Awesome. Yeah, okay. Should we wait for the rest of the video?
Oh, for you. We could, yeah.
Oh, she just texted it.
I guess we can get this started if that's okay with you.
Sure.
Cool.
Cool.
Are we good?
Yeah.
Awesome.
Perfect.
You know, Todd Mintz, thank you so much for sitting down and talking with me.
Yeah, thank you for having me. Pleasure.
I guess my first kind of question is, how would you define AI?
So, Jane, I would define AI kind of broadly as the collection of systems that help with
predictive capabilities and making decisions.
So, historically, there have kind of been two ways of approaching this.
One is through machine learning, which has mainly been using pattern recognition to predict
what may happen in the future.
The second was sort of more what's called rule-based of trying to mimic kind of deductive reasoning
that people use.
Nowadays, most things fall into the machine learning category, especially some of the technology
that we see out there today, like chat, GPT, and those kinds of things.
And does that work primarily in classrooms and opportunities, festivities?
So, depends on the context, right?
So, in research, a lot of the focus has been on developing new techniques for machine learning
and how we can extract insights from data and forming prediction.
But the systems that are used from students are the products of these kinds of methods.
So, for instance, using chat GPT to help with writing.
Use these methods to create the kind of language model that's able to do word prediction.
Same thing with, you know, using next word prediction on your phone.
That's the mechanism behind it that kind of makes it operate.
And what kind of information or where do programs like chat GPT hold their information or data for them?
So, it's a great question.
It really depends on sort of the goal of the particular program.
So, with chat GPT specifically, used the corpus, right?
All kind of written word from the internet from, I believe, around 2019 or 2021.
So, it doesn't have any more up-to-date information than around that year.
But it's still a ton of information.
It's every single kind of written word and every single written article that OpenAI was able to access up to that point
that it used to train a model to figure out its language.
But, for example, other models that might do image recognition or something like mid-journey that takes prompts
and then turns them into photos, they need to have access to photo databases.
So, that's going to be completely different from the way you'd access or the written word database.
And depending on the application might be something else.
So, some of the stuff I work on, for instance, in, like, medical applications,
I go to, like, the hospitals or I go to medical collaborators and go through those approval processes
to make sure that I have, like, the relevant chart data or relevant MRIs and all that kind of thing
in order to form predictive models.
So, a lot of the information is still relative to the new if it's only from 2019.
So, what happens when information that is formed from is misinformation or information that is incorrect or accurate?
So, that's a great question and that's a big concern.
But, I think it's important to think about what is the goal of a system like chat GPT?
So, the goal of the system isn't necessarily to provide information in the form.
It's not a search engine, right? Like, we go on Google, we put in our query, we're looking for an answer to a specific question,
and we're hoping that we get recommended, you know, relevant articles.
Chat GPT is designed to give responses that fit well with the queries we provide.
So, it might not necessarily, so whether the information is correct or not,
might not necessarily be as big of an issue in the way it formulates its answer,
as would someone give an answer to a question that looks like this?
Now, the way that OpenAI obviously, like, deployed it, there's a lot of other different engineering kind of wizardry
that goes around it to make sure that it limits the amount of misinformation that it provides, right?
That it caveats, like, a lot of answers saying, like, oh, this might be beyond my knowledge.
But, really, the way to think about it is the system tries to produce an output that sounds good.
So, we said it limits itself at 2019.
You could ask it, you know, the results of, like, let's say, the 2022 midterm elections.
Even though it doesn't have access to that information, it'll give you some very plausibly sounding good response
about how those elections went, even though, for sure, it has no idea.
You can even ask it for, what's the weather outside?
And it'll give you a response where, oh, this is, if someone were responding to the question, what's the weather like outside,
this would be a response that someone could give, but it might have zero bearing on what's actually happening outside.
So, would you say it depends really on the user and what they want to do with the information and how they want to use the programs?
So, I would say a lot of it has to depend on the user. I think that's a very good observation.
But I think the, it's incumbent on the user to realize, kind of, like, the limitations of what the program is meant to do
and what the task you could do with it is.
But there is, I think this is actually, like, in the past week or the past two weeks,
there's been some recent developments where they've enabled real-time data collection to interface with systems like chat GBT,
for instance, like, being able to give it, like, camera sensors, and then, like, describe the room and, like, what's inside the room.
So, in principle, right, we could design these systems in the very near future with immediate sort of web access
and maybe allow them to, like, set queries to Google to find more relative information and then compile that into answers.
We're not quite there yet. And so, right now, right, it's understanding that the information is limited
and what the tasks that we could actually perform are with that.
So, programs like chat GBT are relatively new only being released last year.
Was there any reason for being released at a particular time like last year?
Or what was kind of the timeline of these programs coming out?
So, right, like, science advances as it does, right?
We need the right kind of technology. We need the right compute.
So, a lot of things had to come together to have something like chat GBT work out.
So, large language models have been around for a while in AI terms, which is, like, probably, like, you know, seven, six years.
But this idea of taking large amounts of language, turning them into representations that computers can understand
and then using them for prediction tasks, we've had some component of that for a while.
So, there have been lots of systems, for instance, trying to figure out, if I look at movie reviews
and I want to figure out, do more critics like or dislike the movie.
So, this is kind of a test called sentiment analysis that use similar kind of language processing.
But notice, this takes data and processes it. It doesn't generate new responses.
There's a whole other side of things, which is the generative models part, which needed to advance of how I take models
and then use them to reproduce kind of new data that seems reasonable.
So, if you remember, even this kind of stuff has been around for a while, we have these deep dream kind of pictures, right,
where you took, you know, classic painting and you said, all right, repaint this with only dog faces
and you have a thing inserting the dog faces. This was like the beginning of generative models.
So, what sort of had to happen was a convergence of these two fields of the large language models,
the generative models coming together and a lot of clever engineering work to make it
so that this can actually be deployed with the given computer resources to make it accessible.
But I think the biggest thing that CHAT GPT had gone for it,
because GPT models have been around for a while. So, the GPT that's currently free is GPT 3.5
and OpenAI has GPT 4, because it's the fourth version of the GPT model.
But, so, let me get back on track here.
Even though these things have been here for a while,
I'm losing my spot here for a second with what I was going with this.
Yeah, can you rephrase the question? Yeah, sorry.
Was there kind of a reason that this program was?
Oh, yeah, thanks. So, right. So, the big thing that changed was actually the CHAT part.
So, the fact that there's a user interface where you can interact with natural language with a GPT model,
that sort of made it really accessible and that kind of democratized access to GPT.
So, even though GPT 1, 2, 3, 4, well, 4 has only been around recently,
but 1, 2, 3 have been around for a while.
Really, you can only deal with them if you had, like, in our computer science degree
and, like, wanted to give command line input,
but the fact that there is a user interface that allows you to chat
is what has, I think, invigorated interest in that design work really kind of elevated it.
Right, absolutely. So, what has that experience been like for you, with your students,
how do students use AI, either in the classroom or out of the classroom,
for assignments in the classes where the high students use AI?
So, sometimes they use AI because I assign them to use AI.
So, like, when I teach, by supply chain class, for example,
and I ask them to do, like, forecasting for, like, demand,
they have to use these machine learning and AI tools in order to do forecasting.
But something that's been really useful is using GPT for editing different types of documents,
be it, you know, I have students, for example, that speak English, like, as a second language.
I myself speak English as a second language. There's some things that don't come naturally
if you're not a native speaker. GPT comes super handy if you give it, you know, a prompt,
you're like, hey, can you edit this to make it sound like more natural?
Or can you edit this to make it sound kind of more concise?
I feel like this is over or long. And it's been really, really great in, basically,
shortening the amount of times of, like, writing certain papers or even writing emails
and responses, making them sound more professional and assisting with communication that way.
Those are the two kind of big things I see.
Absolutely.
Kind of going with that, how, in your experience, have students use AI for academic conduct
or using AI to kind of write essays for them when that wasn't particularly assigned?
So I haven't seen that yet. I don't assign a lot of essays in my classes,
it's an engineering class, and, thankfully, well, not thankfully, I guess,
we'll see whenever this bridge is gapped. But, like, the shared GPT style models, GPT style models,
in general, are not very good at mathematical reasoning, which is very interesting.
So they can reproduce grammar very well. But if you start asking it to solve different math problems
or provide you with proofs for different mathematical propositions,
you're going to get probably nonsense out of it unless it's something like very, very simple
and it's been reproduced thousands and thousands of times. So we're not quite there yet.
But something that would be interesting in the future, and something we've discussed here in the department,
is we teach, you know, both undergrads and PhD students.
We're supposed to be, like, future experts in these things.
Instead of just asking them to reproduce sort of classic results or learn classic material,
presenting them with a proof or a paper that's been created using a generative model
and asking them to find kind of errors in that paper.
Because that's probably going to be where a lot of impact comes in the future,
is as this stuff gets more common, gets sent to peer review journal,
we need reviewers and we need scientists that are on the watch to make sure that things are up to snuff.
So in your opinion, what's the best way to integrate AI safely in whatever how it's being used for,
whether that's the classroom or for personal use for people?
So I don't, I think the best way to think about it is that there's not one single standard for safety.
I think the important thing to consider is, depending on the situation, depending on context,
who the system's interacting with and who's deploying it, what is the measure of safety.
In the classroom, I think we have a relatively straight, more or less straightforward bar to clear here.
If we use it in terms of language editing tools or use it to aid with work processing, I think that's great.
The introduction of Microsoft Word didn't diminish academic work from writing things by hand,
and I think this is just kind of the next step of that.
But other concerns that some folks might have maybe in terms of less savory language
or including different kind of biases and assertions, that's something that we've got to be more on the watch for,
and that's just something that you have to sort of inform the students about it occurring.
And if anything, the incorporation of these tools might actually aid in detecting and stopping plagiarism
because there's a strong tendency of chat GBT to insert citations that don't exist into work.
So it's very easy to find misconduct if the students start citing papers that don't exist,
but that might also make someone a bit more cognizant of like,
oh, okay, this is why I need to make sure that I actually read through this and operate thoughtfully.
So how can people who hope design or influence programs such as to make sure that the information that the program's home from is correct and accurate?
Or what can they do to make sure that steps of misinformation don't happen?
I think there's two different levels of the misinformation, right?
The one level is making sure the information comes in as accurate, and that's difficult.
So if you're trying to learn language, the more examples of language you have, the better.
And so you're not necessarily going to just moderate the content, right?
Just because something's fake news doesn't mean it's not written in proper English.
So that's more challenging, but the other side that's more actionable and needs to be watched out for is that
because chat GBT just reproduces good sounding answers,
it does show the level of like a college educated individual.
So we have this concept called like the Turing test, which is if you interact with like an AI model,
can you discern the difference between it being artificial intelligence versus it being human intelligence?
And so previous systems that have kind of passed this have been at the bar of like,
communicating as if you were like a 13 year old or like a teenager,
and chat GBT communicates at the level of someone with an undergrad degree.
So unless you're probably like a somewhat of an expert in what you're asking it about,
a lot of it sounds very reasonable, even though it might be completely off.
And that's the really challenging aspect.
So I don't think, so one thing that open eyes done, which I think is good is make kind of hard coded decisions
such that GPT or the chat GPT version caveats a lot of what it's saying.
I think that's a good initial step, but I think providing like an annotation system
or providing some kind of other process or more warnings or also public education
about the limitations of these systems might be what we need to avoid just someone
relying completely on generative information versus even going to like a search engine afterwards
to fact check what it is they're saying.
Right, and would you say to people who are more hesitant to using AI generally in either
type of daily life or in schools or people who are just kind of resistant to using it?
So I'm not sure.
I think it's going to be difficult to avoid.
I think there's too many good use cases of AI and different machine learning tools.
It's going to be hard.
It's going to fast approach like the level of avoiding using a computer,
avoiding using the internet, avoiding using electricity.
Everything from the way we interact with like social media on our phones
and everything from like the mail right being sent to us,
the way our credit scores are determined if you want to interact with any sort of institution
unfortunately you're going to have to interact with AI.
Right, so I think it's more of a making piece with the fact that it's there
trying to figure out maybe ways of being informed, being involved,
and kind of lobbying for ways of having additional ways of expressing both like
dissent and expressing concerns to aid in like the design of future models
and making sure that you have a stake in the way things evolve going forward.
And so what do you think the future of AI looks like right now in the near future
in six months period versus five or ten years?
Okay, so this is like always dangerous to make too far-reaching conclusion.
I mean the next six months we'll see.
Advances have been really, really strong with the way these GPT models have been going
especially because it seems that just by increasing the size of the model
and increasing the amount of data that we use for training,
that improves the capability of reproducing language very, very well.
But I think in like the near term what we're going to see is a lot of specializations
of generative models.
We're going to try and see them being deploying various aspects of think of like systems
that might annotate doctors notes in the medical field.
So some kind of generative model that's specialized for like medications
or generative models that are specialized for you know white collar office workers
to write quarterly reports, models that might be used you know.
I don't want to you know journalism is important right,
but think of like standard kind of articles that might be published about
sports results or like summarizing things from like Capitol Hill.
That might be something that is being worked.
I know that those systems have already been in development for a while,
but something like that is going to probably take more hold as time goes on.
In 10 years like who knows.
I think the field has been moving really, really fast.
Some things that have seemed promising ended up being put on the wayside.
Some things that didn't seem promising at all came back.
Like even neural networks which are sort of like the basic model that underlies
a lot of modern day machine learning, they weren't very invented in the 80s
and then they were kind of put in the back burner for almost 30 years.
Before we finally had enough computation problems,
computation power to bring it back and reform the field.
So it's very hard to say what's going to win out in the long term.
But I think we'll see a lot of these specialized tools.
And I think the other thing we'll see is a lot more discussion on what regulation
of AI looks like.
What is the who ownership of data and the relationship between the private consumer
as well as the companies and institutions that use these kinds of tools?
Is there anything else you'd like to add about AI or any use conceptions
that you'd want to clear up that generally people have about AI?
Yeah, I think the most important thing is we use terms like artificial intelligence
and it sounds kind of scary, right?
The image that I contour up is like, you know, Hal 9000 or like terminators,
you know, crazy stuff like that.
I think the important thing to remember at the end of the day,
it's not truly intelligent at the moment as far as we know.
These are just very capable tools that are useful.
They're designed with a purpose, they're designed by people,
and they have limitations.
It's just kind of like the next step in helping us improve, right?
Kind of like how we replaced horsepower with machine power.
We replace, you know, writing things on clay tablets with word processors
and well typewriters than word processors.
This is simply the next step of taking some of these additional tests
that, you know, required us to do before more manually.
We can now automate it in new tools that enable us to do that.
Yeah.
Is that hyperbole?
So I would err on the side of, yes.
But I think it's like caveat it, right?
Coated lead to our eventual demise.
There is a possibility where that can happen, okay?
But I think looking at what might happen in like 100 or 1000 years is important.
I think there's like a few more close term ethical concerns that we need to think about.
So, for example, if we start having integration of autonomous vehicles, right?
What are the justifications we're going to have for preserving the life of the driver
versus preserving lives of pedestrians or other people that are locked by?
As we deploy AI and prediction in like the medical field,
what are the ramifications of having diagnoses made by automated systems?
How does that look like when you integrate it with doctors?
Whose responsibility is it if the prediction is correct and the doctor decides to not go
with the AI prediction, right?
These things are a lot more immediate.
The other thing that's also a bit concerning to me is in the same testimony,
the founder of OpenAI was talking about how there's a need for more regulation
and the slowing down of the process.
And I think it's kind of important to know that he started kind of,
this feels like a little bit like potential regulatory capture in the sense that
as soon as Chad GPT proves to be the generative model that kind of wins out in the market,
that's when OpenAI is more clamoring for that,
but it's obviously hard to say, right?
Like what's his going in there?
Like there are a lot of like legitimate concerns.
But still weighing the democratization of like developing this technology
versus the safety risks of making it too open ended,
that's something that's going to have to be like deliberated
and that's something that's going to have to discuss.
It's a lot more near-term than the existential risk aspect.
Sure.
Is there such a thing as going too fast?
I guess, I guess there might be.
But it's hard to judge the speed of advancement,
like what's incremental versus like what's a paradigm shift, right?
You don't know that you've gone too fast until you've actually gotten there, right?
It's hard to see in between.
Like I said, right, to get to Chad GPT having like global influence
and causing everyone to be interested in AI,
there had to be hundreds of small advancements along the way
before we got there, right?
No one cared really other than some people on Reddit and some people online
about GPT-2 and like the employees at OpenAI,
but now that GPT 3.5 has a chat feature, all of a sudden,
we realize where things are.
So hard, it's hard to judge.
Yeah, of course, no problem.