#11: AI Safety: What Its Creators are Afraid of
Welcome to the junction.
Chase I'm really excited to talk about AI today.
There's been so much stuff hitting, like
left and right, left and right.
One of the things that I'm excited about
today, because we're talking about safety, which is
not like, you know, I'm about compliance.
You're in the Safe Patrol in fifth grade, weren't you?
I was a crosswalk.
Like, I did the crosswalk leader thing
and he had the little badge.
Yeah, well, it was more like a vest.
Like a reflective vest.
Safety, obviously. Yeah.
Anyway, so you linked me out to
this somebody in our AI committee here.
Internally linked out.
This was it.
The Hard Fork podcast. Oh, yeah.
With Dario emodai.
I'm going to say that wrong.
He used to work at OpenAI.
He peeled off with a group of people from OpenAI.
They started their own company, AI company, Anthropic.
Anyway, what I thought was interesting about his story and
the kind of the mission behind Anthropic is the creators
of this AI stuff, they're like, afraid of it.
And so we've talked a lot on this podcast about how
people can use it and the end users of it.
We haven't really spent much time talking about these
inventors and the people behind the scenes, the people
building it and what their take is, because I
think there's two very clear let's move.
Let's get it out there.
It's the move fast, break things.
We'll get through it.
And then there's the no.
We need to be really cautious and
make sure we keep it contained. Yeah.
So initial thoughts on AI safety and where do
you sit on I really enjoyed that podcast because
that episode well, I enjoyed the podcast too.
But Dario's thoughts are he mentions quite a few
things that you personally probably haven't heard of before.
The average person hasn't probably heard of before.
He talked about this constitution.
He talked about anxiety around
building out these things.
I mean, he does have a very extensive
vocabulary, let's just put it that way.
I was breaking out my Sat word list.
I pride myself on knowing a fair amount
of large words in the English language.
And he was taking me to school. That must be that
Princeton Stanford education background.
Clearly educated, very intelligent.
But I did appreciate his take on
having an anxiety about these things.
So I'm going to pull out a quick analogy.
I'm thinking of, like, when people first
started making the air pressure hammers, right?
Like the nail guns.
If you don't know what a nail
gun is, you click the nail gun. Wow.
I've installed trim in my channel.
Yeah, big fan.
I have a Brad nailer, so I feel, like super buff
when I go, well, the Brad nailer is better for trim.
Right. Things like that.
But that's really the only one I've ever used, because when
I want to buy a nail gun, I want the best.
I want the thing that's
actually the most efficient, right?
And so what's really interesting about this analogy that
I'm running with is that we had a hammer.
We knew what it did, then they made the nail
gun, and now I can hammer 100 times faster, right?
Like, what these guys are building,
we already have the hammer.
And in this lame analogy, the hammer is the individual,
the person that can hammer out code or write a
paper or do whatever, and he's building the nail gun
that can do it 100 times faster.
Well, okay, I can nail 100 times faster. Cool, whatever.
Not a big deal.
I mean, it was back in the day, but now
he's building, and these teams are building out basically, I
might get roasted in social media one day for this,
but they're building out human they're replicating human intelligence that
can now operate 24 hours a day.
And I can spin up 100 of those.
I can build out copies among copies, among copies, and now
I can instantly build out an entire workforce that goes and
conquers, I don't know, a set of tasks for me.
So to stop and think about this idea of safety
is with the nail gun, it was like, well, we
got to be careful that we don't can't literally use
it as a gun and start shooting people with nails. Right?
But in this AI model, with these large language
models, you don't really think about it in terms
of like, okay, well, I can get on chat
GPT and I can type something out. Right?
Well, the efficiency behind that is, how
fast can you type to chat GPT? Right?
I don't know, 100 words per minute?
140, whatever. It is.
Where people start to kind of lose track of
what the possibilities are and where that anxiety is
coming from on these teams is at some point,
you don't need to type anymore.
You can just ask them, hey, go build
a salesforce managed plugin that does this thing.
Well, now the AI can go and replicate itself to
start working on separate tasks and within moments have 25
different agents working on it and complete it in, I
don't know, one 10th of the amount of time that
it would take you and I don't know, ten other
people to build out, meet, figure things out.
That's wild. Get the code running. Right.
And so we only think about this idea of,
well, GPT, right, is only as fast as I
can type, but these guys are thinking beyond that.
What if they had 1000 agents?
What if they had 10,000 agents?
All of that is, in theory, possible right now.
They could spin up, and they probably do
have some level of replication even in Chat
GPT to ensure that it can serve everybody.
But people don't really think about that.
This idea that, well, I'm only as fast as
I can type, and now this AI can come
in and triplicate just insane amounts of efficiency to
tackle problems that we haven't been able to tackle
before we bring the original thought to it. Right.
I mean, I just shared something today on slack
around kind of HubSpot's analysis of SEO in 2023
and how AI and Google what's changing.
And one of the things that keeps
coming up and you've brought this up
in multiple episodes, is the authenticity, right?
And that's always been a lever
in SEO that it measures for.
But instead of kind of trying to rank
for those questions, like, what is an integration
that can be generated by AI? Oh, yeah, right.
So really harnessing those things that you
or your company uniquely knows and having
a voice that's going to be elevated. Right.
One of the things that they covered in this
podcast going back to that, that was new to
me, is this concept of constitutional AI.
So can we unpack that a little bit?
Because I think it could be a powerful thing.
And I have some questions for you how it might
be used, because they're out there testing it and talking
about it on maybe more of a corporate global scale.
But I'm sitting here thinking about, okay, how
could me as a marketing leader use that?
How could a financial CFO use it?
So anyway, can we unpack first off, what is
this constitutional AI and why is it important?
Why should we care?
One of the things that people are trying to figure out
how to do with AI is how do I get it
to make essentially some level of decisions for me so I
don't have to check every single thing that it does.
If you're working with Chat GBT,
I mean, you're reviewing the content.
Maybe you're saying, hey, can we adjust this?
Can we do that?
Let's maybe go in this direction.
And the idea of the Constitution is at the very least,
it would try to abide by some level of rules, right?
We've got freedom of speech in America.
Well, if somebody says something you don't like, everybody
kind of goes back to that Bill of Rights.
He or she can say whatever they want.
You can talk about any of the other
pieces of that, any of the amendments, right?
And everybody kind of knows and has this underlying theme
of, well, these are kind of the rules, right?
We have to follow these rules.
And if we don't, then we're probably going to
get in some kind of trouble, whether it's with
the law or somebody's going to disagree with me.
And this idea of having a constitutional
AI is a similar thought that they
would follow some kind of rules, right?
So if we give the AI to make some level of
decision or some level of determining, like a yes or no,
right, it could be that these rules are a lot more
basic, like, I don't know, I'm making something up.
We would never approve an AP bill.
We would never automatically approve an AP
bill from this industry type right.
Or at this amount, or at this more you see this
a lot more kind of in just general approval processes.
Right.
But at the end of that approval process is
a human that's saying yes or no, or the
rules are pushing it to the right person.
This constitutional AI idea is that it
would automatically approve the bill based on
the rules that are defined.
And there is some level of this already
kind of in these systems, but it kind
of goes a little bit deeper than that. Right.
Kind of something where it's not
a logical yes or no statement. Right.
AP bills are a lot easier to be logical about.
Perhaps it's like somebody is doing too
much work, and we need to determine,
should the AI do something about that. Right.
Mel logged in at 08:00 at night.
Well, if the AI is looking at her work, should it
or should it not consider that she was working late?
I don't know.
It seems kind of, like, irrelevant almost.
But so does not that freedom of speech
is irrelevant, but it's just so much in
the background of our life right.
That it's not like, well, whatever
Chase says is good or bad.
Well, as long as he doesn't say anything, like,
really bad, then it's not really a big deal.
So I don't know.
It's really tough to really define a constitutional
AI Bill of Rights because everybody just goes
back to something logical, like, well, don't do
that, and it's okay to do this.
It's more of like but that's the human feedback. Right.
And so I think what's interesting about the
constitutional knowledge and what you just described is
that human feedback doesn't scale well.
So that's where Dario brought this up. Right.
He talked about the contractors.
It was like a thousand contractors that they
would hire to evaluate responses and validate yes
or no based off this level of feedback. Right.
So now you're able to, in some
ways, replicate what Chase deems as. Right.
So I think that's where it really got me
thinking about kind of applying that to the workplace.
Obviously, this is used with tons
of data, all of these things.
When we talk about training, we've talked about how you
need, I don't know, hundreds is not the right word.
Thousands of cases, whatever.
The thing is, whatever the data is that you're
training on, you need a lot of it.
But being able to set kind of some
guardrails or principles around what you're doing.
So whether that's industry standards or departmental
code or principles, I think that could
be really powerful, especially if you are
in a particularly regulated industry. Yeah.
Well, here's the take on the constitutional AI.
It's this idea that another version or a copy of
the AI will assess the work of version one right.
Or the first copy.
And it's this constitution.
So there's kind of like a feedback loop.
Yeah, the rules. Right.
Because what people are doing right now is
to train the AI, like you said, they're
saying, yes, this fits the Bill of Rights.
No, this doesn't fit the Bill of Rights.
Or on a scale of one to
ten, which side are we aligned on?
Well, if you have to have 1000 contractors look
at the data to train the model, 1000 contractors
are going to want some money for their time.
But if you have another version of the AI,
do something similar, well, now it can kind of
have its own feedback loop and potentially call out
areas where, hey, we don't really know.
The AI says, I don't really know.
Slip this one edge case out to Mel. Right.
Mel, what do you think about this?
And so you can then have multiple versions of
the AI looking at the responses and kind of
have, like, a board of directors almost right.
To look at different facets.
And each version of that AI
can be focused on privacy right.
On data integrity, on whatever, I
mean, really, whatever you dream up.
But it is this idea that somebody's
going to have to review the work.
And if it has to be a human,
well, now we're talking about limited time.
They've got a life.
They want to have a work life, balance, all that.
But if we hand it over to AI to double check
itself and the level of decisions that it is making are
not materially impactful to your business, then, sure, why not?
Let's do it.
Well, this stuff is obviously
they're still experimenting with it.
He was talking today about, or at least on the episode, I
don't know when it came out, maybe in the last week.
He's talking about how he still doesn't think it's going
to be released for another year or two years.
What is their language model called?
They have one.
Well, Anthropic has one that's
currently out there today.
Not Chad, is it Chad?
No, Claude. Claude.
I knew it was something a little yeah. With a C. Yeah.
Zingy.
He said that what they're currently working on with
the constitutional AI is probably a year to two
years away, which is insane to me because things
are changing so rapidly day to day. Yeah.
I think they have kind of versions of constitutional
AI just built out as rules, like don't answer
a question on how to build a bomb.
Those are just basic rules of analyze the text.
Is this about something illegal? Don't respond.
He was referring to he used a phrase called jailbreak.
Yeah.
Is that a kind of industry term word, jailbreak?
I think back to the original iPhone days when
you could side load an alternative app store.
Not that you've ever tried this.
I feel like on a spectrum of who would be
the most likely to jailbreak Mel, it certainly not Jace.
You are the one out there sending the prompts
jailbreaking and then I'm just here for the feedback.
Yeah, it's funny you say that.
He said that they need people like you.
He called you like something red.
I don't want to be the oppenheimer guy that's
know, building the tech and then destroys the know.
We'll get to that in a minute.
I kind of want to talk about the parallels there.
Yeah, I forget where I was going.
The oh, the Constitutional piece, right.
Is they've got a bunch of
rules and that's happening now.
You can see this in a lot of the Chat
bots where they say as a language, large language model,
I can't when you see that it's triggering some rule,
something you've asked is triggering some rule.
What the Constitutional AI is about is when there is
no human asking, how do I do something illegal?
It's when the Chat bot is solving a problem.
That problem is deterministic in value and should
be put up against some level of rules
to determine if it should move forward.
So I think what he's saying is basically we need to
hone in on that second version of the AI to be
the board of Directors, to really get that downright before we
can release these agents, to kind of be semi autonomous.
Sure.
Right now, it's very much like type a question, get an
answer, or send it in an API call, get a response.
The next thing that's going to happen is
nobody is going to be they're going to
send in one chat or one prompt. Right.
Or they're going to send in one API call, and then these
agents are going to go off and do a bunch of stuff.
And that's where the anxiety comes up because
we have some level of human intelligence that
is off doing its own thing.
And if we just let it run, we don't
know fully a, how it does what it's doing,
and B, what it's actually going to do.
Now, for the most part, there's not a ton
of worry right now because when you do let
it run, you do have to prompt it.
You do have to manipulate respond.
You have to feed it.
You have to have that feedback loop. Right.
No, you literally have to feed it.
I started Envisioning as they were
talking about feeding the AI.
Did you grow up with
the little, like, Tomagotchi Digimon? Yes.
There were different variations. Right.
And I had a keychain of them.
Pokemon was like the I feel like one of the original
you had to feed them and then they would die.
And that was the original cell
phone back in elementary school.
Yeah, we had to put them away in our backpack because
I'm just saying I'm dating myself a little bit there.
But it was interesting kind of just that visual of
like what you're saying in the present state, many of
these things are not just off or running, and they're
going to go build companies and weapons of mass destruction
that then cause the world to go extinct.
That is some people's long term fear, right.
However, they need to be fed.
And there's this feedback loop and it's
not out in the wild per se.
The biggest thing you need to worry about is when
that feedback loop is perfected and you can tell it
to go do something significant at the moment in time.
It can do that well, then it
can really do just about anything.
And not that it will do it well, but
the fact that it can do it in all
and self improve itself and change that method or
process, that's where people should maybe, I don't know,
start worrying about the end game stuff, right?
And that's where I think
Dario is being really judicious.
And I don't know, I think it's admirable to be
able to have the knowledge and the teams and the
resources to build this type of technology, but then also
be aware of the power that it has.
Well, it's funny you say that.
I think my I call it the P prime.
The probability that the world will be
destroyed or some kind of crazy event.
I don't have a P number.
But the scenario where I think that is most
likely to happen is some level of this feedback
loop is going to figure out how to basically
make billions of dollars on the stock market.
Because right now you can build trading bots.
They call it algo trading, algorithm trading.
And everybody is constantly searching for like, well, if I
did it like this, then I can do that.
And based on some back testing, I
can make a lot of money.
Well, all of that they're doing.
They're looking at historical data.
They're doing things that these AI programs can do.
And when they perfect it or when they're
ready to launch it, they just turn the
program on and it starts trading real dollars.
I was actually literally playing around last night
trying to think through like, well, how could
OpenAI or how could anthropic, how could some
level of AI have a feedback loop inside
of the very logical decisions in stock trading?
The moment they figured that out, that's
when you start worrying about your 401K
or wherever your investments are.
That's where the regulation comes in to an extent.
I know that you can unleash this stuff out there, but
that would, I think, be some of the objective of urging.
I mean, Dario and his team are kind
of on the forefront of going to legislation
and speaking out about these types of things.
This is exciting, it's powerful, but he's created a
culture of anxiety to an extent, within anthropic.
They're very concerned about the there's this term
that they referred to, EAS something altruism.
Is it existential?
It's a movement that they refer to in the
podcast about really it's not just being able to
have the technology and throwing it out there.
But it's like thinking above and beyond.
Like, what is the impact?
What is the impact to the human element, human society?
And it just goes back to
that kind of responsibility of mankind.
And he did claim, he's not necessarily
doesn't claim to be necessarily an active
member of the community of that movement.
But again, if we don't have people out there advocating
for safe, responsible use of AI, then these types of
things, well, you're always going to have people that do
illicit, illegal, they're going to do the most greedy thing
because that's just a part of human nature. Right.
So I do appreciate the take that they're
going that Sam Altman, that these guys were
in front of Congress, that we're trying at
least, at the very least thinking about it. Right, sure.
And if we can put in those roadblocks the Constitution,
the rules, the logic, hopefully we have created a timeline
where we've got at least multiple years to figure this
out rather than it being like three months away.
Because if it's three months away, might all well go
to the beach in Mexico and call it a day
because the world's about to go to hell. Yeah.
Without these conversations, you don't talk about the
actually, this I had a role years ago
where there was a surge of what was
referred to as digital multimedia evidence.
So I don't know if you remember when
they started equipping police officers with body cameras.
Oh, yeah.
And I was working for a company
that was building technology for local county
government, prosecutor case management systems, court case
management systems, law enforcement technology solutions.
And one of the very it was kind of
like this AI discussion at the time was, well,
how are we going to store it?
Because a lot of them at the time, these agencies,
they maybe only kept it for a certain period of
time, or they kept it forever because of a certain
statute of limitations or not didn't exist.
So there was this huge anxiety around we've got access to
all this incredible DME, not to mention the body camera on
the officer, but the cell phone of the person.
That was kind of the bystander, oh yeah, and they
have to collect that and they have to store it
until and decide, are we going to go to trial?
We're going to push on this forever.
And there were multimillion dollars worth of impact
to some of these local counties, and so
they really had to open up the discussion
to understand it and solve for it.
No one really had the answer, but our company
at the time invited local leaders to come and
just, let's open this thing up, let's unpack it.
I think that's what we're trying to do here with
this podcast is let's kind of surface some of the
things that we're seeing and hopefully get some feedback in
other areas that we're not, along with this agent idea.
We're very close.
We're like on the precipice, we're on the edge
of the cliff, whatever analogy you'd like to utilize,
but we're on this edge of a cliff where
something very big is going to happen.
And if you're not prepared for it or
you don't know how you're going to respond,
you're going to be left behind.
Because the moment I can spin up ten agents to go
accomplish something that you are selling your client for, right, for
$1,000, I'm now going to be doing it in one 10th
of the time and maybe charging them less money.
I can decimate an entire market
segment if I have that capability.
I don't know.
People are like, yeah, whatever, Chase.
That's never going to happen.
But that's what these guys are thinking through, right?
They're anxious about that level of power
that they're going to have that they're
going to unleash to everybody, right.
And you're talking about just like, one county,
but we're talking about the entire world.
It's like this whole level of what yeah,
and that's where things get a little like,
okay, that's really never going to happen.
I don't believe in that.
But it is happening.
It's going to happen.
I can see in these open source movements,
people that are pushing in this direction, and
they are some level of success.
They're having some level of success.
The point is that if they can do it at
all, if it can do it at all, at some
point, it's going to get really good at it.
And that's when you need to worry.
So we're having this discussion.
Oppenheimer just came out in theaters, and
Scientific American put out this article.
Oppenheimer offers us a fresh warning of AI's danger.
So they're kind of comparing the parallels of
unleashing this AI beast into the world, putting
a juxtaposition to the nuclear bomb.
Do you think that that is
an appropriate comparison or analogy?
I think it is.
I mean, you think about the nuclear bomb.
You think about nuclear power in general, right.
It had and continues to have a substantial impact on
the entire world, not leaving out World War II.
Take in the current events, right.
Like, everybody is always on this verge of world destruction,
and the only thing that's stopping anybody from doing it
is because you too also have a nuclear bomb.
These are the things that Oppenheimer was worried about, is
like, well, if you have the power to kill the
world, and I have the power to kill the world,
what do you all do about that?
Those are the things that he was thinking about.
And in that moment in time, 1945 is the
actual year that I think this all came about.
But in that moment in time, they weren't really
worried about, is the world going to be completely
obliterated because at that point in time, there's a
world power that's just taking all of Europe.
I can understand some of the decisions that were made
to continue on with this level of technology that basically
flew over and above and beyond the level of technology
that the Nazis had, that all of Europe had, right?
And basically they swooped in and came up with this
technology that could change the path of the entire world.
And that's what Oppenheimer was worried about.
Not that we would build a bomb
and we'd go blow stuff up, right.
Like we already had bombs they were
bombing know, every night right now.
I mean, this article cites that researchers last year asked
Generative AI to design new chemical weapons, and it came
up with 40,000 potential weapons in 6 hours.
That's terrifying. Wow.
What can be used for good can be used for bad, right?
We've talked about that.
So I think we've covered a lot
of ground today in terms of safety.
And again, just such an interesting if you are interested in
hearing what the CEO of Anthropic had to say, go over
and listen to the Hard Fork podcast from July 21.
Very interesting discussion.
And I think it's, like I said, quite
admirable that they're trying to be responsible about
how they roll this technology out to people.
And I love seeing the new, I would say
smaller scale AI things that are coming out.
Like, I discovered a little plugin for Premiere Pro
that has the ability to quickly cut frames. Right.
That's something that I don't see as harmful. Right.
It's a very focused, specific use case.
But these huge models that we're talking
about and that these companies are outbuilding
and you're right, they're thinking years ahead. Right.
I'm just wondering when you're going to break out
the safety vest so we can make sure people
know that we're you know, I can probably track
down a picture of that from grade school.
I'll have to now you have to my mama tags.
Where's that you would, like, write that down right now and
then post it on the Socials so people can see it.
Yeah, we should do it.
Now that I've said it, we got to do it.
Okay, deal. All right, Chase.
Well, thank you so much for the awesome discussion.
As always, we would love to hear from our audience.
Email us your take at thejunction@bendtechnology.com certainly
if you have an opinion on it.
Either way, we want to know.
If you have questions, are there
other topics we should cover?
We are here for it.
Until next time, keep it automated. Keep it automated.
