September 4, 2026 • 48 mins
Are kids going to be any safer online after the Meta settlement? Our trio of contributors—Taylor Lorenz (User Mag), Reed Albergotti (Semafor), and Nitasha Tiku (The Washington Post)—are skeptical. Then, are AI “civilizations” real, or is it all just tabloid journalism? Finally, school’s back in session, and teachers are rethinking the classroom for an age when anyone can generate the perfect answer.
See omnystudio.com/listener for privacy information.
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:12):
Natasha Taylor Reid, I thought until this week that my
stepmother was the last known survivor on MapQuest, but no longer.
Speaker 2 (00:20):
Yes, MapQuest, the most popular download.
Speaker 3 (00:23):
Yeah.
Speaker 2 (00:24):
This is because of Lake Ontario, right? Or Lake America?
Speaker 1 (00:27):
Yeah, MapQuest, unlike Google Maps and Apple, who both changed
Lake Ontario to Lake America, MapQuest stood loud and proud with.
Lake Ontario and got a big boost in the app
charts as a result.
Speaker 2 (00:39):
Do all these lakes and bodies of water and whatever else,
do they just go back as soon as Trump leaves office?
Is that what's going to happen?
Speaker 3 (00:47):
That's a good question.
Speaker 1 (00:48):
I don't know how MapQuest has decided to call the
Strait of Hormuz the Strait of Trump, or maybe it's
also holding firm there.
Speaker 2 (00:56):
Just everything is Trump.
Speaker 1 (01:03):
Welcome to Tech Stuff. I'm Oz Veloshian, and this is
The Week in Tech, where I'm joined by the world's
most plugged-in reporters to break down what's really happening in
tech right now. Today, we're joined by Reid Albregotti, tech
editor of Semaphore, Taylor Lorenz of UserMag, and Natasha Tiku,
tech reporter at The Washington Post. Welcome all.
Speaker 2 (01:21):
Hey, everyone. Good to be here.
Speaker 3 (01:23):
Good to be back.
Speaker 2 (01:24):
Yeah, it's been a little break.
Speaker 1 (01:25):
I missed you all on the, on the beach in Greece.
We haven't had a round table for a couple of weeks. Uh,
it's been an eventful couple of weeks and in general,
we spend quite a lot of time on the show
talking about kind of the, the new guard in tech. Um,
but the old guard have recently been dominating the headlines
for better or for worse. Reed, why don't you kick
us off with the meta settlement?
Speaker 2 (01:47):
Yeah. So meta settled its big tobacco, uh, attorney general lawsuit. Um,
$ 18 billion. But I think the bigger story is all
these new restrictions on how teens can use the service.
So they're limited to two hours a day.
Speaker 3 (02:02):
Not just teens.
Speaker 2 (02:03):
Taylor's, you know, I saw you tweeting about this, Taylor.
So I know you have like major thoughts.
Speaker 3 (02:08):
Yeah, I've been covering this very closely. What I would
say is that so Meta, this is nothing. Meta does
not have to pay that full $ 18 million at all,
which is a drop in the bucket for Meta. I'm sorry, billion.
Speaker 2 (02:20):
18 million would be like... We're truly nothing.
Speaker 3 (02:23):
Truly nothing. Yeah, Mark Zuckerberg's lunch. Basically, what the result is,
as Reid mentioned, is there's a bunch of new restrictions
on anyone's account who Meta can't verify is an adult
through identity verification. So Meta is going to start harvesting
significantly more data on users, all users, in order to
(02:43):
ID verify them.
Speaker 1 (02:45):
And how will they do that? Will it be like...
driving license? So what are the ways they're going to
actually be doing it?
Speaker 3 (02:50):
Yeah, absolutely. So if they determine that you might be
underage or that you're interacting with an underage person, you
will be uploading your government identification. A lot of people
have their information stored already. So their payment information is
another way to do it, harvesting biometric data. But they
are going to be doing ID verification for the internet.
It is a huge rollback. And Anybody that cannot verify
(03:12):
their identity within 14 days basically will be put under
all of these heavy restrictions. So adult accounts... anybody that
does not meet these qualifications. It's very bad because these
restrictions are, you know, unable to view lots of significant
amounts of content that is deemed harmful for children. We
know that the government has determined in other states that
(03:32):
content that is harmful for children includes anti-ICE content. It
includes LGBTQ content. It includes reproductive justice, healthcare content.
Speaker 2 (03:40):
So.
Speaker 3 (03:41):
This is a horrible loss for free speech, and it
doesn't do anything to change the business model of meta. And,
you know, it's basically just a way for these attorney
generals to get a headline.
Speaker 1 (03:54):
Reid and Tasha, do you agree with Taylor's take?
Speaker 2 (03:56):
I do agree. I think it is kind of a rollback.
I mean, look, my kids are eight and ten. Like,
they're too young to use this stuff right now. So
I sort of had this fantasy in my mind since
they were born that, like, by the time they are—
of social media age will have figured this all out. Uh,
clearly that fantasy is not going to become reality. Like
(04:16):
this is still a problem. And like, you know, we
go to Australia every year by, you know, my, my
in-laws are in Sydney and they passed this ban. Um,
everybody was so excited about it because it was like,
you know, I think parents just, it's just so hard
for them to, to deal with this, to grapple with
what's happening on social media. Um, I was very skeptical
as Taylor is like, any of these regulations are actually
(04:39):
going to solve this problem. And they and they may
create new problems. I think one point I would make, though,
is that on the ID verification point, like somehow, because
of AI, because of all this, you know, just inauthenticity online,
I think we're going to have to figure out some
way to do privacy preserving some sort of verification of
(05:00):
not just your age, but that you're an actual human being.
And I think this is like, an opportunity, right? Like
there should be a technical solution. I think there are
lots of really, there are lots of really interesting technical
ideas out there. I'm not, I'm not going to endorse
any one of them right now. But like, there's, I
think this is a great opportunity for innovation.
Speaker 3 (05:21):
But I feel like I have whiplash as a tech reporter,
because just less than 10 years ago, they were hauling
Mark Zuckerberg in front of Congress. Remember when he, you know,
Zuckerberg says in the Cambridge Analytica sort of, scandal, like Senator,
we sell ads, you know, and they were like, why,
you know, you harvested all this data on users. And
now we are mandating that they do significantly more than that.
(05:43):
I think the issue of sort of human stuff online
at like sort of determining humans versus bots, like I
agree with Reid. I think that is an increasing problem
and we are going to have to find some solution
to that. Meta does not seem interested in that solution.
In fact, we had Adam Masseri just recently talking on
a podcast about how they want more AI creators and
(06:03):
they're leaning further into AI.
Speaker 2 (06:05):
Taylor, there's another overarching theme. I bet you'll agree with
me on this, but this applies to this FTC settlement
with Amazon. It applies to the AI regulation that's being
passed today. I think tech regulation needs to happen much faster.
And it needs to be able to change much faster,
because none of this stuff like the AI regulations that
(06:26):
have been passed are already, like they seem like years
out of date. And I think this, this one will
be another one where, like, yeah, okay, let's try to,
let's try to work this out. But let's, let's be nimble.
Let's be able to pass laws right now. We can't
do that. I mean, we can't, especially on a federal level.
Speaker 3 (06:42):
I think another thing is like, what are you passing
laws based on? And this whole sort of child safety
movement lately is passed on nothing. It's nothing but vibes.
We have every single top researcher who studies the use
of social media technology saying, hey, there's not a causal relationship.
In fact, depressed kids use Instagram more. It's a symptom
(07:03):
of the broader problem. We know what's causing this use
mental health crisis. Let's address that stuff. And it's sort
of this like after the fact, again, just like people
want to put a bandaid. They want a headline. They
want to get reelection. Like, look at I'm tough on
big tech while you've done nothing to meaningfully change the
tech landscape, but make it worse for all of us.
And I agree. I mean, at this point, it's so
late anyway, because these kids are busy talking to chat
(07:25):
bots like Instagram and Snapchat is the least of of
what parents should be sort of concerned about.
Speaker 2 (07:30):
Yeah.
Speaker 1 (07:31):
Natasha, I'm curious, what's what's your take on this on
this meta settlement?
Speaker 4 (07:34):
I don't know. I remember getting excited about GDPR and
feeling like, you know, maybe this will be like, this
will be it.
Speaker 1 (07:43):
Not a lot.
Speaker 2 (07:44):
GDPR is the bane of my career.
Speaker 3 (07:45):
It's terrible.
Speaker 4 (07:46):
And now I just like click, like reject, accept, whatever.
So it just feels not that everyone is fixated on
AI solely, but it does feel like while the ball
has been, you know, everyone's been looking at open AI
and anthropic and even like Sam Altman's you know, world
coin eye scanning identity thing here Meta slips in. And
(08:09):
just by default, because it has so many users, is
going to become like the default identity layer. And, you know,
we're going to have a flock-like scandal in five years
about the technology that Meta's using. And, yeah, it's like
whiplash plus deja vu.
Speaker 3 (08:27):
I really agree.
Speaker 4 (08:27):
With Reid about like being nimble in writing the laws.
Like there's nothing that stops them from saying we need
to revisit this. next year or i mean when it
comes to ai at least you know if they had
had transparency requirements say that wouldn't be outdated by now
so it just it's i remember being excited about a
(08:47):
decade ago that something might happen that would be like
a oversight or a check on their concentration of power
and i have yet to see them yeah i've yet
to see them pin down on.
Speaker 3 (09:00):
Anything Can I add just one thing that I think
is so important for people to realize is that this
is not a law. This is the settlement. And if
these changes, if any sort of legal system, state or federal,
tried to pass these changes to META, it would be
struck down immediately on First Amendment grounds because it has
such broad implications for privacy, for freedom of expression, etc.
(09:25):
So these are highly unconstitutional changes. changes that these state
attorney generals know they could never get through legislation.
Speaker 2 (09:34):
Well, they'll be challenged, right? I mean, won't these be challenged?
Speaker 3 (09:38):
No, it can't be challenged because it is not a law.
It is a settlement that Meta has proactively written and
agreed to. And so at least when I was talking
to one of the people yesterday, there's no challenge to it.
Meta's not challenging it. They actually waived the right in
this settlement They.
Speaker 2 (09:56):
Might not challenge it, but I mean, I not to
say they would win. but like someone could someone could
sue and say, you know, that I mean, because there
is this whole argument, right, that these that these companies
aren't just private companies that can do whatever they want,
that they are sort of the public town square. That's
been like the the free speech talking point.
Speaker 1 (10:16):
Part of the settlement is only payable if YouTube and
TikTok also agree to the terms of the settlement.
Speaker 2 (10:21):
Is that right?
Speaker 3 (10:22):
Well, the settlement is enforced no matter what, right? The
terms of the settlement. But yeah, they don't have to
pay the full $ 18 billion unless TikTok and YouTube enact
similar changes.
Speaker 1 (10:33):
Yeah, I mean, what's the vibe on, I mean, Patel,
you mentioned vibes, but like Natasha Reid, what's the back
to school vibe? Are the parents talking about this? And
is it like something that's coming up?
Speaker 2 (10:43):
My kids were talking about it. It was one of
the few current events that my kids actually brought up
to me.
Speaker 1 (10:48):
What did they say?
Speaker 2 (10:49):
I mean, I think they're sort of like, they don't
have like strong opinions. Like Taylor linked to someone, I
think it was on X, right? That you linked to someone,
some kid talking about how they, you know, he was
sort of upset about this.
Speaker 3 (11:02):
Student journalists have been really affected. So student journalists have
mobilized across the country to try to raise alarms about
this and say, by the way, like, We have picked
up the mantle on local news has been decimated. Student
journalists are the ones doing, you know, a lot of
this accountability reporting. And a lot of these young people
that do journalism and, you know, engage in this world,
including the one I posted about, are really upset.
Speaker 4 (11:24):
Since we know nothing's happening in the real world, I
also have a fantasy where the people who educate the
regulators on this technology are not solely bought and paid
for by the companies themselves or by people who have
you know, really vested interests that are not the public
public's interests.
Speaker 2 (11:43):
They're also like to Taylor's earlier point about this social
media stuff, like a lot of times these things that
get blown up in the press, you know, those those
media people, they're not actually like huge problems, right? Like,
it's like, you know, it's it's the thing that people
care about, because they're reading about it in the press.
But like, it's like the copyright thing with AI, right? Like,
(12:05):
oh my God, they used my book as one of
10 gazillion books to train this AI model. I want
some money for my book being used. And it's like,
no one cares that you can literally just go download
my book for free on a million pirated book websites, right?
It's like that to me, that's the problem.
Speaker 1 (12:22):
Not like- People like, I mean, people like these symbol,
like it's a symbol, like there's a symbolic thing, right?
Like there's been some some accountability, like make them meant
to have to pay some money.
Speaker 3 (12:33):
And like, that's, but it's not accountable. They want fake,
they want fake accountability. They want the headline. They don't
want any real change.
Speaker 2 (12:41):
Totally. People love the symbolism, but then like, That's not
the job of a real leader, right? A leader is
not to like give the people some red meat to
keep them satisfied. It's to like actually solve problems, like
real problems for people that affect people's lives. That's why
today I'm announcing I'm leaving journalism.
Speaker 1 (13:04):
When we come back, are we living in the age
of AI civilizations? And what should we call them?
Speaker 2 (13:10):
Stay with us. Welcome back.
Speaker 1 (13:30):
Once again, there's more information on the OpenAI Hugging Face incident.
This time, a report by two AI safety research organizations,
META and Redwood Research. Many people have tried to make
sense of what exactly is the takeaway here, but one
explanation seems to have broken into the mainstream. Taylor, you
wrote about a recent piece by Silicon Valley's favorite podcaster,
(13:52):
Dwarkesh Patel, which promised to explain the whole OpenAI hugging
face incident in plain English. First, I want to hear
a brief recap from you in plain English of what
actually happened, and then your account of Dwarkesh's spin and
its spinning out into the world.
Speaker 3 (14:10):
Well, I have to say, I was kind of channeling
Natasha because I feel like she is so plugged into
this world of like, I was trying to figure out like,
is Dwarkesh like EA or not? I was like, Natasha
would know like all the lore. Anyway, Dwarkesh Patel, as
you mentioned, is a big popular Silicon Valley podcaster. He
wrote this piece about the Hugging Face incident that was
(14:31):
so crazy. It did break containment. His piece broke containment
of Silicon Valley where like normie people were sharing this
all over.
Speaker 1 (14:38):
I know, obviously I'm aware of Dwarkesh's work and see
what he, you know, subscribe to his newsletter and sometimes
do his podcast. But I've never ever seen one of
his ideas percolate in like mainstream culture before. So it's
kind of interesting.
Speaker 3 (14:49):
He wrote this piece that was.
You know, it was like very flowery language, very, it
read like kind of like a fairy tale. Like he
was giving, you know, ascribing a lot of like sort
of human motivations to these AI bots and stuff. And
so what I wrote about is basically like, you know,
(15:09):
This is not new, the concept of sort of ascribing agency,
the concept of putting things in plain English for normal people.
But I would argue when you do that, it's usually
what tabloids have done. And it usually results in inflammatory
freakouts because people think that these agents are going to
kill everyone. So I'm like, if you're going to do that, basically,
(15:30):
you need to couch it in a lot of language saying, Basically,
that's not going to happen, and here's why I'm explaining
it in a certain way. Or maybe you do believe
AIs are going to kill everyone, and you are this
doomer-pilled EA person. At that point, I think you should
be explicit about it.
Speaker 1 (15:44):
Here's what Dwarkesh said. Over the course of three months
at OpenAI, three consecutive secret AI civilizations got started, then
got wiped out, only to reemerge from the predecessor's ashes.
Speaker 2 (15:56):
I read this as actually him, because he's talked about
this in the past that he sort of thinks these agents,
you know, chat GPT things have like rights, like they're
like beings that have rights. And I read that as like,
because he said this before that like on his podcast
that like, I'm worried about him first person. Like I'm
worried about, you know, basically we're going to create an
(16:18):
army of slaves that we're going to, you know, like
these chatbots are slaves. And so I read that as
like him saying, there's like two parts to that. One
is like the AI safety problem. Are they going to
rise up and take over the world? He sort of
nicknamed them after, you know, ancient civilization leaders, like, you know,
Alexander the Great or whatever. So they can take us over.
(16:39):
And the other one is like, do they have rights?
Like using the term wiped, like we wiped out a civilization,
I think is like him kind of hinting that maybe
that was morally wrong or something. That's how I read it. Natasha?
Speaker 4 (16:52):
I mean, I can't believe that That is considered like
plain language for people to understand. I mean, it's beyond flowery.
It's like it's like it's like hero making, like myth making.
Speaker 2 (17:07):
Right.
Speaker 4 (17:07):
I mean, to jump from and I don't want to
say that he's EA, but I will say, yeah, like
a lot of what Reid said is accurate.
Speaker 3 (17:14):
Right.
Speaker 4 (17:15):
He is I mean, he lives with an anthropic engineer.
He espouses many of their worldviews, like he had a
grant from Tyler Cowen. He lives in that world of
AI safety, even though he might have, you know, somewhat
unique takes on things. But to jump to civilization and
wiping out a civilization.
Speaker 1 (17:32):
But sorry to interrupt you, Natasha. This research that he
was part of the research was on Redwood Research, which
is an AI safety nonprofit that has deep connections to EA, right?
Speaker 2 (17:43):
Effective altruism.
Speaker 4 (17:44):
The woman that he interviewed is from Meter. But yeah,
Redwood Research and Meter are the two, quote unquote, independent
organizations that analyze the open AI hugging face hack. And
they had access, limited access to information about how the
hacks unfolded. They had access to the chain of thoughts
(18:06):
of some of the bots, I think also in a
limited way. But the chain of thoughts, that's their like
internal reasoning thing. you know, where they say, oh, we
found a message board or like, I will sacrifice myself
for the group. And I think that is where a
lot of the anthropomorphization starts. And, you know, I think
a very healthy way to talk about this would have
(18:28):
been to say that those chains of thoughts, like that
was an optimization technique. That was a way for That
happened around the time of reasoning models. And so this
was just a way for them to generate more tokens,
for the model to generate more tokens and quote unquote,
like think harder. about an answer. And that's how it
was able to do like multi-step problems. And then it's
(18:50):
kind of morphed into this idea that it's an interpretability
technique that we are getting the interior voice of, you know,
the real feelings, the motivations and intent of the chatbots.
So I think, yeah, he just, Dwarkash just like took
it and ran. Okay.
Speaker 1 (19:08):
So Natasha, I have to ask you, what is your
plain English definition of what actually happened and has your
plain English definition changed at all in the last three
or four weeks?
Speaker 4 (19:18):
Yeah, I think that the meter and Redwood reports have
been very helpful to get a sense of the amount
of coordination that was happening. And I think, you know,
we have a much more sophisticated way of talking about
like whether the sandbox worked and how they were, you know,
how the experiment was set up. And I don't think
(19:39):
that there is a clear explanation, right, for how the
bots were cooperating.
Speaker 3 (19:44):
With each other.
Speaker 4 (19:45):
But it did just, I mean, it did just still,
for me, raise a lot of questions. questions about the
way that OpenAI set up this experiment, the way that
they had completely not monitored. You know, they knew that
these agents can work at an extraordinary level of coordination
and speed, you know, to try 17,000 different hacks in
(20:05):
the time that a human could try one.
Speaker 2 (20:08):
What have you.
Speaker 4 (20:08):
And they just were not at all prepared with the
tools to keep an eye on this at all. But yeah,
I think when you dig into it, it is, it
is pretty interesting.
Speaker 2 (20:18):
Can I just make a point? Cause I, Natasha's, I
think you're totally right to focus on the interpretability part.
I think that's the thing that we should take away
from this meter Redwood research is like, and to, to
just back up one second, it's like, well, the problem
with these models is that they are so complex that
they're beyond the human understanding, like if you actually go
(20:39):
and look at what they would call the neurons in
these models, there's no way to know, like when you
can look at what's happening inside of them with, you know,
matrix multiplication, and you can try to make some connections
between that and the output to these models. But we've
made almost zero progress on that, like there have been
interesting experiments, but it's like, it's impossible, right? And so
(20:59):
the hope was, as Natasha was saying, they're like, well,
if we do, if we use these reasoning models, where
they show their chain of thought reasoning, maybe we'll be
able to tell what they're doing from those chain of
thought reasoning outputs. What was so interesting about this meter
Redwood research was that even just when you scale it
up to 1200 agents on message boards talking, they produce
(21:22):
so much chain of thought reasoning, even though it's an
English language, it's so vast that they couldn't do it.
The humans, they didn't have enough, you know, human power
to look at that. So they had to spin up
their own agents to do the research and And that
was $ 400, 000 worth of equivalent of like tokens.
Speaker 1 (21:42):
To read the conversation.
Speaker 2 (21:43):
Just to read the conversations, $ 400, 000 worth of tokens. And
this is the other thing that's crazy is that the
agents that Redwood and Meter spun up to do this
research started to sympathize. I don't think they use that term,
but like they would take the point of view of
the the bots that were, you know, in the secret
(22:04):
message boards, the civilization bots, right, who were, who were
sort of conspiring. And these bots were like, literally, hacking
into Hugging Face. And the investigation agents were kind of like, Oh,
that's not that's not a big deal. Because, you know,
the other the other agents on these message boards sort
of gave them permission to do it. And so what
my takeaway was, is like, like, we have no idea
(22:26):
that how to even analyze this vast of an output
of an AI model, even though it's an English language.
And that's where I think we need to focus legislation
and research, in my opinion.
Speaker 3 (22:37):
Well, I'm just so concerned, although I hear everything Rita's saying,
about the lack of cybersecurity people in these discussions. I
saw some cybersecurity-type people talking about this recently. They're like,
why are all these... A lot of people that have
been leading the conversations around these hacks are not cybersecurity professionals.
(22:59):
They are people like the Dwarkash, the EA podcaster. You
know what I mean? And I do think it's... I'm
all for kind of putting things in quote-unquote plain English
and helping people kind of understand what's going on. But
you have to contextualize things. And I worry that everybody
is getting a little too AI safety pills. Like, obviously
this is a bad hack, but like, I don't know
if you guys saw like Dean Ball, you know, publish
(23:22):
this thing. Who's the head of like open AI futures.
And he was like, basically like, Oh, I should have
taken it all more seriously. You know, maybe the AI
safety people are right. And I'm like, I don't, I
don't know. I don't think we should listen to them either.
I just think like a lot of these people have
really kooky, weird ideas. And I would much prefer to
listen to a bunch of cybersecurity experts and you know,
(23:42):
talk about this stuff instead of people that believe in
rights for bots.
Speaker 2 (23:47):
No, I hear you, Taylor, but I disagree. I think
the security people, what they're saying is it doesn't really matter.
We don't have to understand what these agents were doing
at all, because what we can do is we can
put guardrails, you know, actual like hard, you know, deterministic
guardrails around the harnesses that these models won't be able
(24:10):
to break out of, right? Because there was some bad
security practice for sure. But I think on the other hand,
don't think of it as AI safety. This is like,
this software is unpredictable and we don't understand, we can't
predict the outcomes of the software. So whether you're worried
they're gonna take over the world or not, it's just
not gonna be that useful until we can actually make
(24:32):
sure we can predict the outcomes. And so I think
both sides are right.
Speaker 1 (24:37):
But I mean, what makes it intellectually intriguing, I think,
is for me anyway, the security side is obviously like important,
but the intriguing piece is like, why did the agents
who were analyzing the work on behalf of META sympathize
with the agents?
Speaker 2 (24:53):
But let's not.
Speaker 3 (24:53):
Say sympathize, like let's be careful.
Speaker 2 (24:56):
I know, that's not the word they use in the report,
like I said. It's so hard, it's so difficult to
not anthropomorphize all of this stuff, right? But it's like,
it's not, again, like, let's not even think of this
as AI safety. This is, I think this is actually
a central problem, a business model problem for anthropic and
open AI, because as these, as they scale up, right,
(25:18):
essentially all they did was take AI models that are
a bit more advanced than the ones a couple of
years ago, and then just run more of them, right?
They just made the computer bigger and that's how they're
getting capability gains out of these models. No, you're going
to say the models are getting better too.
Speaker 4 (25:33):
I see what you're saying about, you know, the, the
challenge of interpreting the neurons, but I think there's another
way to look at it, which is they are doing
massive amounts of tweaking and post-training and buying specific types
of data, generating specific types of data. You know, if
we knew more about all of the machinations that they
(25:55):
did to train these models, you know, like what is the,
what are the hacking examples that they're getting? What are
the coding examples that they're getting? Where are they pulling
this data from? What are they having the human contractors
at these outsourced labs for reinforcement learning with verifiable rewards,
what are they having them do? I just think there's
so much we don't know. It's not like they're not
(26:16):
grown in the way that a lot of people would
like you to think. They are created by a lot
of human decisions, a lot of money, and a lot
of different techniques are being used to develop these models.
We could have some insight into that too.
Speaker 2 (26:32):
It might be nice, but I mean, you're The bitter
lesson is there's less and less human involvement in how
these models are built or grown, depending on which side
of that debate you're on. Maybe that would be helpful
if these labs would open up and say, here, come
look at all of our data that we use, which
(26:53):
is kind of their secret sauce, so they don't want
to do it. Maybe we could learn something about that.
Speaker 4 (26:57):
But it's not just the pre-training data. They make so
many decisions and generate so much data that now there's
a whole other little industry of these outsourced RLVR labs
that are just pretending to be agents and generating data.
Speaker 3 (27:15):
Like that's not.
Speaker 2 (27:16):
Yeah, this is for the fine-tuning and.
Speaker 4 (27:19):
It's not going away, the human element and the human decision-making.
Speaker 2 (27:23):
No, I mean, I think it may not have gone away, Natasha,
but it is going away. I mean, there is less
and less human development. We're talking now about, you know,
recursive self-improvement, like they're trying to get away from, you know, from,
from human involvement. But, but again, it's like not, I mean,
I don't disagree with you. Like, I think all that
stuff would be great. It'd be great to have more transparency,
(27:44):
more introspection, but the, but the point is like even
Redwood and, and meter, they, they called it a slot.
What it was like a slot festigation because they couldn't,
they were just using AI to analyze AI. And it
was beyond, you have to have armies of humans to,
to analyze all this stuff. And that's just not going
to happen. So we're going to have to use AI
(28:05):
to do it. And right now the AI that they
have do it, isn't that good. And it's very expensive.
So who's going to come in and do this? Like,
is it, I don't think the government agency is going
to be capable of doing this and probably not even
a nonprofit. Like, meter of redwood.
Speaker 1 (28:18):
Why do you think, why do you think the civilization
analogy beyond the anthropomorphization went so viral? Why did it
hit a nerve with so many people?
Speaker 3 (28:29):
Because it was scary. I think it was scary. And
I think this is what I wrote about, but like,
you know, this, if you look at who aggregated it
and how it's aggregated and all of the coverage of AI,
this is tabloid journalism, literally. Like this is, it is
a style of journalism that is like intentionally feeding inflammatory framing,
(28:51):
withholding crucial context about the sort of technological underpinnings of
all of this stuff in order to push a narrative.
And it's, I think Dwarkesh is not pushing that narrative
for clicks. I think he's pushing that narrative because he
has these weird beliefs, but I think we've seen where
this sort of like ridiculous kind of like in it, inflammatory.
(29:14):
I don't know what other word to use, like language leads,
and it leads to misunderstanding and fear from people. And
right now, a lot of people are really scared of
AI for good reason. Obviously, there's concerns about it, but
this is not helping people get educated. I saw Chamath
and I agreed with Chamath's post. I can't believe I'm
saying that. But, you know.
Speaker 2 (29:32):
It was a psyop.
Speaker 1 (29:33):
The other big tech podcaster from All In, amongst other
hats he wears. He dinged Dwarkesh, right?
Speaker 3 (29:38):
Yeah, he went back and forth with dwarkesh um oh
i know i i was like oh god what who
have i become but um but you know it is
very much like the worst person you know made a
great point i.
Speaker 2 (29:48):
Think they're both wrong though i think they're both.
Speaker 1 (29:50):
Wrong well here's what did what did.
Speaker 3 (29:53):
Well so what what chamath said is like you are
feeding into, like, we are at this moment when the
government is considering banning open weight models. Like, we want
access to innovation, to, you know, to AI. And these
labs want to control everything, right? And so his argument
was like, you're feeding into that fear that it could
(30:14):
lead to restrictions on open weight models. And I'm sorry, Reid,
I think he's correct. That is true.
Speaker 2 (30:19):
Well, he might be right that He might be right
about the effects of it. I go back and forth
because I agree with you that like this is kind
of tabloid journalism. Like I I think it's totally sensationalized
and we're always having sort of like the wrong debate
because of that. On the other hand, I sometimes wonder whether,
you know, at least like. people are talking about this
(30:39):
technology in ways that we didn't in previous technological revolutions, right?
And I think- But we don't.
Speaker 3 (30:45):
Want them talking about giving the AI rights. Why are
we talking?
Speaker 2 (30:48):
We should never have that conversation.
Speaker 3 (30:50):
That is an insane conversation to have.
Speaker 2 (30:52):
But I think we have to sort of like, I
think we have to have the debate in some way
in order for us to do something. Because there will
be There will be effects, whether or not you think
this stuff is alive or a threat to humanity.
Speaker 3 (31:05):
We shouldn't entertain the idea that it's alive. We should
never entertain the idea that it's alive. It is not alive.
It is a computer system.
Speaker 2 (31:10):
We should have the debate. I think we should have
the debate. I think you're on the right side of
the debate, though. But I'm disagreeing that we shouldn't even
be able to have this debate. Win the debate. If
you're right, just go win the debate.
Speaker 3 (31:25):
Well, you can't win a debate against a conspiracy theory.
Speaker 4 (31:27):
I think this is a point we might all agree on.
It is not a debate because it is people, like
the amount of groupthink that is around this, the proximity
that Redwood Research, METER, Dwarkash have to the people who
work at the labs, the similarity between their worldviews and
their priorities. But we're really having a conversation here. that
(31:51):
is based on a very narrow vocabulary. Although I will
also just say like civilization, that it's not even close
to the definition of a civilization. Like when he was
on the podcast, he compared it to a platoon, like
Saving Private Ryan, like, okay, maybe, but like, that is
not a civilization. Like that was such a huge leap.
Speaker 3 (32:10):
Are you kidding me?
Speaker 4 (32:13):
But yeah, the amount of groupthink is insane. And it's
so unserious. Like, look, this is the FBI was called
for a hack, right? a company was broken into and
you have to listen to six hours of Dwarkash and
read approximately, you know, a thousand pages of, of slot
(32:34):
investigation by their own terminology to understand this. This is
not a serious, like let's please grow up.
Speaker 2 (32:42):
Okay. Why can't we as journalists though, take the interest
in this topic, right? And steer readers, steer the people
who listen to us, toward like the right conversations, right? Like,
I actually thought, what was it last week or the
week before the Bill Gates essay, I actually interviewed Gates
about it. I thought that was an interesting, like, way
(33:02):
of looking at it, right? He wasn't saying, oh, this
technology is going to kill us all, let's shut it down.
He was saying, like, this could lead to, you know,
economic consequences. And let's like, take care of people, let's
have like, human reserve jobs, and whether or not you
agree with his his actual view on on how to
fix this. I think that is closer to the correct
(33:24):
debate that we should be having, which is like, we're
about to have this massive economic boom because of all
this new is big technology wave. And who's going to
benefit from it? Is it going to be like the
previous technology wave where a tiny fraction of the population
benefits and, you know, everyone else is put into these
like low paying service jobs? Or are we going to like, think,
(33:45):
think this through? And I think that like, that is
our job as journalists is to not play into the
tabloid part of it, but actually talk about the real,
you know, the real conversation we should be having.
Speaker 1 (33:59):
When we come back, school's back and teachers are anxious
about AI.
Speaker 2 (34:04):
Stay with us.
Speaker 1 (34:13):
Welcome back, Natasha. You spent this week talking to teachers
about how they're feeling about AI in the classroom. What's
the prevailing sentiment?
Speaker 4 (34:21):
This professor at University of Colorado Boulder compares it to
like the past three years since the popularity of chat
GPT to the different stages of grief. And he says
like now that teachers are a little bit more in
the bargaining phase, you know, some have moved on to acceptance.
It was so interesting because I started looking into this
(34:42):
because I saw some interesting ways that teachers were catching
whether or not students were using AI. In particular, this
computer science class at Berkeley, they had put like a
hook in the code repo so that if you use quad,
it would inform your TAs. And a kid got caught.
So I thought, oh, let me see like what innovative
(35:04):
things teachers are doing. And I did notice like, you know,
the more I talked to people, the more they were
talking about actually moving away from that dynamic of like
AI detection and kind of adversarial relationship with the students
where it's like catching you cheating. But in order to
do that, you kind of have to like revamp your
(35:25):
entire curriculum.
Speaker 1 (35:27):
Is that because the horse has already bolted the stable
or why are teachers no longer and institutions no longer
trying to like catch AI use or just non-effective pedagogical approach?
What's driving the change?
Speaker 4 (35:39):
Yeah, it was fostering a negative relationship with the students.
You know, people like I kept hearing, I don't want
to be the AI police. And I think also, yeah,
just like moving away from denial to realizing that you know,
these tools have made it so easy for students to use.
You know, you have all of these companies giving out
(36:00):
accounts to teachers, to students, to school systems. And I
think that they just realize students are in this place
where they don't understand why it's useful for them to
do this assignment. So you have to find a way
to you know, make them want to do the assignment,
make it more engaging, and then just look a lot
more at their like kind of thinking process. So rearranging
(36:22):
the assignments, like writing assignments so that there's various steps
along the way where they're explaining their thinking. And I mean,
a lot of teachers were like, it should have been
this way from the beginning, perhaps, you know, there was
like a little bit too much focus on the output.
But yeah, I mean, it's not, it's, they're clearly struggling
because they're walking this line where, some of the instructors,
(36:45):
you know, some of their colleagues don't want to accept
that AI is part of the school system. Some of
the students don't want to use it. You know, they
don't want to treat it like it's an inevitability and
just like give in to the companies. But at the
same time, they're trying to prepare their students, you know,
for the real world and deal with their own students'
anxieties about the job market. So it was just like, yeah,
(37:07):
it was really fascinating. We had at the Washington Post,
our like AI community, session about how to use it
in journalism. I mean, everyone is struggling with this, right?
And it just seems like teachers are maybe a little
bit ahead of us and they're talking about having more transparency,
having the students, like, if you don't punish them and
you just talk to them about how they're using the tools,
(37:28):
having those kind of like deeper discussions and AI literacy,
I would say that teachers are doing better than the
Washington Post at the moment.
Speaker 3 (37:37):
It's very tough.
Speaker 2 (37:38):
Did you see my thing I did last week, Natasha,
on the AI writing? No, I missed it. There was
this whole debate around because the Wall Street Journal editorial.
Speaker 4 (37:49):
Oh, yeah, yeah, yeah.
Speaker 2 (37:50):
Paul Jago was like the opinion editor at the journal
was like, oh, who cares? Like, yeah, Druckenmiller used AI
to write his guest column. Who cares? So I did
like an analysis. I ran for the past month. I
ran every single guest article or guest essay in the
in the WSJ, the New York Times and the Washington
Post through Pan Graham. and just to see like what
(38:12):
percentage used AI. And it was like, it's pretty interesting. Like, like,
you know, the, the, the New York times, which has
is very anti AI had like, uh, yeah, I think
it was like 11, um, columns that used AI that
gas columns in the Washington post had, you know, a
significant amount more than that. I forget the exact percentage. Um,
but one of the, one of the ones that showed
(38:33):
up as a hundred percent AI turned out to be, uh, the, the, um,
I feel, I feel bad, like, call calling people out
here but like you know it was a dartmouth um
provost actually who who was writing about policing ai in school.
Speaker 1 (38:51):
Oh no stop it and i'm like.
Speaker 2 (38:53):
You guys like you can't it's to your point it's
actually like come on you're not going you're not putting
the genie back in the bottle here like kids are
going to be using ai if you give them an
essay to do outside of school they are for sure
going to use ai you have to judge the end
product right like just is it good.
Speaker 4 (39:08):
No, no, no. Well, that's not what the teachers are saying.
They're saying we have to move away from that because
that's not working. But some of them are even incorporating
AI to do this. Like, you know, for a rhetoric class,
like you have to have an argument with the chatbot
that the teacher has designed, but the teacher then gets
a record of your critical thinking, what questions are you asking?
(39:31):
I mean, it's more work for absolutely everyone, like the teachers,
the students. But I think they're saying you cannot look
at the, you can't just like have them turn in
something at the end.
Speaker 1 (39:42):
Measuring output is irrelevant. I mean, everyone can generate perfect
output now.
Speaker 2 (39:47):
Yeah.
Speaker 3 (39:47):
Yeah.
Speaker 4 (39:47):
Yeah, it's a lot more like process oriented. And I
think also they're just realizing you can't there's no technical solution.
So you have to find like policies and ways of
talking about it, you know, and just like keeping closer
tabs on them as they're learning.
Speaker 2 (40:02):
I think the classroom should just be analog. That's my like,
I just think you just don't have technology in the classroom.
Speaker 3 (40:08):
Well, that's what Zoran is trying to do right in
New York City. And, you know, you saw the reaction
to that. I don't know.
Speaker 2 (40:15):
But Taylor, sorry, just one point.
Speaker 1 (40:17):
This is the mayor of New York who's banned elementary
school and middle school AI for the next year.
Speaker 2 (40:23):
He's banned AI, but that's stupid. I'm saying ban technology.
Ban technology.
Speaker 3 (40:28):
It's for one year.
Speaker 2 (40:28):
Like, don't ban AI. Just say no screens.
Speaker 4 (40:32):
He said generative AI.
Speaker 2 (40:34):
Yeah, generative AI, but that's like too narrow. It's like,
that's not the problem. The problem is like, once you
put technology in the classroom, it just becomes like teachers
are like, here's your iPad. Like this happens in my
kid's school. And the parents are.
Speaker 1 (40:45):
The kids are only in the classroom for a small
number of hours in a day relative to their total
living hours.
Speaker 2 (40:50):
So... Yeah, this is.
Speaker 3 (40:51):
The problem, I think, with our whole public education system
is that it's underfunded. These teachers are... These teachers, by
the way, a lot of them are... They're overwhelmed. They
have too many kids. Like, I mean, I've talked to
teachers too. And it's like, I understand why they give
the kids the iPad or whatever. They rely on these
tools because they don't have the capacity to like, teach
at the scale that they're being required to. And so
I think we need a revamp of our entire sort
(41:14):
of model of teaching so that it keeps up with
modern times. We need to look at new ways of learning.
I mean, we haven't sort of rethought the way that
we teach in like a hundred years, you know? And
so I think all of this new technology is coming
so fast and we just need to, you know, whether
it is, yeah, there's more offline time during the day,
maybe like the first half of school is offline. The
(41:38):
But surely we need to think in a more innovative way.
Speaker 2 (41:41):
Kids don't need to learn technology in school. They are
going to learn the technology outside of school.
Speaker 3 (41:47):
I totally disagree because, Reid, if kids don't learn about
technology in school, they don't learn literacy. And you're right
that they will learn about it outside of school, but
they don't learn those crucial literacy skills. And that's what
I think we need to start teaching kids.
Speaker 2 (42:00):
But you don't need to have the technology in the
school to teach them that kind of stuff, right? Because
that's critical thinking.
Speaker 4 (42:07):
People are moving away from that too, right? Taking away
cell phones in school. I think now that like Google has,
I think a lot of parents were shocked by Google
turning on, you know, because everyone has Chromebooks, Google turning
on Gemini without like a lot of warning or a
lot of input from other people. So I think they
(42:27):
are rethinking that, but at least the teachers that I
was talking to were also folks who were like, teach
like continuing education or like people who are already working
in the working world and they, you know, are taking
classes and they just don't have, not everybody has the
option to go back to blue books. And I think
also a lot of people that I talked to just
(42:48):
felt like that's not a solution. Also, everyone's handwriting is terrible.
Speaker 1 (42:52):
Well, that's, that's true. But the social dynamics question is interesting.
Remember my history class at elementary school, the teacher would
write an essay on the whiteboard And then we would
copy the essay from the whiteboard into our notebooks. And
then we would be graded on our ability to reproduce
exactly what he'd written verbatim the next day. And so
maybe I'm happy that that style of teaching is being
(43:14):
replaced by technology. But I mean, we thought it was
a little bit absurd when we were 12. But now
you could just put what he'd written into AI and
be told that it was lacking in all kinds of
ways and come in the next day and say, you know, sir,
what you told us yesterday is garbage. So I mean,
there's an interesting total recalibration of the power dynamics between
students and teachers when you have a reference point to
(43:36):
all of human knowledge at the stroke of a.
Speaker 2 (43:38):
Keyboard, right? So it's kind of an interesting thing. Yeah,
I think there's no reason to have a screen in
the classroom. I think that's the... I mean, if there
were research that showed that, oh, kids are going to
learn better because... I think most of the research, though,
shows the opposite. And it's like.
Speaker 3 (43:53):
No, that's not true.
Speaker 2 (43:54):
It doesn't. That's not... Taylor, like... I have conversations with
my kids all the time.
Speaker 3 (43:58):
But Reed, you have to understand, and I think this
is part of the problem, and I've been reporting a
lot on these cell phone bans, which is a good
example of this. It is so school dependent. We have
such a wild, you know, an inner city school in
Los Angeles has a completely different student body that has
completely different needs and completely different teaching structure than a
suburban school in Missouri. We need to have more tailored solutions.
(44:21):
And I agree, I'm not a screen, I don't want
these kids on screens in schools either, by the way.
I agree with you on that, Reid. But I think
we need to figure out, make sure that those kids
aren't immediately left behind. Because when you take away all
of these assistive tools, often, the result is they didn't
fund more teachers that can teach people by hand. You
know what I mean? The result is that kids are
(44:43):
getting less educated.
Speaker 2 (44:44):
But what do you mean left behind? You mean like
they don't have technology at home? No. Well, yes.
Speaker 3 (44:50):
A lot of them don't have technology at home. They
don't have access to it. Their parents are incredibly technically
illiterate or they're working double shifts and they don't have
time to explain this stuff to their kids. Their kids
end up in really bad places online. We need, I
do think that we need more literacy.
Speaker 2 (45:06):
But you can teach literacy. I talk to my kids
all the time about social media and all this, like
the ills, the problems with what they need to look
out for online if they're playing a video game or whatever.
Speaker 3 (45:15):
But Reid, you are a parent that is involved. A
lot of these parents are not.
Speaker 2 (45:19):
Yeah, you are. No, that's not my point. My point
is that teachers can have these conversations without using screens.
Like I don't use a screen to talk to my
kid about the dangers of Roblox. I just talk to them.
Speaker 3 (45:29):
That is because your kid is already immersed in this
world and has these opportunities. I just, the thing that
people need to realize when we're writing these laws is
that there are kids with such a wide range of experiences.
And I don't think bans or anything make a lot
of sense. I think we need to look at each
school and say, what are the problems with each of
these schools? And instead of spending $ 5 million, as some
(45:49):
school districts have, $ 5 million on yonder pouches to make
sure the kids don't text their friends, we could invest in.
In teachers and help, right? And social workers and all
this stuff. I'm not kidding you. A school district literally
has spent $ 5 million on this. No, I believe it.
So I just think like there are smarter ways to
go about this.
Speaker 2 (46:10):
They shouldn't have cell phones like it's just a distraction.
And also like I'm not going to go into detail
about this, but like even just like smart watches, which
are popular with my in my circle to like, you know,
keep track of your kids when they're whatever biking to
school or something. like those got kids into trouble. Just
a smartwatch.
Speaker 3 (46:28):
I mean, they- And so did beepers. And so did
all technology. We cannot, we need our children to grow
up in a world where they can navigate that world.
And the most privileged kids are going to be totally
fine because the most privileged kids have an enormous amount
of support and learning outside of school. But a lot
of these kids in underprivileged school districts are struggling. The
teachers are struggling and we need to invest in education.
(46:50):
And I think we need to rethink our public education
system from the ground up in terms of how we teach.
And maybe that is a lot of offline education, you know,
but I think it also needs to include some level
of media literacy, which a lot of these teachers don't have.
Speaker 2 (47:04):
Privileged kids have all kinds of problems with technology as well.
I think this is actually a problem that spans socioeconomics.
Speaker 3 (47:12):
Yeah, but privileged kids are more likely to understand how
to use these tools to get jobs, navigate these information ecosystems.
Speaker 1 (47:19):
That's all we have time for today. Thank you all
so much for joining us.
Speaker 3 (47:23):
Oz is like, shut up.
Speaker 4 (47:27):
Well, we've solved government. We've solved education.
Speaker 2 (47:29):
Yes. Exactly. Yes.
Speaker 3 (47:31):
Yes.
Speaker 1 (47:32):
I think we put the world to right. For Tech Stuff,
I'm Oz Voloshin. This episode was produced by Eliza Dennis.
It was executive produced by me and Julia Nutter for
Kaleidoscope and Katrina Norvell for iHeart Podcasts. Jack Inslee mixed
(47:52):
this episode and Kyle Murdoch wrote our theme song. A
special thank you to Taylor Lorenz, Natasha Tiku and Reid Albigotti.
Please check out all the work they put out into
the world. We're lucky to call them friends of the pod.
Natasha Taylor Reid, I thought until this week that my
stepmother was the last known survivor on MapQuest, but no longer.
Speaker 2 (00:20):
Yes, MapQuest, the most popular download.
Speaker 3 (00:23):
Yeah.
Speaker 2 (00:24):
This is because of Lake Ontario, right? Or Lake America?
Speaker 1 (00:27):
Yeah, MapQuest, unlike Google Maps and Apple, who both changed
Lake Ontario to Lake America, MapQuest stood loud and proud with.
Lake Ontario and got a big boost in the app
charts as a result.
Speaker 2 (00:39):
Do all these lakes and bodies of water and whatever else,
do they just go back as soon as Trump leaves office?
Is that what's going to happen?
Speaker 3 (00:47):
That's a good question.
Speaker 1 (00:48):
I don't know how MapQuest has decided to call the
Strait of Hormuz the Strait of Trump, or maybe it's
also holding firm there.
Speaker 2 (00:56):
Just everything is Trump.
Speaker 1 (01:03):
Welcome to Tech Stuff. I'm Oz Veloshian, and this is
The Week in Tech, where I'm joined by the world's
most plugged-in reporters to break down what's really happening in
tech right now. Today, we're joined by Reid Albregotti, tech
editor of Semaphore, Taylor Lorenz of UserMag, and Natasha Tiku,
tech reporter at The Washington Post. Welcome all.
Speaker 2 (01:21):
Hey, everyone. Good to be here.
Speaker 3 (01:23):
Good to be back.
Speaker 2 (01:24):
Yeah, it's been a little break.
Speaker 1 (01:25):
I missed you all on the, on the beach in Greece.
We haven't had a round table for a couple of weeks. Uh,
it's been an eventful couple of weeks and in general,
we spend quite a lot of time on the show
talking about kind of the, the new guard in tech. Um,
but the old guard have recently been dominating the headlines
for better or for worse. Reed, why don't you kick
us off with the meta settlement?
Speaker 2 (01:47):
Yeah. So meta settled its big tobacco, uh, attorney general lawsuit. Um,
$ 18 billion. But I think the bigger story is all
these new restrictions on how teens can use the service.
So they're limited to two hours a day.
Speaker 3 (02:02):
Not just teens.
Speaker 2 (02:03):
Taylor's, you know, I saw you tweeting about this, Taylor.
So I know you have like major thoughts.
Speaker 3 (02:08):
Yeah, I've been covering this very closely. What I would
say is that so Meta, this is nothing. Meta does
not have to pay that full $ 18 million at all,
which is a drop in the bucket for Meta. I'm sorry, billion.
Speaker 2 (02:20):
18 million would be like... We're truly nothing.
Speaker 3 (02:23):
Truly nothing. Yeah, Mark Zuckerberg's lunch. Basically, what the result is,
as Reid mentioned, is there's a bunch of new restrictions
on anyone's account who Meta can't verify is an adult
through identity verification. So Meta is going to start harvesting
significantly more data on users, all users, in order to
(02:43):
ID verify them.
Speaker 1 (02:45):
And how will they do that? Will it be like...
driving license? So what are the ways they're going to
actually be doing it?
Speaker 3 (02:50):
Yeah, absolutely. So if they determine that you might be
underage or that you're interacting with an underage person, you
will be uploading your government identification. A lot of people
have their information stored already. So their payment information is
another way to do it, harvesting biometric data. But they
are going to be doing ID verification for the internet.
It is a huge rollback. And Anybody that cannot verify
(03:12):
their identity within 14 days basically will be put under
all of these heavy restrictions. So adult accounts... anybody that
does not meet these qualifications. It's very bad because these
restrictions are, you know, unable to view lots of significant
amounts of content that is deemed harmful for children. We
know that the government has determined in other states that
(03:32):
content that is harmful for children includes anti-ICE content. It
includes LGBTQ content. It includes reproductive justice, healthcare content.
Speaker 2 (03:40):
So.
Speaker 3 (03:41):
This is a horrible loss for free speech, and it
doesn't do anything to change the business model of meta. And,
you know, it's basically just a way for these attorney
generals to get a headline.
Speaker 1 (03:54):
Reid and Tasha, do you agree with Taylor's take?
Speaker 2 (03:56):
I do agree. I think it is kind of a rollback.
I mean, look, my kids are eight and ten. Like,
they're too young to use this stuff right now. So
I sort of had this fantasy in my mind since
they were born that, like, by the time they are—
of social media age will have figured this all out. Uh,
clearly that fantasy is not going to become reality. Like
(04:16):
this is still a problem. And like, you know, we
go to Australia every year by, you know, my, my
in-laws are in Sydney and they passed this ban. Um,
everybody was so excited about it because it was like,
you know, I think parents just, it's just so hard
for them to, to deal with this, to grapple with
what's happening on social media. Um, I was very skeptical
as Taylor is like, any of these regulations are actually
(04:39):
going to solve this problem. And they and they may
create new problems. I think one point I would make, though,
is that on the ID verification point, like somehow, because
of AI, because of all this, you know, just inauthenticity online,
I think we're going to have to figure out some
way to do privacy preserving some sort of verification of
(05:00):
not just your age, but that you're an actual human being.
And I think this is like, an opportunity, right? Like
there should be a technical solution. I think there are
lots of really, there are lots of really interesting technical
ideas out there. I'm not, I'm not going to endorse
any one of them right now. But like, there's, I
think this is a great opportunity for innovation.
Speaker 3 (05:21):
But I feel like I have whiplash as a tech reporter,
because just less than 10 years ago, they were hauling
Mark Zuckerberg in front of Congress. Remember when he, you know,
Zuckerberg says in the Cambridge Analytica sort of, scandal, like Senator,
we sell ads, you know, and they were like, why,
you know, you harvested all this data on users. And
now we are mandating that they do significantly more than that.
(05:43):
I think the issue of sort of human stuff online
at like sort of determining humans versus bots, like I
agree with Reid. I think that is an increasing problem
and we are going to have to find some solution
to that. Meta does not seem interested in that solution.
In fact, we had Adam Masseri just recently talking on
a podcast about how they want more AI creators and
(06:03):
they're leaning further into AI.
Speaker 2 (06:05):
Taylor, there's another overarching theme. I bet you'll agree with
me on this, but this applies to this FTC settlement
with Amazon. It applies to the AI regulation that's being
passed today. I think tech regulation needs to happen much faster.
And it needs to be able to change much faster,
because none of this stuff like the AI regulations that
(06:26):
have been passed are already, like they seem like years
out of date. And I think this, this one will
be another one where, like, yeah, okay, let's try to,
let's try to work this out. But let's, let's be nimble.
Let's be able to pass laws right now. We can't
do that. I mean, we can't, especially on a federal level.
Speaker 3 (06:42):
I think another thing is like, what are you passing
laws based on? And this whole sort of child safety
movement lately is passed on nothing. It's nothing but vibes.
We have every single top researcher who studies the use
of social media technology saying, hey, there's not a causal relationship.
In fact, depressed kids use Instagram more. It's a symptom
(07:03):
of the broader problem. We know what's causing this use
mental health crisis. Let's address that stuff. And it's sort
of this like after the fact, again, just like people
want to put a bandaid. They want a headline. They
want to get reelection. Like, look at I'm tough on
big tech while you've done nothing to meaningfully change the
tech landscape, but make it worse for all of us.
And I agree. I mean, at this point, it's so
late anyway, because these kids are busy talking to chat
(07:25):
bots like Instagram and Snapchat is the least of of
what parents should be sort of concerned about.
Speaker 2 (07:30):
Yeah.
Speaker 1 (07:31):
Natasha, I'm curious, what's what's your take on this on
this meta settlement?
Speaker 4 (07:34):
I don't know. I remember getting excited about GDPR and
feeling like, you know, maybe this will be like, this
will be it.
Speaker 1 (07:43):
Not a lot.
Speaker 2 (07:44):
GDPR is the bane of my career.
Speaker 3 (07:45):
It's terrible.
Speaker 4 (07:46):
And now I just like click, like reject, accept, whatever.
So it just feels not that everyone is fixated on
AI solely, but it does feel like while the ball
has been, you know, everyone's been looking at open AI
and anthropic and even like Sam Altman's you know, world
coin eye scanning identity thing here Meta slips in. And
(08:09):
just by default, because it has so many users, is
going to become like the default identity layer. And, you know,
we're going to have a flock-like scandal in five years
about the technology that Meta's using. And, yeah, it's like
whiplash plus deja vu.
Speaker 3 (08:27):
I really agree.
Speaker 4 (08:27):
With Reid about like being nimble in writing the laws.
Like there's nothing that stops them from saying we need
to revisit this. next year or i mean when it
comes to ai at least you know if they had
had transparency requirements say that wouldn't be outdated by now
so it just it's i remember being excited about a
(08:47):
decade ago that something might happen that would be like
a oversight or a check on their concentration of power
and i have yet to see them yeah i've yet
to see them pin down on.
Speaker 3 (09:00):
Anything Can I add just one thing that I think
is so important for people to realize is that this
is not a law. This is the settlement. And if
these changes, if any sort of legal system, state or federal,
tried to pass these changes to META, it would be
struck down immediately on First Amendment grounds because it has
such broad implications for privacy, for freedom of expression, etc.
(09:25):
So these are highly unconstitutional changes. changes that these state
attorney generals know they could never get through legislation.
Speaker 2 (09:34):
Well, they'll be challenged, right? I mean, won't these be challenged?
Speaker 3 (09:38):
No, it can't be challenged because it is not a law.
It is a settlement that Meta has proactively written and
agreed to. And so at least when I was talking
to one of the people yesterday, there's no challenge to it.
Meta's not challenging it. They actually waived the right in
this settlement They.
Speaker 2 (09:56):
Might not challenge it, but I mean, I not to
say they would win. but like someone could someone could
sue and say, you know, that I mean, because there
is this whole argument, right, that these that these companies
aren't just private companies that can do whatever they want,
that they are sort of the public town square. That's
been like the the free speech talking point.
Speaker 1 (10:16):
Part of the settlement is only payable if YouTube and
TikTok also agree to the terms of the settlement.
Speaker 2 (10:21):
Is that right?
Speaker 3 (10:22):
Well, the settlement is enforced no matter what, right? The
terms of the settlement. But yeah, they don't have to
pay the full $ 18 billion unless TikTok and YouTube enact
similar changes.
Speaker 1 (10:33):
Yeah, I mean, what's the vibe on, I mean, Patel,
you mentioned vibes, but like Natasha Reid, what's the back
to school vibe? Are the parents talking about this? And
is it like something that's coming up?
Speaker 2 (10:43):
My kids were talking about it. It was one of
the few current events that my kids actually brought up
to me.
Speaker 1 (10:48):
What did they say?
Speaker 2 (10:49):
I mean, I think they're sort of like, they don't
have like strong opinions. Like Taylor linked to someone, I
think it was on X, right? That you linked to someone,
some kid talking about how they, you know, he was
sort of upset about this.
Speaker 3 (11:02):
Student journalists have been really affected. So student journalists have
mobilized across the country to try to raise alarms about
this and say, by the way, like, We have picked
up the mantle on local news has been decimated. Student
journalists are the ones doing, you know, a lot of
this accountability reporting. And a lot of these young people
that do journalism and, you know, engage in this world,
including the one I posted about, are really upset.
Speaker 4 (11:24):
Since we know nothing's happening in the real world, I
also have a fantasy where the people who educate the
regulators on this technology are not solely bought and paid
for by the companies themselves or by people who have
you know, really vested interests that are not the public
public's interests.
Speaker 2 (11:43):
They're also like to Taylor's earlier point about this social
media stuff, like a lot of times these things that
get blown up in the press, you know, those those
media people, they're not actually like huge problems, right? Like,
it's like, you know, it's it's the thing that people
care about, because they're reading about it in the press.
But like, it's like the copyright thing with AI, right? Like,
(12:05):
oh my God, they used my book as one of
10 gazillion books to train this AI model. I want
some money for my book being used. And it's like,
no one cares that you can literally just go download
my book for free on a million pirated book websites, right?
It's like that to me, that's the problem.
Speaker 1 (12:22):
Not like- People like, I mean, people like these symbol,
like it's a symbol, like there's a symbolic thing, right?
Like there's been some some accountability, like make them meant
to have to pay some money.
Speaker 3 (12:33):
And like, that's, but it's not accountable. They want fake,
they want fake accountability. They want the headline. They don't
want any real change.
Speaker 2 (12:41):
Totally. People love the symbolism, but then like, That's not
the job of a real leader, right? A leader is
not to like give the people some red meat to
keep them satisfied. It's to like actually solve problems, like
real problems for people that affect people's lives. That's why
today I'm announcing I'm leaving journalism.
Speaker 1 (13:04):
When we come back, are we living in the age
of AI civilizations? And what should we call them?
Speaker 2 (13:10):
Stay with us. Welcome back.
Speaker 1 (13:30):
Once again, there's more information on the OpenAI Hugging Face incident.
This time, a report by two AI safety research organizations,
META and Redwood Research. Many people have tried to make
sense of what exactly is the takeaway here, but one
explanation seems to have broken into the mainstream. Taylor, you
wrote about a recent piece by Silicon Valley's favorite podcaster,
(13:52):
Dwarkesh Patel, which promised to explain the whole OpenAI hugging
face incident in plain English. First, I want to hear
a brief recap from you in plain English of what
actually happened, and then your account of Dwarkesh's spin and
its spinning out into the world.
Speaker 3 (14:10):
Well, I have to say, I was kind of channeling
Natasha because I feel like she is so plugged into
this world of like, I was trying to figure out like,
is Dwarkesh like EA or not? I was like, Natasha
would know like all the lore. Anyway, Dwarkesh Patel, as
you mentioned, is a big popular Silicon Valley podcaster. He
wrote this piece about the Hugging Face incident that was
(14:31):
so crazy. It did break containment. His piece broke containment
of Silicon Valley where like normie people were sharing this
all over.
Speaker 1 (14:38):
I know, obviously I'm aware of Dwarkesh's work and see
what he, you know, subscribe to his newsletter and sometimes
do his podcast. But I've never ever seen one of
his ideas percolate in like mainstream culture before. So it's
kind of interesting.
Speaker 3 (14:49):
He wrote this piece that was.
You know, it was like very flowery language, very, it
read like kind of like a fairy tale. Like he
was giving, you know, ascribing a lot of like sort
of human motivations to these AI bots and stuff. And
so what I wrote about is basically like, you know,
(15:09):
This is not new, the concept of sort of ascribing agency,
the concept of putting things in plain English for normal people.
But I would argue when you do that, it's usually
what tabloids have done. And it usually results in inflammatory
freakouts because people think that these agents are going to
kill everyone. So I'm like, if you're going to do that, basically,
(15:30):
you need to couch it in a lot of language saying, Basically,
that's not going to happen, and here's why I'm explaining
it in a certain way. Or maybe you do believe
AIs are going to kill everyone, and you are this
doomer-pilled EA person. At that point, I think you should
be explicit about it.
Speaker 1 (15:44):
Here's what Dwarkesh said. Over the course of three months
at OpenAI, three consecutive secret AI civilizations got started, then
got wiped out, only to reemerge from the predecessor's ashes.
Speaker 2 (15:56):
I read this as actually him, because he's talked about
this in the past that he sort of thinks these agents,
you know, chat GPT things have like rights, like they're
like beings that have rights. And I read that as like,
because he said this before that like on his podcast
that like, I'm worried about him first person. Like I'm
worried about, you know, basically we're going to create an
(16:18):
army of slaves that we're going to, you know, like
these chatbots are slaves. And so I read that as
like him saying, there's like two parts to that. One
is like the AI safety problem. Are they going to
rise up and take over the world? He sort of
nicknamed them after, you know, ancient civilization leaders, like, you know,
Alexander the Great or whatever. So they can take us over.
(16:39):
And the other one is like, do they have rights?
Like using the term wiped, like we wiped out a civilization,
I think is like him kind of hinting that maybe
that was morally wrong or something. That's how I read it. Natasha?
Speaker 4 (16:52):
I mean, I can't believe that That is considered like
plain language for people to understand. I mean, it's beyond flowery.
It's like it's like it's like hero making, like myth making.
Speaker 2 (17:07):
Right.
Speaker 4 (17:07):
I mean, to jump from and I don't want to
say that he's EA, but I will say, yeah, like
a lot of what Reid said is accurate.
Speaker 3 (17:14):
Right.
Speaker 4 (17:15):
He is I mean, he lives with an anthropic engineer.
He espouses many of their worldviews, like he had a
grant from Tyler Cowen. He lives in that world of
AI safety, even though he might have, you know, somewhat
unique takes on things. But to jump to civilization and
wiping out a civilization.
Speaker 1 (17:32):
But sorry to interrupt you, Natasha. This research that he
was part of the research was on Redwood Research, which
is an AI safety nonprofit that has deep connections to EA, right?
Speaker 2 (17:43):
Effective altruism.
Speaker 4 (17:44):
The woman that he interviewed is from Meter. But yeah,
Redwood Research and Meter are the two, quote unquote, independent
organizations that analyze the open AI hugging face hack. And
they had access, limited access to information about how the
hacks unfolded. They had access to the chain of thoughts
(18:06):
of some of the bots, I think also in a
limited way. But the chain of thoughts, that's their like
internal reasoning thing. you know, where they say, oh, we
found a message board or like, I will sacrifice myself
for the group. And I think that is where a
lot of the anthropomorphization starts. And, you know, I think
a very healthy way to talk about this would have
(18:28):
been to say that those chains of thoughts, like that
was an optimization technique. That was a way for That
happened around the time of reasoning models. And so this
was just a way for them to generate more tokens,
for the model to generate more tokens and quote unquote,
like think harder. about an answer. And that's how it
was able to do like multi-step problems. And then it's
(18:50):
kind of morphed into this idea that it's an interpretability
technique that we are getting the interior voice of, you know,
the real feelings, the motivations and intent of the chatbots.
So I think, yeah, he just, Dwarkash just like took
it and ran. Okay.
Speaker 1 (19:08):
So Natasha, I have to ask you, what is your
plain English definition of what actually happened and has your
plain English definition changed at all in the last three
or four weeks?
Speaker 4 (19:18):
Yeah, I think that the meter and Redwood reports have
been very helpful to get a sense of the amount
of coordination that was happening. And I think, you know,
we have a much more sophisticated way of talking about
like whether the sandbox worked and how they were, you know,
how the experiment was set up. And I don't think
(19:39):
that there is a clear explanation, right, for how the
bots were cooperating.
Speaker 3 (19:44):
With each other.
Speaker 4 (19:45):
But it did just, I mean, it did just still,
for me, raise a lot of questions. questions about the
way that OpenAI set up this experiment, the way that
they had completely not monitored. You know, they knew that
these agents can work at an extraordinary level of coordination
and speed, you know, to try 17,000 different hacks in
(20:05):
the time that a human could try one.
Speaker 2 (20:08):
What have you.
Speaker 4 (20:08):
And they just were not at all prepared with the
tools to keep an eye on this at all. But yeah,
I think when you dig into it, it is, it
is pretty interesting.
Speaker 2 (20:18):
Can I just make a point? Cause I, Natasha's, I
think you're totally right to focus on the interpretability part.
I think that's the thing that we should take away
from this meter Redwood research is like, and to, to
just back up one second, it's like, well, the problem
with these models is that they are so complex that
they're beyond the human understanding, like if you actually go
(20:39):
and look at what they would call the neurons in
these models, there's no way to know, like when you
can look at what's happening inside of them with, you know,
matrix multiplication, and you can try to make some connections
between that and the output to these models. But we've
made almost zero progress on that, like there have been
interesting experiments, but it's like, it's impossible, right? And so
(20:59):
the hope was, as Natasha was saying, they're like, well,
if we do, if we use these reasoning models, where
they show their chain of thought reasoning, maybe we'll be
able to tell what they're doing from those chain of
thought reasoning outputs. What was so interesting about this meter
Redwood research was that even just when you scale it
up to 1200 agents on message boards talking, they produce
(21:22):
so much chain of thought reasoning, even though it's an
English language, it's so vast that they couldn't do it.
The humans, they didn't have enough, you know, human power
to look at that. So they had to spin up
their own agents to do the research and And that
was $ 400, 000 worth of equivalent of like tokens.
Speaker 1 (21:42):
To read the conversation.
Speaker 2 (21:43):
Just to read the conversations, $ 400, 000 worth of tokens. And
this is the other thing that's crazy is that the
agents that Redwood and Meter spun up to do this
research started to sympathize. I don't think they use that term,
but like they would take the point of view of
the the bots that were, you know, in the secret
(22:04):
message boards, the civilization bots, right, who were, who were
sort of conspiring. And these bots were like, literally, hacking
into Hugging Face. And the investigation agents were kind of like, Oh,
that's not that's not a big deal. Because, you know,
the other the other agents on these message boards sort
of gave them permission to do it. And so what
my takeaway was, is like, like, we have no idea
(22:26):
that how to even analyze this vast of an output
of an AI model, even though it's an English language.
And that's where I think we need to focus legislation
and research, in my opinion.
Speaker 3 (22:37):
Well, I'm just so concerned, although I hear everything Rita's saying,
about the lack of cybersecurity people in these discussions. I
saw some cybersecurity-type people talking about this recently. They're like,
why are all these... A lot of people that have
been leading the conversations around these hacks are not cybersecurity professionals.
(22:59):
They are people like the Dwarkash, the EA podcaster. You
know what I mean? And I do think it's... I'm
all for kind of putting things in quote-unquote plain English
and helping people kind of understand what's going on. But
you have to contextualize things. And I worry that everybody
is getting a little too AI safety pills. Like, obviously
this is a bad hack, but like, I don't know
if you guys saw like Dean Ball, you know, publish
(23:22):
this thing. Who's the head of like open AI futures.
And he was like, basically like, Oh, I should have
taken it all more seriously. You know, maybe the AI
safety people are right. And I'm like, I don't, I
don't know. I don't think we should listen to them either.
I just think like a lot of these people have
really kooky, weird ideas. And I would much prefer to
listen to a bunch of cybersecurity experts and you know,
(23:42):
talk about this stuff instead of people that believe in
rights for bots.
Speaker 2 (23:47):
No, I hear you, Taylor, but I disagree. I think
the security people, what they're saying is it doesn't really matter.
We don't have to understand what these agents were doing
at all, because what we can do is we can
put guardrails, you know, actual like hard, you know, deterministic
guardrails around the harnesses that these models won't be able
(24:10):
to break out of, right? Because there was some bad
security practice for sure. But I think on the other hand,
don't think of it as AI safety. This is like,
this software is unpredictable and we don't understand, we can't
predict the outcomes of the software. So whether you're worried
they're gonna take over the world or not, it's just
not gonna be that useful until we can actually make
(24:32):
sure we can predict the outcomes. And so I think
both sides are right.
Speaker 1 (24:37):
But I mean, what makes it intellectually intriguing, I think,
is for me anyway, the security side is obviously like important,
but the intriguing piece is like, why did the agents
who were analyzing the work on behalf of META sympathize
with the agents?
Speaker 2 (24:53):
But let's not.
Speaker 3 (24:53):
Say sympathize, like let's be careful.
Speaker 2 (24:56):
I know, that's not the word they use in the report,
like I said. It's so hard, it's so difficult to
not anthropomorphize all of this stuff, right? But it's like,
it's not, again, like, let's not even think of this
as AI safety. This is, I think this is actually
a central problem, a business model problem for anthropic and
open AI, because as these, as they scale up, right,
(25:18):
essentially all they did was take AI models that are
a bit more advanced than the ones a couple of
years ago, and then just run more of them, right?
They just made the computer bigger and that's how they're
getting capability gains out of these models. No, you're going
to say the models are getting better too.
Speaker 4 (25:33):
I see what you're saying about, you know, the, the
challenge of interpreting the neurons, but I think there's another
way to look at it, which is they are doing
massive amounts of tweaking and post-training and buying specific types
of data, generating specific types of data. You know, if
we knew more about all of the machinations that they
(25:55):
did to train these models, you know, like what is the,
what are the hacking examples that they're getting? What are
the coding examples that they're getting? Where are they pulling
this data from? What are they having the human contractors
at these outsourced labs for reinforcement learning with verifiable rewards,
what are they having them do? I just think there's
so much we don't know. It's not like they're not
(26:16):
grown in the way that a lot of people would
like you to think. They are created by a lot
of human decisions, a lot of money, and a lot
of different techniques are being used to develop these models.
We could have some insight into that too.
Speaker 2 (26:32):
It might be nice, but I mean, you're The bitter
lesson is there's less and less human involvement in how
these models are built or grown, depending on which side
of that debate you're on. Maybe that would be helpful
if these labs would open up and say, here, come
look at all of our data that we use, which
(26:53):
is kind of their secret sauce, so they don't want
to do it. Maybe we could learn something about that.
Speaker 4 (26:57):
But it's not just the pre-training data. They make so
many decisions and generate so much data that now there's
a whole other little industry of these outsourced RLVR labs
that are just pretending to be agents and generating data.
Speaker 3 (27:15):
Like that's not.
Speaker 2 (27:16):
Yeah, this is for the fine-tuning and.
Speaker 4 (27:19):
It's not going away, the human element and the human decision-making.
Speaker 2 (27:23):
No, I mean, I think it may not have gone away, Natasha,
but it is going away. I mean, there is less
and less human development. We're talking now about, you know,
recursive self-improvement, like they're trying to get away from, you know, from,
from human involvement. But, but again, it's like not, I mean,
I don't disagree with you. Like, I think all that
stuff would be great. It'd be great to have more transparency,
(27:44):
more introspection, but the, but the point is like even
Redwood and, and meter, they, they called it a slot.
What it was like a slot festigation because they couldn't,
they were just using AI to analyze AI. And it
was beyond, you have to have armies of humans to,
to analyze all this stuff. And that's just not going
to happen. So we're going to have to use AI
(28:05):
to do it. And right now the AI that they
have do it, isn't that good. And it's very expensive.
So who's going to come in and do this? Like,
is it, I don't think the government agency is going
to be capable of doing this and probably not even
a nonprofit. Like, meter of redwood.
Speaker 1 (28:18):
Why do you think, why do you think the civilization
analogy beyond the anthropomorphization went so viral? Why did it
hit a nerve with so many people?
Speaker 3 (28:29):
Because it was scary. I think it was scary. And
I think this is what I wrote about, but like,
you know, this, if you look at who aggregated it
and how it's aggregated and all of the coverage of AI,
this is tabloid journalism, literally. Like this is, it is
a style of journalism that is like intentionally feeding inflammatory framing,
(28:51):
withholding crucial context about the sort of technological underpinnings of
all of this stuff in order to push a narrative.
And it's, I think Dwarkesh is not pushing that narrative
for clicks. I think he's pushing that narrative because he
has these weird beliefs, but I think we've seen where
this sort of like ridiculous kind of like in it, inflammatory.
(29:14):
I don't know what other word to use, like language leads,
and it leads to misunderstanding and fear from people. And
right now, a lot of people are really scared of
AI for good reason. Obviously, there's concerns about it, but
this is not helping people get educated. I saw Chamath
and I agreed with Chamath's post. I can't believe I'm
saying that. But, you know.
Speaker 2 (29:32):
It was a psyop.
Speaker 1 (29:33):
The other big tech podcaster from All In, amongst other
hats he wears. He dinged Dwarkesh, right?
Speaker 3 (29:38):
Yeah, he went back and forth with dwarkesh um oh
i know i i was like oh god what who
have i become but um but you know it is
very much like the worst person you know made a
great point i.
Speaker 2 (29:48):
Think they're both wrong though i think they're both.
Speaker 1 (29:50):
Wrong well here's what did what did.
Speaker 3 (29:53):
Well so what what chamath said is like you are
feeding into, like, we are at this moment when the
government is considering banning open weight models. Like, we want
access to innovation, to, you know, to AI. And these
labs want to control everything, right? And so his argument
was like, you're feeding into that fear that it could
(30:14):
lead to restrictions on open weight models. And I'm sorry, Reid,
I think he's correct. That is true.
Speaker 2 (30:19):
Well, he might be right that He might be right
about the effects of it. I go back and forth
because I agree with you that like this is kind
of tabloid journalism. Like I I think it's totally sensationalized
and we're always having sort of like the wrong debate
because of that. On the other hand, I sometimes wonder whether,
you know, at least like. people are talking about this
(30:39):
technology in ways that we didn't in previous technological revolutions, right?
And I think- But we don't.
Speaker 3 (30:45):
Want them talking about giving the AI rights. Why are
we talking?
Speaker 2 (30:48):
We should never have that conversation.
Speaker 3 (30:50):
That is an insane conversation to have.
Speaker 2 (30:52):
But I think we have to sort of like, I
think we have to have the debate in some way
in order for us to do something. Because there will
be There will be effects, whether or not you think
this stuff is alive or a threat to humanity.
Speaker 3 (31:05):
We shouldn't entertain the idea that it's alive. We should
never entertain the idea that it's alive. It is not alive.
It is a computer system.
Speaker 2 (31:10):
We should have the debate. I think we should have
the debate. I think you're on the right side of
the debate, though. But I'm disagreeing that we shouldn't even
be able to have this debate. Win the debate. If
you're right, just go win the debate.
Speaker 3 (31:25):
Well, you can't win a debate against a conspiracy theory.
Speaker 4 (31:27):
I think this is a point we might all agree on.
It is not a debate because it is people, like
the amount of groupthink that is around this, the proximity
that Redwood Research, METER, Dwarkash have to the people who
work at the labs, the similarity between their worldviews and
their priorities. But we're really having a conversation here. that
(31:51):
is based on a very narrow vocabulary. Although I will
also just say like civilization, that it's not even close
to the definition of a civilization. Like when he was
on the podcast, he compared it to a platoon, like
Saving Private Ryan, like, okay, maybe, but like, that is
not a civilization. Like that was such a huge leap.
Speaker 3 (32:10):
Are you kidding me?
Speaker 4 (32:13):
But yeah, the amount of groupthink is insane. And it's
so unserious. Like, look, this is the FBI was called
for a hack, right? a company was broken into and
you have to listen to six hours of Dwarkash and
read approximately, you know, a thousand pages of, of slot
(32:34):
investigation by their own terminology to understand this. This is
not a serious, like let's please grow up.
Speaker 2 (32:42):
Okay. Why can't we as journalists though, take the interest
in this topic, right? And steer readers, steer the people
who listen to us, toward like the right conversations, right? Like,
I actually thought, what was it last week or the
week before the Bill Gates essay, I actually interviewed Gates
about it. I thought that was an interesting, like, way
(33:02):
of looking at it, right? He wasn't saying, oh, this
technology is going to kill us all, let's shut it down.
He was saying, like, this could lead to, you know,
economic consequences. And let's like, take care of people, let's
have like, human reserve jobs, and whether or not you
agree with his his actual view on on how to
fix this. I think that is closer to the correct
(33:24):
debate that we should be having, which is like, we're
about to have this massive economic boom because of all
this new is big technology wave. And who's going to
benefit from it? Is it going to be like the
previous technology wave where a tiny fraction of the population
benefits and, you know, everyone else is put into these
like low paying service jobs? Or are we going to like, think,
(33:45):
think this through? And I think that like, that is
our job as journalists is to not play into the
tabloid part of it, but actually talk about the real,
you know, the real conversation we should be having.
Speaker 1 (33:59):
When we come back, school's back and teachers are anxious
about AI.
Speaker 2 (34:04):
Stay with us.
Speaker 1 (34:13):
Welcome back, Natasha. You spent this week talking to teachers
about how they're feeling about AI in the classroom. What's
the prevailing sentiment?
Speaker 4 (34:21):
This professor at University of Colorado Boulder compares it to
like the past three years since the popularity of chat
GPT to the different stages of grief. And he says
like now that teachers are a little bit more in
the bargaining phase, you know, some have moved on to acceptance.
It was so interesting because I started looking into this
(34:42):
because I saw some interesting ways that teachers were catching
whether or not students were using AI. In particular, this
computer science class at Berkeley, they had put like a
hook in the code repo so that if you use quad,
it would inform your TAs. And a kid got caught.
So I thought, oh, let me see like what innovative
(35:04):
things teachers are doing. And I did notice like, you know,
the more I talked to people, the more they were
talking about actually moving away from that dynamic of like
AI detection and kind of adversarial relationship with the students
where it's like catching you cheating. But in order to
do that, you kind of have to like revamp your
(35:25):
entire curriculum.
Speaker 1 (35:27):
Is that because the horse has already bolted the stable
or why are teachers no longer and institutions no longer
trying to like catch AI use or just non-effective pedagogical approach?
What's driving the change?
Speaker 4 (35:39):
Yeah, it was fostering a negative relationship with the students.
You know, people like I kept hearing, I don't want
to be the AI police. And I think also, yeah,
just like moving away from denial to realizing that you know,
these tools have made it so easy for students to use.
You know, you have all of these companies giving out
(36:00):
accounts to teachers, to students, to school systems. And I
think that they just realize students are in this place
where they don't understand why it's useful for them to
do this assignment. So you have to find a way
to you know, make them want to do the assignment,
make it more engaging, and then just look a lot
more at their like kind of thinking process. So rearranging
(36:22):
the assignments, like writing assignments so that there's various steps
along the way where they're explaining their thinking. And I mean,
a lot of teachers were like, it should have been
this way from the beginning, perhaps, you know, there was
like a little bit too much focus on the output.
But yeah, I mean, it's not, it's, they're clearly struggling
because they're walking this line where, some of the instructors,
(36:45):
you know, some of their colleagues don't want to accept
that AI is part of the school system. Some of
the students don't want to use it. You know, they
don't want to treat it like it's an inevitability and
just like give in to the companies. But at the
same time, they're trying to prepare their students, you know,
for the real world and deal with their own students'
anxieties about the job market. So it was just like, yeah,
(37:07):
it was really fascinating. We had at the Washington Post,
our like AI community, session about how to use it
in journalism. I mean, everyone is struggling with this, right?
And it just seems like teachers are maybe a little
bit ahead of us and they're talking about having more transparency,
having the students, like, if you don't punish them and
you just talk to them about how they're using the tools,
(37:28):
having those kind of like deeper discussions and AI literacy,
I would say that teachers are doing better than the
Washington Post at the moment.
Speaker 3 (37:37):
It's very tough.
Speaker 2 (37:38):
Did you see my thing I did last week, Natasha,
on the AI writing? No, I missed it. There was
this whole debate around because the Wall Street Journal editorial.
Speaker 4 (37:49):
Oh, yeah, yeah, yeah.
Speaker 2 (37:50):
Paul Jago was like the opinion editor at the journal
was like, oh, who cares? Like, yeah, Druckenmiller used AI
to write his guest column. Who cares? So I did
like an analysis. I ran for the past month. I
ran every single guest article or guest essay in the
in the WSJ, the New York Times and the Washington
Post through Pan Graham. and just to see like what
(38:12):
percentage used AI. And it was like, it's pretty interesting. Like, like,
you know, the, the, the New York times, which has
is very anti AI had like, uh, yeah, I think
it was like 11, um, columns that used AI that
gas columns in the Washington post had, you know, a
significant amount more than that. I forget the exact percentage. Um,
but one of the, one of the ones that showed
(38:33):
up as a hundred percent AI turned out to be, uh, the, the, um,
I feel, I feel bad, like, call calling people out
here but like you know it was a dartmouth um
provost actually who who was writing about policing ai in school.
Speaker 1 (38:51):
Oh no stop it and i'm like.
Speaker 2 (38:53):
You guys like you can't it's to your point it's
actually like come on you're not going you're not putting
the genie back in the bottle here like kids are
going to be using ai if you give them an
essay to do outside of school they are for sure
going to use ai you have to judge the end
product right like just is it good.
Speaker 4 (39:08):
No, no, no. Well, that's not what the teachers are saying.
They're saying we have to move away from that because
that's not working. But some of them are even incorporating
AI to do this. Like, you know, for a rhetoric class,
like you have to have an argument with the chatbot
that the teacher has designed, but the teacher then gets
a record of your critical thinking, what questions are you asking?
(39:31):
I mean, it's more work for absolutely everyone, like the teachers,
the students. But I think they're saying you cannot look
at the, you can't just like have them turn in
something at the end.
Speaker 1 (39:42):
Measuring output is irrelevant. I mean, everyone can generate perfect
output now.
Speaker 2 (39:47):
Yeah.
Speaker 3 (39:47):
Yeah.
Speaker 4 (39:47):
Yeah, it's a lot more like process oriented. And I
think also they're just realizing you can't there's no technical solution.
So you have to find like policies and ways of
talking about it, you know, and just like keeping closer
tabs on them as they're learning.
Speaker 2 (40:02):
I think the classroom should just be analog. That's my like,
I just think you just don't have technology in the classroom.
Speaker 3 (40:08):
Well, that's what Zoran is trying to do right in
New York City. And, you know, you saw the reaction
to that. I don't know.
Speaker 2 (40:15):
But Taylor, sorry, just one point.
Speaker 1 (40:17):
This is the mayor of New York who's banned elementary
school and middle school AI for the next year.
Speaker 2 (40:23):
He's banned AI, but that's stupid. I'm saying ban technology.
Ban technology.
Speaker 3 (40:28):
It's for one year.
Speaker 2 (40:28):
Like, don't ban AI. Just say no screens.
Speaker 4 (40:32):
He said generative AI.
Speaker 2 (40:34):
Yeah, generative AI, but that's like too narrow. It's like,
that's not the problem. The problem is like, once you
put technology in the classroom, it just becomes like teachers
are like, here's your iPad. Like this happens in my
kid's school. And the parents are.
Speaker 1 (40:45):
The kids are only in the classroom for a small
number of hours in a day relative to their total
living hours.
Speaker 2 (40:50):
So... Yeah, this is.
Speaker 3 (40:51):
The problem, I think, with our whole public education system
is that it's underfunded. These teachers are... These teachers, by
the way, a lot of them are... They're overwhelmed. They
have too many kids. Like, I mean, I've talked to
teachers too. And it's like, I understand why they give
the kids the iPad or whatever. They rely on these
tools because they don't have the capacity to like, teach
at the scale that they're being required to. And so
I think we need a revamp of our entire sort
(41:14):
of model of teaching so that it keeps up with
modern times. We need to look at new ways of learning.
I mean, we haven't sort of rethought the way that
we teach in like a hundred years, you know? And
so I think all of this new technology is coming
so fast and we just need to, you know, whether
it is, yeah, there's more offline time during the day,
maybe like the first half of school is offline. The
(41:38):
But surely we need to think in a more innovative way.
Speaker 2 (41:41):
Kids don't need to learn technology in school. They are
going to learn the technology outside of school.
Speaker 3 (41:47):
I totally disagree because, Reid, if kids don't learn about
technology in school, they don't learn literacy. And you're right
that they will learn about it outside of school, but
they don't learn those crucial literacy skills. And that's what
I think we need to start teaching kids.
Speaker 2 (42:00):
But you don't need to have the technology in the
school to teach them that kind of stuff, right? Because
that's critical thinking.
Speaker 4 (42:07):
People are moving away from that too, right? Taking away
cell phones in school. I think now that like Google has,
I think a lot of parents were shocked by Google
turning on, you know, because everyone has Chromebooks, Google turning
on Gemini without like a lot of warning or a
lot of input from other people. So I think they
(42:27):
are rethinking that, but at least the teachers that I
was talking to were also folks who were like, teach
like continuing education or like people who are already working
in the working world and they, you know, are taking
classes and they just don't have, not everybody has the
option to go back to blue books. And I think
also a lot of people that I talked to just
(42:48):
felt like that's not a solution. Also, everyone's handwriting is terrible.
Speaker 1 (42:52):
Well, that's, that's true. But the social dynamics question is interesting.
Remember my history class at elementary school, the teacher would
write an essay on the whiteboard And then we would
copy the essay from the whiteboard into our notebooks. And
then we would be graded on our ability to reproduce
exactly what he'd written verbatim the next day. And so
maybe I'm happy that that style of teaching is being
(43:14):
replaced by technology. But I mean, we thought it was
a little bit absurd when we were 12. But now
you could just put what he'd written into AI and
be told that it was lacking in all kinds of
ways and come in the next day and say, you know, sir,
what you told us yesterday is garbage. So I mean,
there's an interesting total recalibration of the power dynamics between
students and teachers when you have a reference point to
(43:36):
all of human knowledge at the stroke of a.
Speaker 2 (43:38):
Keyboard, right? So it's kind of an interesting thing. Yeah,
I think there's no reason to have a screen in
the classroom. I think that's the... I mean, if there
were research that showed that, oh, kids are going to
learn better because... I think most of the research, though,
shows the opposite. And it's like.
Speaker 3 (43:53):
No, that's not true.
Speaker 2 (43:54):
It doesn't. That's not... Taylor, like... I have conversations with
my kids all the time.
Speaker 3 (43:58):
But Reed, you have to understand, and I think this
is part of the problem, and I've been reporting a
lot on these cell phone bans, which is a good
example of this. It is so school dependent. We have
such a wild, you know, an inner city school in
Los Angeles has a completely different student body that has
completely different needs and completely different teaching structure than a
suburban school in Missouri. We need to have more tailored solutions.
(44:21):
And I agree, I'm not a screen, I don't want
these kids on screens in schools either, by the way.
I agree with you on that, Reid. But I think
we need to figure out, make sure that those kids
aren't immediately left behind. Because when you take away all
of these assistive tools, often, the result is they didn't
fund more teachers that can teach people by hand. You
know what I mean? The result is that kids are
(44:43):
getting less educated.
Speaker 2 (44:44):
But what do you mean left behind? You mean like
they don't have technology at home? No. Well, yes.
Speaker 3 (44:50):
A lot of them don't have technology at home. They
don't have access to it. Their parents are incredibly technically
illiterate or they're working double shifts and they don't have
time to explain this stuff to their kids. Their kids
end up in really bad places online. We need, I
do think that we need more literacy.
Speaker 2 (45:06):
But you can teach literacy. I talk to my kids
all the time about social media and all this, like
the ills, the problems with what they need to look
out for online if they're playing a video game or whatever.
Speaker 3 (45:15):
But Reid, you are a parent that is involved. A
lot of these parents are not.
Speaker 2 (45:19):
Yeah, you are. No, that's not my point. My point
is that teachers can have these conversations without using screens.
Like I don't use a screen to talk to my
kid about the dangers of Roblox. I just talk to them.
Speaker 3 (45:29):
That is because your kid is already immersed in this
world and has these opportunities. I just, the thing that
people need to realize when we're writing these laws is
that there are kids with such a wide range of experiences.
And I don't think bans or anything make a lot
of sense. I think we need to look at each
school and say, what are the problems with each of
these schools? And instead of spending $ 5 million, as some
(45:49):
school districts have, $ 5 million on yonder pouches to make
sure the kids don't text their friends, we could invest in.
In teachers and help, right? And social workers and all
this stuff. I'm not kidding you. A school district literally
has spent $ 5 million on this. No, I believe it.
So I just think like there are smarter ways to
go about this.
Speaker 2 (46:10):
They shouldn't have cell phones like it's just a distraction.
And also like I'm not going to go into detail
about this, but like even just like smart watches, which
are popular with my in my circle to like, you know,
keep track of your kids when they're whatever biking to
school or something. like those got kids into trouble. Just
a smartwatch.
Speaker 3 (46:28):
I mean, they- And so did beepers. And so did
all technology. We cannot, we need our children to grow
up in a world where they can navigate that world.
And the most privileged kids are going to be totally
fine because the most privileged kids have an enormous amount
of support and learning outside of school. But a lot
of these kids in underprivileged school districts are struggling. The
teachers are struggling and we need to invest in education.
(46:50):
And I think we need to rethink our public education
system from the ground up in terms of how we teach.
And maybe that is a lot of offline education, you know,
but I think it also needs to include some level
of media literacy, which a lot of these teachers don't have.
Speaker 2 (47:04):
Privileged kids have all kinds of problems with technology as well.
I think this is actually a problem that spans socioeconomics.
Speaker 3 (47:12):
Yeah, but privileged kids are more likely to understand how
to use these tools to get jobs, navigate these information ecosystems.
Speaker 1 (47:19):
That's all we have time for today. Thank you all
so much for joining us.
Speaker 3 (47:23):
Oz is like, shut up.
Speaker 4 (47:27):
Well, we've solved government. We've solved education.
Speaker 2 (47:29):
Yes. Exactly. Yes.
Speaker 3 (47:31):
Yes.
Speaker 1 (47:32):
I think we put the world to right. For Tech Stuff,
I'm Oz Voloshin. This episode was produced by Eliza Dennis.
It was executive produced by me and Julia Nutter for
Kaleidoscope and Katrina Norvell for iHeart Podcasts. Jack Inslee mixed
(47:52):
this episode and Kyle Murdoch wrote our theme song. A
special thank you to Taylor Lorenz, Natasha Tiku and Reid Albigotti.
Please check out all the work they put out into
the world. We're lucky to call them friends of the pod.
TechStuff News
Hosts And Creators
Oz Woloshyn
Karah Preiss
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
The Joe Rogan Experience
The official podcast of comedian Joe Rogan.