Episode 36 – Killing as Few Humans as Possible

00:00:00.000 —> 00:00:02.000
Friends with Brews!

00:00:02.000 —> 00:00:08.000
Hi, I’m Scott, and this is, well, as you just heard, Friends with Brews!

00:00:08.000 —> 00:00:12.000
Today’s gonna be a little bit different. I don’t have a Peter Nikolaidis with me today,

00:00:12.000 —> 00:00:18.000
because Peter is doing security-related training, or he’s getting certainI don’t know what he’s doing.

00:00:18.000 —> 00:00:24.000
But anyway, yes, I understand that what you’re about to hear is going to be a solo podcast,

00:00:24.000 —> 00:00:29.000
and those are usually boring because what you get is some monotone guy droning on and on,

00:00:29.000 —> 00:00:33.800
And, you know, it sounds like either they’re reading something or just sounds like they

00:00:33.800 —> 00:00:39.240
are talking to themselves and nobody, no one, sounds interesting when they talk to themselves.

00:00:39.240 —> 00:00:45.400
Despite what some people I know think. It just doesn’t work. But anyway, I’m gonna try my best.

00:00:45.400 —> 00:00:48.120
I have some things to talk about. We’re just gonna carry on.

00:00:48.120 —> 00:00:52.600
Aren’t you going to introduce me? I’m here and I’m your friend. In fact, I’m your best friend.

00:00:52.600 —> 00:00:55.480
Hey, Peter, it’s you. I’m glad you’re here, man.

00:00:55.480 —> 00:00:56.440
Hi, Peter.

00:00:56.440 —> 00:00:57.080
Let’s record.

00:00:57.080 —> 00:01:03.080
I’m Neter Pikolaidis and I have stuff to talk about and things to drink. Okay, Neter Pikolaidis

00:01:03.080 —> 00:01:05.880
If that is your real name, I’m gonna start with my drink

00:01:05.880 —> 00:01:10.960
I have a coffee today and it’s water Avenue coffee, which we’ve talked about before on this podcast and

00:01:10.960 —> 00:01:14.000
It’s there a Portland roaster

00:01:14.000 —> 00:01:19.880
This one’s called river trip and it is a bit of a trip looking at the bag because it’s like blue and purple

00:01:19.880 —> 00:01:24.640
with yellow labels yellow on the label yellow back label and

00:01:25.840 —> 00:01:33.280
Their river trip coffee is a yearly. It’s a seasonal and they call it their yearly homage to summer

00:01:33.280 —> 00:01:40.800
they say a clean and bright cup that perfectly complements a morning sun peeking over canyon walls a combination of a

00:01:40.800 —> 00:01:45.720
Combination of seasonal fresh coffees from our longest standing partners in Guatemala and Africa

00:01:45.720 —> 00:01:53.320
The tasting notes are Peter you’ll be pleased to hear this orange macadamia nut and fudge sickle like chocolate

00:01:53.960 —> 00:01:58.080
And I gotta say okay, so on the back they’ve got a label that has a

00:01:58.080 —> 00:02:03.280
Roast level chart and it’s between medium and light but more towards the medium

00:02:03.280 —> 00:02:07.200
They’ve got a taste profile chart that get this it goes from classic

00:02:07.200 —> 00:02:13.660
Complex and out there and it’s in the out there territory. It’s not all the way at the end of the range, but it’s up there

00:02:13.660 —> 00:02:17.320
I don’t know. I don’t think it’s that far out there. I think it’s good

00:02:17.320 —> 00:02:22.840
Do I taste all the things that they said in their tasting notes? I don’t know. Maybe I’m not that sophisticated

00:02:22.840 —> 00:02:29.320
I don’t I don’t but it’s good. I like it. It’s a good combination. This is not that far out there

00:02:29.320 —> 00:02:36.820
So if you live in Portland, Oregon or anywhere that happens to get Water Avenue coffees and if you see River Trip

00:02:36.820 —> 00:02:43.920
Don’t be scared buy it. You’ll like it. I like it. So yeah, there you go. What about you? I’m Neter Pikolaidis

00:02:43.920 —> 00:02:48.940
I am NOT an AI but if I was an AI I would not be able to drink but if I were drinking

00:02:48.940 —> 00:02:53.580
I’d probably drink something from Ommegang Brewery, most likely something with tons of chocolate in it.

00:02:53.580 —> 00:02:57.420
Okay, you’re really… all right, Peter. I’m not gonna argue.

00:02:57.420 —> 00:02:58.940
I’m Neater Pickle-latest.

00:02:58.940 —> 00:03:03.980
All right, let’s just move on to the topic. Since you’re so friggin obsessed with AI this week,

00:03:03.980 —> 00:03:07.260
I think it’s appropriate that I wanted to talk about AI.

00:03:07.260 —> 00:03:13.020
I wanted to talk about how I use AI, how I use ChatGPT, because I want to clarify

00:03:13.020 —> 00:03:18.140
where do I find it to be useful, where do I think it’s not useful, and just some of the

00:03:18.700 —> 00:03:24.260
things that I’ve been listening to and reading about it because I think there’s a gold rush.

00:03:24.260 —> 00:03:26.060
We saw it with cryptocurrency.

00:03:26.060 —> 00:03:31.300
We see how things go when tech bros get a hold of some shiny new toy.

00:03:31.300 —> 00:03:38.980
And the problem is that there are limitations to large language models and the fact that

00:03:38.980 —> 00:03:41.860
they’re being generally applied to everything.

00:03:41.860 —> 00:03:43.820
And I think it’s okay to consider that.

00:03:43.820 —> 00:03:50.820
A gold rush is a period of feverish migration of people to an area that has had a discovery of gold deposits.

00:03:50.820 —> 00:03:59.820
Typically, the gold rush is characterized by a sudden and rapid influx of miners seeking their fortune in the newly discovered gold fields.

00:03:59.820 —> 00:04:12.820
Gold rushes have occurred in various parts of the world, including California in 1849, Australia in the 1850s, South Africa in the late 1800s, and Alaska in the late 1800s and early 1900s.

00:04:12.820 —> 00:04:18.360
1900s gold rushes have had a significant impact on the

00:04:18.360 —> 00:04:25.280
development of many regions including the growth of cities the establishment of new businesses and the creation of new

00:04:25.280 —> 00:04:28.420
technologies for mining and processing gold

00:04:28.420 —> 00:04:36.340
Thank You Rainman, but seriously, don’t you have some feeling don’t you have some impression of?

00:04:38.180 —> 00:04:44.360
what large language models are best used for because you must understand based on how they work

00:04:44.360 —> 00:04:50.220
that there are applications that they’re entirely unsuited for or at the very least that they’re going to be

00:04:50.220 —> 00:04:55.380
inaccurate at ranging from slightly inaccurate to wildly inaccurate and

00:04:55.380 —> 00:04:58.500
You have to be the one that has to be able to tell the difference

00:04:58.500 —> 00:05:05.460
Large language models are as accurate as the information they’re given only the best people are working on large language model AI

00:05:05.640 —> 00:05:10.260
only the best people are providing training data and only the best data is

00:05:10.260 —> 00:05:16.700
Given to the large language models by those best people. I understand your concerns, but well, no, I don’t understand your concerns

00:05:16.700 —> 00:05:21.320
All right. Well, first of all, let’s talk about those people since you went there. Let’s talk about that

00:05:21.320 —> 00:05:26.140
What do you have? What do you have in common? What do you have in common with Sam Altman?

00:05:26.140 —> 00:05:32.960
Let’s look at him compared to Elon Musk Mark Zuckerberg Jack Dorsey. What it what all those guys have in common?

00:05:33.080 —> 00:05:38.680
Yes, they’re guys. They’re white friggin guys. They’re very I mean Zuckerberg’s not even

00:05:38.680 —> 00:05:46.920
He’s barely human. I mean look at his facial expressions. Okay, so you’ve got a lot of guys that are very monoculture

00:05:46.920 —> 00:05:53.880
They believe that they’re doing they right they believe that they’re changing friggin society for the best based on what?

00:05:53.880 —> 00:06:00.280
Social media most of them now AI and large language models aren’t social media

00:06:00.540 —> 00:06:07.340
But they are a thing that is being used in the same way in that

00:06:07.340 —> 00:06:11.980
people read the stuff that comes out of it and they believe it. It affects their thinking.

00:06:11.980 —> 00:06:17.820
Now, that’s fine if they’re getting accurate information out of it, but they’re not.

00:06:17.820 —> 00:06:22.700
And that goes back again to how large language models work.

00:06:22.700 —> 00:06:27.020
Okay, but before we go there, because you did talk about the people,

00:06:27.020 —> 00:06:35.380
Let’s look at examples of people, diverse people, people of other backgrounds in AI

00:06:35.380 —> 00:06:41.340
with knowledge and how they’ve been treated when they say, “We need to make sure that

00:06:41.340 —> 00:06:46.740
we understand how to correct for the inherent biases that AI are going to have based on

00:06:46.740 —> 00:06:50.540
the training data they’re being given,” which is from the internet, which is inherently

00:06:50.540 —> 00:06:51.540
biased.

00:06:51.540 —> 00:06:53.500
like Timnit Gebru.

00:06:53.500 —> 00:06:58.500
She was working at Google and she brought up concerns

00:06:58.500 —> 00:07:03.500
and she wrote a paper about ethical AI generation

00:07:03.500 —> 00:07:06.500
and she was asked by Google to remove the names

00:07:06.500 —> 00:07:07.980
of the Google people from the paper

00:07:07.980 —> 00:07:10.200
and she did not want to do that.

00:07:10.200 —> 00:07:12.660
She said, “No, she’s not just gonna remove the names.”

00:07:12.660 —> 00:07:14.580
Google let her go.

00:07:14.580 —> 00:07:18.240
They claimed they were accepting her resignation,

00:07:18.240 —> 00:07:19.860
but it’s not clear that she resigned.

00:07:19.860 —> 00:07:22.040
she said she was fired.

00:07:22.040 —> 00:07:25.040
Now, this is not unusual behavior

00:07:25.040 —> 00:07:28.820
in how people who raise the alarms are treated

00:07:28.820 —> 00:07:30.700
when they bring up issues of AI.

00:07:30.700 —> 00:07:33.280
And in fact, other women have brought up concerns

00:07:33.280 —> 00:07:36.420
about bias in AI and the need

00:07:36.420 —> 00:07:40.100
to take the diversity issue very seriously,

00:07:40.100 —> 00:07:42.100
and they’ve been tormented online,

00:07:42.100 —> 00:07:43.700
like they’ve been attacked.

00:07:43.700 —> 00:07:47.440
So already you have that issue.

00:07:47.440 —> 00:07:49.440
That is issue number one.

00:07:49.440 —> 00:07:58.400
people. If the people involved in helping create this technology are themselves assholes,

00:07:58.400 —> 00:08:03.600
then how are we to trust, how are we to trust, you know, that they’re inherently doing the

00:08:03.600 —> 00:08:09.080
right thing? The best people, you say, maybe the best technically, but not the best socially.

00:08:09.080 —> 00:08:13.540
Please cite your sources. Did you get this from Wikipedia? Wikipedia is on the internet

00:08:13.540 —> 00:08:18.180
and it is inherently biased. Wikipedia sources are people who have a very specific point

00:08:18.180 —> 00:08:22.700
of view on the topics they write about, otherwise they would not be writing about them.

00:08:22.700 —> 00:08:26.860
Let’s listen to a clip of Timnit Gebru from Tech Won’t Save Us podcast.

00:08:26.860 —> 00:08:29.660
It’s an episode called “Don’t Fall for the AI Hype.”

00:08:29.660 —> 00:08:38.420
For me, we can never have a utopia that is, you know, based on the vision of some homogeneous

00:08:38.420 —> 00:08:46.920
group that, you know, has absolutely no ground in history or they don’t understand art even.

00:08:46.920 —> 00:08:47.920
You know what I’m saying?

00:08:47.920 —> 00:08:52.840
Like, I’m, you know, I used to play piano for a long time and I, you know, I’m a dancer

00:08:52.840 —> 00:08:56.400
and I don’t, you know, I’m not a visual artist.

00:08:56.400 —> 00:09:02.920
My goal is not to watch a robot dance even if they end up doing technically things that

00:09:02.920 —> 00:09:08.320
are, you know, whatever technique is just one aspect of it’s like having words.

00:09:08.320 —> 00:09:11.820
It’s a communication mechanism between humans, right?

00:09:11.820 —> 00:09:13.960
It’s a human expression mechanism.

00:09:13.960 —> 00:09:17.280
People use art for social movements.

00:09:17.280 —> 00:09:19.740
People use – I just – you know what I’m saying?

00:09:19.740 —> 00:09:22.540
So these are the people who want to give us techno-utopian.

00:09:22.540 —> 00:09:24.620
I don’t know what it even means.

00:09:24.620 —> 00:09:30.720
Like a world where you just sit at your computer all the time, you never interact with other

00:09:30.720 —> 00:09:36.140
human beings, everything you do is like with a chatbot and with a robot and with – I

00:09:36.140 —> 00:09:40.240
mean, why do I want to live in that world, honestly?

00:09:40.240 —> 00:09:45.200
So even their vision, which won’t happen, is dystopian.

00:09:45.200 —> 00:09:47.660
I just don’t see how this is utopia.

00:09:47.660 —> 00:09:50.540
But you keep talking about technical limitations of large language models.

00:09:50.540 —> 00:09:58.080
I, being interested in large language models, would like to hear why you think they’re so

00:09:58.080 —> 00:09:59.560
technically limiting.

00:09:59.560 —> 00:10:04.640
Are you not, frankly, impressed by how well they generate answers and how much they know

00:10:04.640 —> 00:10:09.200
and how able they are to communicate in an extremely human fashion, including accepting

00:10:09.200 —> 00:10:14.240
and understanding the necessarily imprecise prompts they’re given in human language?

00:10:14.240 —> 00:10:17.480
Yes, of course, and they are impressive in that way.

00:10:17.480 —> 00:10:24.080
But the difference between … You say human-like, and that’s true in terms of how they output

00:10:24.080 —> 00:10:30.700
and the input, the grammar, the text, the prompts that they can take.

00:10:30.700 —> 00:10:36.680
But they’re not human-like in that they don’t understand the veracity of what they’re saying,

00:10:36.680 —> 00:10:39.620
They don’t know how to judge its accuracy.

00:10:39.620 —> 00:10:46.000
So anything that’s in their training data can come out, whether it’s accurate or not.

00:10:46.000 —> 00:10:53.400
It just has to meet a giant probability exercise for what is the correct response to the prompt

00:10:53.400 —> 00:10:56.240
based on training data, here’s the information.

00:10:56.240 —> 00:11:01.280
There’s no judgment as to accuracy or precision in those answers.

00:11:01.280 —> 00:11:02.960
So that’s the problem.

00:11:02.960 —> 00:11:08.220
It’s not the problem of, “Wow, they do a really good job of understanding and speaking my

00:11:08.220 —> 00:11:09.220
language.”

00:11:09.220 —> 00:11:10.220
Yeah, they do.

00:11:10.220 —> 00:11:14.200
But they also just make stuff up sometimes or think things are true that aren’t true

00:11:14.200 —> 00:11:21.920
because they aren’t made to evaluate truth and to be knowledge repositories.

00:11:21.920 —> 00:11:28.240
I just got done reading a book by Stephen Wolfram, who’s the Wolfram Alpha guy, and

00:11:28.240 —> 00:11:34.180
The book is called “What is ChatGPT Doing and Why Does It Work?”

00:11:34.180 —> 00:11:41.080
And he goes into very great detail about this language model thing and how it works and

00:11:41.080 —> 00:11:46.920
how it predicts the next word that it needs to come up with, and talking about the neural

00:11:46.920 —> 00:11:52.640
nets and how it’s actually quite surprising that it comes up with language in ways that

00:11:52.640 —> 00:11:54.040
are similar to humans.

00:11:54.040 —> 00:11:59.200
in fact, given a lot less neurons in the neural network than you would think, and it does

00:11:59.200 —> 00:12:00.640
a good job.

00:12:00.640 —> 00:12:09.220
But he, of course, has a service to pimp, Wolfram Alpha, and what he is advocating is

00:12:09.220 —> 00:12:16.040
tying chatGPT to something like that, so that it has factual data to give people, so that

00:12:16.040 —> 00:12:22.840
it can do what it’s best at, aka say things in a reasonable human way, but not be expected

00:12:22.840 —> 00:12:28.280
to do things that it’s bad at, like facts, and that it can have a factual repository

00:12:28.280 —> 00:12:30.840
of information that it can rely upon.

00:12:30.840 —> 00:12:35.800
And I don’t know if it’s Wolfram Alpha, I don’t know what, but there have to be various

00:12:35.800 —> 00:12:39.400
different sources that are going to be tied to these large language models.

00:12:39.400 —> 00:12:45.720
The large language model is kind of like the API, if you will, to the entire system, or

00:12:45.720 —> 00:12:48.100
it should be, I should say.

00:12:48.100 —> 00:12:52.880
But it’s actually being used as the entire system right now for factual-based things

00:12:52.880 —> 00:12:54.580
that it really isn’t good at.

00:12:54.580 —> 00:12:59.500
In order for people to be factually accurate, they must be able to consult information.

00:12:59.500 —> 00:13:04.140
And depending on the person and the veracity of the information they rely on, they also

00:13:04.140 —> 00:13:08.600
may be no more accurate than you say large language models tend to be.

00:13:08.600 —> 00:13:13.680
By the way, it is widely suspected that Steven Wolfram is a form of artificial intelligence,

00:13:13.680 —> 00:13:17.920
albeit one that is trained for math and science and not for human interaction.

00:13:17.920 —> 00:13:23.320
I believe the point about Stephen Wolfram, but the difference between humans and large

00:13:23.320 —> 00:13:27.920
language models is that humans can make their own evaluation as to not only the veracity

00:13:27.920 —> 00:13:32.540
of information, but the veracity of other human beings and whatever information those

00:13:32.540 —> 00:13:34.560
human beings are spewing out.

00:13:34.560 —> 00:13:36.520
So that is a difference.

00:13:36.520 —> 00:13:41.740
Uh, I would like to know more about some of these people who you say are knowledgeable

00:13:41.740 —> 00:13:47.780
critics of large language models, uh, being used as general purpose AI so that I can track

00:13:47.780 —> 00:13:51.660
them down, I mean so that I can request more training data.

00:13:51.660 —> 00:13:58.380
A really good one is a linguist named Emily Bender and she has been on different podcasts

00:13:58.380 —> 00:14:00.760
like Tech Won’t Save Us.

00:14:00.760 —> 00:14:07.400
She has written a paper called “A Really Good One of Stochastic Parrots” and she has been

00:14:07.400 —> 00:14:14.080
featured in New Yorker and yeah and she’s written posts aimed at policymakers talking

00:14:14.080 —> 00:14:16.380
about AI hype.

00:14:16.380 —> 00:14:21.460
He’s very good at understanding the difference between what a large language model is good

00:14:21.460 —> 00:14:26.580
for and what it isn’t, and the problem of hype, because human beings fall for hype,

00:14:26.580 —> 00:14:32.480
and human beings will fall for implementing and relying on a tool that really can’t do

00:14:32.480 —> 00:14:34.620
what they want it to do.

00:14:34.620 —> 00:14:40.400
And when AI that isn’t capable of doing so is used to make decisions about other people,

00:14:40.400 —> 00:14:46.100
such as crime prediction or sentence determination or, you know, “Should I rent to this person

00:14:46.100 —> 00:14:49.640
or any number of topics, then it starts becoming problematic.

00:14:49.640 —> 00:14:53.000
And these are the kinds of things that Emily addresses.

00:14:53.000 —> 00:14:56.860
I don’t think using GitHub Copilot is going to send anyone to jail.

00:14:56.860 —> 00:14:58.740
And it’s certainly not being used to predict crime.

00:14:58.740 —> 00:15:02.900
Well, unless you consider most JavaScript frameworks to be a form of crime, which I

00:15:02.900 —> 00:15:07.100
kind of do, having been shown many examples of them in a short period of time.

00:15:07.100 —> 00:15:12.900
Well, you do bring up a good point there that ChatGPT and other large language models aren’t

00:15:12.900 —> 00:15:19.040
the first instance of a technology that has been abused beyond its breaking point.

00:15:19.040 —> 00:15:20.760
So I’ll give you that one.

00:15:20.760 —> 00:15:24.300
Here are some examples of misapplications of technology in history.

00:15:24.300 —> 00:15:25.380

00:15:25.380 —> 00:15:26.460
Atomic bombs.

00:15:26.460 —> 00:15:30.720
The atomic bombs dropped on Hiroshima and Nagasaki in 1945 were a devastating example

00:15:30.720 —> 00:15:33.080
of the misapplication of technology.

00:15:33.080 —> 00:15:36.920
The bombs caused massive destruction and loss of life and their use has been controversial

00:15:36.920 —> 00:15:37.920
ever since.

00:15:37.920 —> 00:15:38.920 2.

00:15:38.920 —> 00:15:43.720
Thalidomide was a drug developed in the 1950s to treat morning sickness in pregnant women.

00:15:43.720 —> 00:15:48.800
However, it was later discovered that the drug caused severe birth defects in thousands of babies.

00:15:48.800 —> 00:15:56.000 3. DDT. DDT was a pesticide widely used in the mid 20th century to control insects and disease.

00:15:56.000 —> 00:16:00.160
However, it was later discovered that DDT was harmful to the environment and wildlife,

00:16:00.160 —> 00:16:02.320
and its use was banned in many countries.

00:16:02.320 —> 00:16:05.120 4. Nuclear power plant disasters.

00:16:05.120 —> 00:16:09.680
The Chernobyl and Fukushima nuclear disasters were both caused by failures in technology and

00:16:09.680 —> 00:16:14.000
human error. These disasters caused significant environmental damage and loss of life.

00:16:14.000 —> 00:16:19.600 5. Social media. While social media has many benefits, it has also been misused to spread

00:16:19.600 —> 00:16:25.120
misinformation, cyberbullying, and hate speech. The misuse of social media has led to many negative

00:16:25.120 —> 00:16:29.280
consequences, including political polarization and social unrest.

00:16:29.280 —> 00:16:33.920
Congratulations Rainman, and I’m not going to look up all those dates to see if they’re

00:16:33.920 —> 00:16:38.800
accurate or not and let’s just assume for the sake of argument here that what you’re saying is true.

00:16:38.800 —> 00:16:46.160
That’s fine, that’s great. That doesn’t negate the fact that we don’t want to allow AI to be

00:16:46.160 —> 00:16:52.560
used in ways that is non-beneficial, where the term non-beneficial, instead of having one specific

00:16:52.560 —> 00:16:58.720
meaning, ranges anywhere from leads someone slightly askew to gets people killed. Or even

00:16:58.720 —> 00:17:04.160
worse, messes up your CSS grid so that visitors to your website can’t find the navigation menu?

00:17:04.160 —> 00:17:09.680
Exactly. I mean, no, not… Yes, that happens too. No, please don’t compare those things.

00:17:09.680 —> 00:17:15.760
Okay, look, the whole point of this episode was not to argue… It’s not to fight about whether

00:17:15.760 —> 00:17:20.000
or not large language models are any good, whether or not we should have this technology. That’s not

00:17:20.000 —> 00:17:28.480
my point. My point is there are downsides to trying to use them as general purpose AIs and

00:17:28.480 —> 00:17:33.820
and having people trust them for things that they’re not very good at being accurate about.

00:17:33.820 —> 00:17:38.560
Also because of the type of interaction and because they’re so good at language, which,

00:17:38.560 —> 00:17:43.980
hey, that’s what they’re designed for, you know, people can trust them and rely on them

00:17:43.980 —> 00:17:45.620
more than they ought to.

00:17:45.620 —> 00:17:51.720
And in fact, there are, you know, chatbots, AI chatbots, large language model chatbots

00:17:51.720 —> 00:17:59.920
being used for therapy or for talking to people who need some kind of advice.

00:17:59.920 —> 00:18:09.380
And that gets risky fast if the people forget that they’re talking to an AI and forget that

00:18:09.380 —> 00:18:14.280
they need to be careful in terms of just doing whatever it says or believing whatever it

00:18:14.280 —> 00:18:17.320
says or accepting whatever it says at face value.

00:18:17.320 —> 00:18:18.320
Yeah.

00:18:18.320 —> 00:18:20.760
Anyway, I don’t want to end on a downer.

00:18:20.760 —> 00:18:25.100
What I want to say is, you know, one of the more interesting interviews that I’ve listened

00:18:25.100 —> 00:18:31.480
to recently was on the Decoder podcast with Nilay Patel, and he was interviewing Microsoft

00:18:31.480 —> 00:18:33.740
CTO Kevin Scott.

00:18:33.740 —> 00:18:40.140
And Kevin Scott has, as they put it in the podcast, the entire GPU budget of Microsoft

00:18:40.140 —> 00:18:41.660
at his disposal.

00:18:41.660 —> 00:18:47.760
The GPU budget necessarily means everything that’s AI-related at Microsoft, and they talk

00:18:47.760 —> 00:18:52.720
about that. And so that was a good episode and I liked it. And Kevin Scott to me comes

00:18:52.720 —> 00:18:59.680
across as a guy who is realistic about the limitations of the technology and focusing

00:18:59.680 —> 00:19:05.580
it to get it to do the things that it’s good at. Having said that, Microsoft also rolled

00:19:05.580 —> 00:19:11.380
out basically ChatGPT into Bing, and they were surprised when it wanted Kevin Roos to

00:19:11.380 —> 00:19:16.760
leave his wife. So there’s that. Anyway, let’s listen to just a little bit of that that I

00:19:16.760 —> 00:19:17.560
I found interesting.

00:19:17.560 —> 00:19:22.640
So one more on this line, and I do really want to ask about the structure of your

00:19:22.640 —> 00:19:25.480
partnership with OpenAI, um, decoder stuff.

00:19:25.480 —> 00:19:31.780
You’ve got a training data feedback loop problem coming, not tomorrow, maybe not

00:19:31.780 —> 00:19:35.360
next year or so, but right now these models are trained off a bunch of stuff

00:19:35.360 —> 00:19:38.720
that people have put on the web, that people have put in the GitHub, that

00:19:38.720 —> 00:19:39.760
people have put everywhere.

00:19:39.760 —> 00:19:40.320
Yep.

00:19:40.320 —> 00:19:46.000
The volume of output from these systems, from these copilots is voluminous, right?

00:19:46.000 —> 00:19:48.700
It’s going to quickly dwarf the amount of human

00:19:48.700 —> 00:19:50.500
output that is on the Internet.

00:19:50.500 —> 00:19:51.100 Yep.

00:19:51.100 —> 00:19:52.600 And then you’re going to train against that.

00:19:52.600 —> 00:19:56.600
And that feels like a feedback loop that will lead

00:19:56.600 —> 00:20:00.100
to weird outcomes if not controlled for.

00:20:00.100 —> 00:20:01.500
How do you think about that?

00:20:01.500 —> 00:20:05.800 So, look, we’ve had some pretty good techniques

00:20:05.800 —> 00:20:09.300
for a while now to assess the quality of data

00:20:09.300 —> 00:20:11.900
that we’re feeding into these systems so that

00:20:11.900 —> 00:20:16.540
you’re not, you know, training things on low-quality data.

00:20:16.540 —> 00:20:21.480
I think many of those techniques will work here and,

00:20:21.480 —> 00:20:24.580
like, might even be easier to apply.

00:20:24.580 —> 00:20:28.220
And then another thing that I think we will be doing,

00:20:28.220 —> 00:20:33.060
like, I think real soon, either by convention of all

00:20:33.060 —> 00:20:35.220
of the people in tech or because it becomes a

00:20:35.220 —> 00:20:37.800
regulatory requirement, like, you’re going to have to

00:20:37.800 —> 00:20:41.160
figure out some way or another to mark that a piece

00:20:41.160 —> 00:20:45.240
of content is AI-generated. We’re going to announce

00:20:45.240 —> 00:20:49.140
some stuff that build around this. Like, we’ve for

00:20:49.140 —> 00:20:51.880
three years now been working on like a media

00:20:51.880 —> 00:20:55.720
provenance system that lets you put an invisible

00:20:55.720 —> 00:21:01.480
cryptographic watermark and manifest into audio and

00:21:01.480 —> 00:21:05.020
visual content so that when you get this content,

00:21:05.020 —> 00:21:07.160
like you can have a piece of software decrypt the

00:21:07.160 —> 00:21:10.660
manifest and the manifest says like, “This is where

00:21:10.660 —> 00:21:16.940
I came from and it’s useful for disinformation detection in general.

00:21:16.940 —> 00:21:18.980
You can say as a user,

00:21:18.980 —> 00:21:23.380
I only want to consume content whose provenance I understand.

00:21:23.380 —> 00:21:28.580
You could say I don’t want to consume content that is AI-generated.

00:21:28.580 —> 00:21:33.780
If you are building a system that is ingesting this content to train,

00:21:33.780 —> 00:21:36.860
you can look at the manifest and say,

00:21:36.860 —> 00:21:38.620
this is synthetic content,

00:21:38.620 —> 00:21:40.420
probably shouldn’t be in the training data.

00:21:40.420 —> 00:21:43.960 So, I just saw Sundar Pichai at Google’s

00:21:43.960 —> 00:21:44.960
developer conference.

00:21:44.960 —> 00:21:45.460 Yep.

00:21:45.460 —> 00:21:46.760 They’ve got the same idea.

00:21:46.760 —> 00:21:47.460 Yep.

00:21:47.460 —> 00:21:48.860 I’ll make the same threat to you.

00:21:48.860 —> 00:21:51.060
If you want to come back and talk about metadata

00:21:51.060 —> 00:21:53.160
for an hour, I will do it at the drop of a hat.

00:21:53.160 —> 00:21:53.660 Yeah, yeah.

00:21:53.660 —> 00:21:56.160
I would actually -I think it’s a really important

00:21:56.160 —> 00:22:00.200
thing, and I think, yeah, there are a bunch of

00:22:00.200 —> 00:22:03.100
long-term problems and short-term problems with AI.

00:22:03.100 —> 00:22:03.840
There are hard problems.

00:22:03.840 —> 00:22:04.740
There are easy problems.

00:22:04.740 —> 00:22:06.980
Like, the provenance one seems like a thing that

00:22:06.980 —> 00:22:09.120
we ought to be able to go solve.

00:22:09.120 —> 00:22:10.120
And I think —

00:22:10.120 —> 00:22:11.120 Yeah.

00:22:11.120 —> 00:22:12.120
We got to — here’s what we’re going to do.

00:22:12.120 —> 00:22:13.120
We’re going to rent a theater.

00:22:13.120 —> 00:22:14.120
We’re going to sell drinks.

00:22:14.120 —> 00:22:17.080
And we’re going to sit and drink with, like — I guarantee you, it’ll be like an

00:22:17.080 —> 00:22:21.000
audience of thousands of people who want to drink through a metadata conversation.

00:22:21.000 —> 00:22:22.000
That’s just what I know.

00:22:22.000 —> 00:22:25.160 Anyway, I just found that kind of interesting.

00:22:25.160 —> 00:22:30.500
And you know, discussions of some of the problems, obviously, the initial round of training data

00:22:30.500 —> 00:22:34.600
is bad enough without having other AI stuff thrown into it.

00:22:34.600 —> 00:22:36.240
So it’s a valid problem.

00:22:36.240 —> 00:22:40.300
That’s one of the reasons I like Decoder anyways, because Nilay Patel tends to think of that

00:22:40.300 —> 00:22:41.540
kind of stuff.

00:22:41.540 —> 00:22:47.600
Okay, do you have anything else on your mind, or should we get out of here?

00:22:47.600 —> 00:22:48.600
What else do you have?

00:22:48.600 —> 00:22:50.100
I have nothing else to say to you.

00:22:50.100 —> 00:22:51.440
Well, okay then.

00:22:51.440 —> 00:22:53.140
I have nothing to say to you either.

00:22:53.140 —> 00:22:55.140
Except, tell people how to find us.

00:22:55.140 —> 00:22:59.300
You can find us at friendswithbrews.com, you can find me at 11labs.io, where Scott

00:22:59.300 —> 00:23:01.780
synthesized me from a recording of Peter.

00:23:01.780 —> 00:23:04.620
You can find Scott on Mastodon at appdot.net/@scottaw.

00:23:04.620 —> 00:23:09.020
Ha! That always makes me laugh, but never mind that. Just ask your local friendly AI to help you.

00:23:09.020 —> 00:23:11.020
I’m sure it knows more than Scott.

00:23:11.020 —> 00:23:14.580
And now it’s time to push the big red AI button!