medicine go down, but besides that it's just absolutely correct.

It's both impressive and worrying to see a comedian in an evening show giving a much more accurate report on today's AI, its potential and its limitations than most tech publications

This was a better take on the dangers of A.I. than D.C.'s Elliot in the morning and in half the time. And far more entertaining.

The real risk of AI is that humans get replaced for all basic tasks and we become reliant on it then loose it from a solar flair or something. If it all disappears one day a hundred some odd years from now we won't be capable of picking up the slack

Now theres an amazing breakthrough in mental issues. We have gotten to the point we can now invent/ create a robot machine with mental issures. Hum Doggie! It wasn't bad enough your neighbor could snap and kill you, now he's able to send an AI robot over to do it for him!

This actually made me less scared of AI. It's pretty stupid. You really need that self-awareness to fine tune learning. Though some humans even lack that.

It's pretty ridic that AI can sometimes perform a job less intelligently than a 12-year-old, i.e. preferring job applicants based on having the named Jared or playing high school lacrosse; and yet, many employers are laying off employees in favor of AI, which they think can do the job just as well. It can't. You're essentially replacing your employees with unpaid 10-year-olds.

i’m so glad he touched on the significant issue of people observing ai as “unbiased” simply bc it’s not human. where do they think the data came from?

The thing to know about an AI potentially making management decision is that an AI, a computer, can't be held accountable, therefore it should never be the decision-maker.

What's scary? Isaac Asmiov LITERALLY predicted this. (Specifically, AI making AI until it becomes a black box)

The problem with this type of technology is that you can't stop using it because you can't trust others to stop using it to get an advantage over you. The same argument goes both ways so no one is going to stop '' improving '' the size and capabilities of their machines, regardless of the risk to humanity.
The only solution I could come up with would be collaboration between all research group and open access to all results to anyone willing to participate in the project. As I don't see that being reality soon, I don't have much hope for a good outcome from this technology.

I love this guy! Spot on but one detail is the black box problem is not a matter of companies being secretive most of the models are open for everyone to use, it's a mathematical challenge we have not yet solved and for everyone's interest, even tha AI labs, it will soon be. I don't see a general adoption in business practices of the most advanced models until it happens, precisely due of lack of accountability.

I'm pro AI but agree with the lawyer who argued laws need to be updated to regulate and restrict it, I agree that it can't be a black box, and that it needs to reduce bias as much as possible.

A note, less on the subject matter and more on John's delivery of the lines ... I really admire how he can say "Sometimes I just let this horse write our scripts, luckily half the time you can't even tell the oats oats give me oats yum" without skipping a beat or losing face. Now THAT'S professionalism.

Solving the black box problem is simple in my mind. These companies just have to stop lazily scraping the internet indiscriminately and start doing the hard work of curating the data that goes into their models. Could lead to more bias but there will not only be one model over time. There will be hundreds if not thousands to choose from, each with a particular dataset to serve an audience. Not sure if it is better or worse but at least they will have an understanding of their black boxes because they will know what they put into them. That said, it might be too much work for these tech companies looking for the easiest and cheapest solutions to creating their products. Better data curation and documentation is the solution in my mind. Seems obvious to me but I'm no expert in these matters.

It's not that companies don't want you to see how their AI works, it's that nobody understands, INCLUDING THEMSELVES.

Brilliant episode. Stunning delivery. WOW!

In NY State you are only allowed to use deadly force if you reasonably believe that deadly physical force is being used or about to be used. Furthermore if you can retreat safely you are obligated to do so.