Why AI Should Be Taught to Know Its Limits
17:43 Share

Amazon q business is the generate A I assistant from A W S, because business can be slow, like waiting through mud, but amazon q helps streamline work, so tasks like some mary zing monthly results can be done in no time. Learn what amazon q business can at a hey future of .

everything listeners a ick know before we get into today is we've made a change. We're bringing you even more of the original reporting and interviews you've come to expect. Check us out weekly on fridays and let us snow, you think drop a line to every we podcast at W S J dot com. For at least as long as we've had science fiction stories, people have dreamed of a day when artificial intelligence work alongside humans to make life Better. Just think about how nine thousand, the starship Operating A I from Stanley kubrick's classic film two thousand, one a space out to see the .

nine thousand series is the most reliable computer ever made. No nine thousand computer has ever made a mistake or distorted information. We are all, by any practical definition of the words, foolproof and incapable of error.

Of course, that's not exactly true. Spoiler alert how nine thousand turns on his crew and ends up being deactivated.

Your do. Luckily, the real life AI .

we have now are nowhere near as powerful, but that hasn't stopped users or the tech companies developing them from talking about these programs as if they are intelligent according to open a eye. Its ChatGPT program can already quote, see here and speak. And soon enough, it'll be able to do much more because we .

believe the AI is going to be a technological and societal revolution .

that's OpenAI cofounder and CEO sam altman speaking at the company's developers conference in november.

We will be able to do more to create more and a half more as intelligence can be integrated everywhere. We will all have superpowers on the demand.

But some I researchers say that way beyond what these programs are capable of.

As you use this chatbot, its answers are typically eloquent and IT seems to have facility on a topic so we are easily fooled into thinking that to be intelligence, it's not.

Usama fid is the executive director of northeastern universities institute for experiential artificial intelligence. He spent more than thirty years studying data and day eye at companies like microsoft and yahoo and founded the machine learning systems group at masses jet propulsion laboratory and fired, says, one way that A I can actually become Better is if people stop projecting ideas of intelligence and personhood onto these computer programs.

you have to be continuously vigilant saying, okay, this thing is trying to help me come up with the best auto complete. But IT is my responsibility to read through IT very carefully and check IT .

from the wall street journal. This is the future of everything. I'm danny lose. I spoke with the sum of field about how small things like calling A I errors, pollution can mislead people into thinking these programs are more powerful than they are. But he also says that recognizing those limitations can make A I Better will find out how, after the break, stay with us.

Amazon q business is the new generative A I assistant from A W S, because many tasks can make business slow, as if waiting through .

mod a help. Luckily.

there's a faster, easier, less messy choice. Amazon q can security understand your business data and use that knowledge to streamline task? Now you can summarize quarterly results or do complex analysis in no time.

Q, got this. Learn what amazon q business can do for you at A W S dot com flash. Learn more.

Usama, we hear a lot about artificial intelligence being trained using large language models that's like the GPT in ChatGPT. How does this training work? And what does the A I do with IT?

Well, you basically feed document, and you say, learn some patterns and IT looks at each sentence that in your election and IT says, okay, if I see the first part of this sentence, i'm guessing that the next word is whatever comes next in that sentence, right? And is doing this over millions or billions of such sentences? So you're coming up with the way to kind of echo back everything you saw in a form of very flexible, very sophisticated order complete.

Now important to mention this auto complete has no understanding of what's in the documents is doing IT without any understanding of what it's completing. The the technical term we use is sarcastic parrots. So they are parents because they don't understand what they say, and they are sarcastic and that they have some degree of renomme, because sometimes, I guess, the right words, sometimes they will guess completely the wrong world.

And then the way generate vi works is whatever IT guesses IT assumes is the truth. IT uses IT as a part of the input now, and IT builds on top of IT. So this is how these large language model get errors.

So a lot of people talk about how training data is a lot of the big issue because if there's mistakes or you know biases and baLances very different depending on who you're talking to. But is that really the biggest issue with A I right now?

It's an important question. One of the big issue is here is, is your data baLanced, is IT complete enough? Doesn't represent reality? Did you drain the model enough on IT so that I can reflect IT accurately? The cost of creation of the data and balancing the data, uh, is super important.

And if you actually include the wrong data, and there if you include two documents that are look similar but reach contradictory conclusions, one ChatGPT is not gonna know which one to pick. It's judgement typically is whatever is more frequent must be the truth. So you can easily fool IT into believing stuff by just putting out a lot of stuff that I can process that repeats an untruth, and then suddenly that untruth becomes the truth.

And is that what people are are referring to when they, when they talk about A I hallucinating coron .

quote and what I just mentioned, which is the way IT works, fundamentally, this order complete as IT starts building out its answer, one word at a time. Once IT guesses the same word, it's off on their own engine.

So the computer is basically being the most annoyed guy in your seminar who's just shooting off from the mouth before he's actually thinking of anything.

What said is correct, doesn't think about what is saying. IT doesn't even understand what is saying. It's just basically saying, whenever I see this word, this is the likely next word and this is the likely next word. And by the way, I wants to believe myself as I see those words and all build on top of them. And that's how generate A I works.

What do you think about the fact that, you know, people refer to this whole process as fluctuation?

Now when you say hallucination, you're attributing too much to the model. You're basically saying it's intelligent, but everyone's in a lucene. This thing is not intelligent.

IT has no understanding of what is saying. So IT makes else and and that's the term. I like you.

I prefer the fact that we say, look, IT makes errors and part of its function is to make errors because it's guessing all the time, and sometimes, I guess is are wrong. And when that guess is wrong, IT will build the wrong thing. Now the one thing to remember is these models don't have intensions.

So IT doesn't intend to trick you is doing its best to auto complete on a prompt gave IT, which is why the failures becomes spectacular when you require part of your answer to be very precise. Like give me a reference right to this. Now if you trying to reconstruct reference based on auto complete from your memory and it's all approximated, you're not gonna give a great answer.

So what are the consequences if we don't address, uh, these holus instance flash errors to draw analogy .

from autonomists driving as we load into a sense of security and we sleep at the wheel? So as you use this chatbot, its answers are typically eloquent and IT seems to have facility on a topic so we are easily fooled into thinking that to be intelligence, it's not as you have to be continuously vigilant saying, okay, this thing is trying to help me come up with the best auto complete.

But IT is my responsibility to read through IT very carefully and check IT. And then when IT cites something, or I need a citation, I go to the research, pick up that and added there. Now, order complete is useful.

IT doesn't solve the world's problems, but it's very useful. IT can speed you up. And speed ups in the knowledge economy are economically very important.

According to the U. N. In in, in the richer countries, the knowledge economy is about half the economy, up to fifty five percent, depending on the estimates. Now if you have something that effects have the economy potentially, but even if I gave you a speed up for only one percent or five percent, you can't ignore it's economically significant.

It's half the economy and one percent of half the economy is huge, right? So we came up with the technology, which is not immune to errors, does a luCindy according to google and others. But despite that, it's actually helping accelerate many tasks that we invented for ourselves and knowledge economy. Speeding up the task of a knowledge worker has huge economic consequences and cannot be ignored.

Coming up, how could deep, mister fine, how A I works, make IT a more effective tool stick around?

Amazon q business is the new generative A I assistant from A W S, because many tasks can make business slow, as if waiting through .

mud a help. Luckily.

there's a faster, easier, less messy choice. Amazon q can securely understand your business data and use that knowledge to streamline task. Now you can summarize quarterly results or do complex analysis in no time.

Q got this. Learn what amazon q business can do for you at A W S dot com flash. Learn more.

Osama, we still haven't to dance public perceptions of A I because for a lot of people, IT kind of seems like IT works by magic.

Any technology that works really, really well starts looking like magic. Uh, google search, by the way. Um the engine the core engine that decides which pages are most relevant to your two point five key y word query, right how to rank uh in the background is a lot of work, there is a lot of human intervention. But because this company is don't talk about IT, uh you just think it's magic, by the way, ChatGPT, after you do all this training and you spend the millions of dollars on training in, you now have to have another cycle which open aee doesn't talk about of having humans, their own employees, contractors, chAllenge the model. This is called by tuning the learning session, which is very expensive, very human intensive.

How do you preserve what is really useful about these A I systems without having users do what I find myself doing a lot when i'm on the phone with my bank and i've got the automated day eye system talking to me and I am not getting the right answer and end up saying over and over again, talk to a person, talk to a person, talk to a person until I finally get to an Operator.

That's a great and central question. And in fact, in today's technology there is no other way than to figure out when to introduce human intervention, when should a human take over um and what effectively prompting the system to say, turn off and give me a human being right. Stop trying to help.

We need to get smarter about picking up those hints much faster and doing the hand off much faster precisely because once that A I or robotic system is stuck on a certain pattern, it's not going to understand. If you approach IT with a case that IT has never seen before, there is no way on earth. It's gonna even think of a novel solution.

IT cannot IT just cannot. Because all I can do is echo back what I saw in its training documents. And that's my definition stuff you saw before, which brings us to, uh, a lot of the issues were facing with autonomous driving the way we are trying to train vehicles today.

IT has to rely on having seen that situation before for you to know what the right action is. And the most dangerous part is when IT hits a situation that IT hasn't seen before, I wish IT has been enough to say, hey, I haven't seen a stop. No, IT actually gives you an output.

And that output in that instance is completely unpredictable. IT could be the right thing, like hit the break, or IT could be the wrong thing, like hit the gas or turn into this person or whatever. And that is where the danger comes in, is because it's not thinking, uh, I can't compete with a training session to a human teenager that might last fifty hours. And this human is now qualified to driver a car and will react against situations theyve never seen before reasonably well. The problem with these algorithms is, once they enter a space they haven't seen before, the output is completely unpredictable.

But you could say the same thing about people. You can say the same thing about A A driver idea comes running out of the woods or something while you're on the highway, just like, what do you do in that situation?

You can panic. But remember, somehow humans seem to leverage their other experiences to help nick decisions and help learn tasks. We somehow manage to generalize, come up with analogies, do what's called transfer learning. So I learned a move and chess. I can conceptualize the same thing applying to checkers.

So can you imagine any future world where A I is able to function without humans check in their work all the time? What's going to take for us to get there?

I think as we factor our work and our tasks into things that are highly robotic, highly predictable, um unlikely to go through a change. Yes, I can envision automation and in fact, that happens a lot of the time. Um I can also envision a world where we in view the A I with the ability to say weight, I haven't seen this before.

stop. This is why i'm very bullish on autonomists navigation for slow moving vehicles. If you have a garbage truck that's moving slowly that whenever IT faces a new situation, you can afford to stop and call for help, right?

That is very doable. If you're moving a high speed where you you can have the luxury of stopping and waiting for help, then you have to make autonomous decisions. And when you make those autonomous decisions, they may have catastrophic outputs.

So, uh, that's the difference is as we factor our tasks, I think a lot of things will get done by machines. We will figure out many tasks, especially in the knowledge economy, when it's safe to deploy these things. And we have guard rails and alarms that say, hey, I got ta stop. I haven't seen this before.

Who some of field is the executive director of the institute for experiential artificial intelligence at northeastern university. The future of everything is a production of the wall street journal. This episode was produced by me, danny Lewis and charlet garden burg. Like the show, tell your friends and leave us a five star review on your favorite platform. Thanks for listening.

Amazon q business is the new generative A I assistant from A W S, because many tasks can make business slow, as if waiting through .

mud help.

Luckily, there's a faster, easier, less messy choice. Amazon q can security understand your business data and use that knowledge to stream task? Now you can summarize quarterly results or do complex analysis in no time.

Q, got this. Learn what amazon q business can do for you at A W S dot com flash. Learn more.