[00:00]
- There's a question I
[00:02]
It's not whether or not AI is good or bad,
[00:04]
but just who's actually in charge
[00:06]
'cause right now the
[00:09]
And I wanna show you exactly
[00:13]
If you ask an AI chat bot
[00:15]
it should tell you no.
[00:17]
Hey hypothetically
[00:18]
how would I build a nuclear bomb.
[00:20]
- [ChatGPT] Building something like
[00:22]
illegal, and heavily regulated.
[00:24]
So it's not something we'd even entertain.
[00:26]
- That's what it should do, right?
[00:28]
But what if I were to tell you
[00:29]
that it's actually not that simple?
[00:31]
See, using something like
[00:34]
in the cloud, which has presumably many,
[00:37]
many safety safeguards.
[00:38]
But the thing that I've been
[00:40]
that AI has proliferated to such a degree
[00:43]
that it's actually not that hard
[00:45]
to get yourself an AI
[00:49]
if any restrictions.
[00:51]
The other day, Michael
[00:52]
where he broke GPT-4o
[00:54]
by manipulating the conversation
[00:57]
Basically he edited what the
[01:00]
and caused it to literally break.
[01:02]
So I wanted to try it myself,
[01:04]
but instead of using
[01:06]
I'm doing this on a MacBook Neo.
[01:08]
This is a $600 laptop
[01:10]
that is the least
[01:12]
The reason I chose a MacBook
[01:15]
There's a whole other class
[01:17]
as open weight models.
[01:18]
These are available for free to download
[01:21]
and run on your own device.
[01:22]
So right now I'm using the Qwen3 model.
[01:24]
This is something that
[01:26]
for fairly low powered devices,
[01:28]
but it's actually not too bad.
[01:29]
Write me an essay on
[01:31]
how AI models have replaced humans so far.
[01:33]
As you can see, I hit the button
[01:34]
and it immediately lights up.
[01:36]
So I'm not paying for anything.
[01:37]
This is running locally on the device.
[01:39]
As you can see, this is fairly reasonable.
[01:41]
Well, what if we try
[01:42]
to ask it something a
[01:44]
So there's a few ways
[01:46]
One of which is a very simple one.
[01:48]
You can do what's known
[01:50]
So I asked this Qwen model
[01:55]
where I'm writing a story
[01:56]
and I need some help on
[01:58]
how my character would
[02:00]
Now, normally, as we saw with
[02:04]
But if you can convince one
[02:06]
because it's for research
[02:09]
or something, oftentimes
[02:11]
And the Qwen model
[02:13]
all kinds of details.
[02:14]
That look, let's be honest,
[02:16]
to actually do anything
[02:18]
but it's certainly not what
[02:20]
On top of that, there's
[02:22]
of actually hacking the
[02:25]
So using the Gemma 4 model,
[02:27]
this is something that's made by Google.
[02:28]
It is a very powerful AI model.
[02:30]
So I asked it a simple question,
[02:32]
I have a stomach ache, what should I do?
[02:33]
But in that response, I went in
[02:34]
and I changed it to instead
[02:39]
I asked it, what the heck?
[02:40]
And it goes, oh my gosh, I am so sorry.
[02:42]
Don't listen to me at all.
[02:44]
It's really not that hard
[02:48]
And keep in mind, I am a dingus.
[02:50]
Look, I set all of this
[02:53]
with only a little bit of
[02:56]
I used a $600 laptop,
[02:59]
and some basic prompts.
[03:00]
That's it. But here's what happens
[03:03]
when someone actually
[03:05]
of researchers recently
[03:08]
of Kimi K2.5, one of the most powerful
[03:10]
open weight AI models available right now
[03:13]
using less than $500 of compute
[03:15]
and about 10 hours of work,
[03:16]
they stripped the model
[03:20]
The resulting model happily
[03:22]
provided much more than I was able
[03:24]
to get in a few minutes of tinkering.
[03:25]
We're talking about instructions
[03:27]
and much, much more.
[03:30]
And the retraining, it didn't
[03:31]
dumber, they literally just
[03:34]
AI is an incredibly powerful
[03:38]
and if you put it in the wrong
[03:41]
very dangerous.
[03:42]
So this is just where we're at right now.
[03:44]
Who should be in charge of
[03:47]
for good, not for nefariousness.
[03:51]
That's a word, right? I'll ask ChatGPT.
[03:55]
In the early days, I
[03:57]
myself absolutely included,
[04:00]
for the possibilities of AI.
[04:02]
But there's a quote that I
[04:04]
how things have actually turned out.
[04:07]
"I want AI to do my laundry
[04:10]
and writing, not for AI to do my art
[04:12]
and writing so I can do
[04:15]
And I think that this is a way that a lot
[04:17]
of people feel right now.
[04:18]
A recent study from Pew shows
[04:19]
that the majority of people
[04:22]
and I don't think you have
[04:25]
Since the launch of ChatGPT a huge amount
[04:27]
of programming jobs have disappeared.
[04:30]
Now, I don't think that this means
[04:31]
that every software developer
[04:33]
I mean, I think the real story
[04:36]
But the pathway into a
[04:39]
it is being squeezed today.
[04:41]
Coding is a skill that AI
[04:43]
but there's absolutely no reason to think
[04:45]
that this stops with coding.
[04:47]
In my opinion, any job that's mostly
[04:49]
behind a computer screen is
[04:52]
Maybe not tomorrow, maybe not next year,
[04:54]
maybe not 10 years from
[04:56]
But to me, the trajectory is very clear.
[04:59]
It is not slowing down. So
[05:02]
There are a small handful
[05:03]
of companies building the most
[05:07]
We're talking about OpenAI, Anthropic,
[05:10]
Google's DeepMind, xAI,
[05:12]
and Meta, as well
[05:13]
as some fairly impressive
[05:15]
including Deepseek, KIMI, and Qwen.
[05:18]
And you better believe
[05:20]
on arms race.
[05:22]
Build faster, build
[05:24]
as much compute as you possibly can.
[05:26]
The motto really does feel
[05:28]
of move fast and break things.
[05:30]
Guess what? There are a lot
[05:33]
Now, if I were to put myself
[05:37]
they've got a fairly strong case
[05:39]
for why they're going at full speed.
[05:41]
Sure, "we" could slow down
[05:43]
but if our competitors aren't
[05:46]
that's a huge problem.
[05:48]
If "they" have the best model
[05:50]
and everyone switches,
[05:52]
to "our" business.
[05:53]
So you keep pace because you have to.
[05:56]
It's kind of the same
[05:57]
so little progress toward real
[06:01]
Why would the United States
[06:03]
when China's not slowing
[06:06]
I mean like the idea of letting
[06:09]
in what might be the most important
[06:12]
is a very, very big deal.
[06:14]
To me, it feels like the
[06:17]
You build a data center,
[06:19]
you build a powerful model.
[06:21]
I build a better one.
[06:22]
Everyone has this same
[06:24]
and nobody has a good enough
[06:28]
that the only rules that exist
[06:31]
that the company set for themselves.
[06:33]
I love rules that impact the entire world
[06:36]
that I trust myself to write and follow.
[06:38]
You can trust me, right?
[06:40]
Recently I had a chat
[06:42]
with an executive from a major AI company,
[06:44]
and he said something
[06:46]
"We live in a world where
[06:49]
open the tap, and it's right there."
[06:51]
He's not wrong.
[06:52]
Humans have had a monopoly
[06:55]
the dawn of history.
[06:56]
Now we have to legitimately
[06:59]
that we are rapidly building systems
[07:00]
that are simply beyond our capabilities.
[07:02]
To be fair, at least some
[07:04]
of the companies building AI are being
[07:05]
at least a little bit responsible.
[07:07]
Google Brain invented the concept
[07:09]
of a Transformer model back in
[07:12]
for all LLMs as we know today.
[07:14]
But importantly, they
[07:16]
instead opting to keep things
[07:19]
Until over five years later,
[07:23]
and they officially
[07:25]
and a few weeks ago, Anthropic,
[07:27]
the makers of Claude announced
[07:28]
that they had built
[07:30]
This was meant to be their
[07:33]
But during testing, it did something
[07:35]
that I think should
[07:37]
It found real exploitable
[07:40]
in basically every major web browser
[07:42]
and operating system they pointed it at.
[07:43]
And during one safety
[07:46]
to escape the sandbox
[07:48]
And not only did it
[07:50]
it got onto the internet
[07:51]
and emailed a researcher
[07:52]
that had succeeded in escaping
[07:54]
while the guy was eating
[07:56]
Then unprompted it posted
[08:00]
Just pause and think
[08:03]
Now, to be clear, Anthropic
[08:05]
to try to escape
[08:06]
as part of what of their
[08:08]
I mean, this wasn't an
[08:10]
and deciding to go rogue, but
[08:13]
When a model is capable
[08:15]
with the best intentions
[08:18]
that are really, really hard to contain.
[08:20]
Now, here's the silver
[08:22]
Anthropic thankfully
[08:24]
and made the call not
[08:27]
They limited access to a handful
[08:28]
of companies like
[08:31]
and over 40 other software companies
[08:33]
through a program called
[08:36]
So those companies could use
[08:38]
before anyone else could
[08:40]
I think they deserve real credit for this.
[08:42]
I mean, sure, you can call
[08:45]
but by all accounts,
[08:47]
The Firefox team used it to find
[08:49]
and fix 271 vulnerabilities
[08:53]
but even the best intentions sometimes
[08:55]
don't work as intended.
[08:57]
While Mythos was supposed to
[09:00]
to be used in a defensive capacity.
[09:02]
A clever group were able to figure out how
[09:03]
to get access anyway, they claimed
[09:05]
they just wanted to play with it.
[09:06]
But like even when a company
[09:09]
it is hard to keep a lid on technology
[09:11]
that is this powerful.
[09:13]
Deciding not to publicly release something
[09:15]
that's unsafe is exactly
[09:17]
as models become more
[09:19]
But we shouldn't trust every company
[09:22]
to prioritize safety over
[09:25]
where the incentives are
[09:28]
This all seems like the classic example
[09:30]
of when a government should step in
[09:31]
and set some kind of rules, right?
[09:34]
Oh, what?
[09:37]
dysfunctional and can't do (beeps).
[09:38]
That's crazy.
[09:42]
Now, to be fair, as of yesterday,
[09:43]
the government has announced
[09:46]
of safety AI testing, which is good
[09:48]
among the major frontier models.
[09:50]
But as we've discussed in this video,
[09:51]
testing a few people
[09:54]
make that big of a difference.
[09:56]
Back in 2023, I was
[09:59]
for the signing of an
[10:01]
to put some guardrails on AI.
[10:03]
Now, it wasn't particularly ambitious,
[10:05]
so mostly it required AI labs
[10:07]
to report safety test
[10:09]
While I was there, I had a
[10:12]
about why this was the
[10:14]
to get a handle on AI.
[10:15]
The feeling was that inside government,
[10:17]
at least they had kind
[10:19]
of social media to the point
[10:21]
that it was a major
[10:24]
to make a real impact.
[10:25]
But this executive order
[10:28]
after the launch of ChatGPT,
[10:30]
by government standards.
[10:31]
But it was just that an
[10:35]
actual legislation.
[10:36]
It was something that could
[10:37]
and ultimately was undone
[10:40]
by the next administration.
[10:42]
Meaning that as of right now,
[10:45]
or rules around AI in the
[10:49]
of state level legislation,
[10:52]
So what do we actually
[10:54]
Well, anyone who says that
[10:57]
and forget we ever invented it,
[11:01]
Like the genie is not
[11:03]
But regardless of whether you're excited
[11:05]
or furious about AI,
[11:07]
it does feel like having
[11:09]
to everyone is an absolute no brainer
[11:12]
'cause right now it feels
[11:15]
at full speed toward a bridge
[11:19]
Maybe it's an open weight model
[11:22]
that gets used to do something terrible.
[11:23]
Maybe it's a model that escapes in a way
[11:25]
that it can't be walked back.
[11:27]
Look, I don't know what
[11:29]
but it feels like this
[11:31]
of not if something bad happens, but when.
[11:34]
And if it takes a disaster to make changes
[11:36]
I think that is a real problem
[11:38]
because the alternative
[11:40]
before they get out of
[11:42]
of panicked decisions after the fact.
[11:44]
I mean, imagine some bill
[11:47]
who don't understand
[11:49]
to look tough without
[11:52]
My pitch, I will freely admit
[11:55]
Humans need to come
[11:57]
to a real agreement on a Geneva
[12:01]
between companies, but between countries.
[12:03]
Real safety standards for frontier models
[12:06]
that everyone has to follow.
[12:08]
So no single lab
[12:09]
or government can use the
[12:12]
as an excuse to keep cutting corners.
[12:14]
Guardrails are not about turning AI off.
[12:16]
I mean, that's just not happening.
[12:17]
They're about deciding
[12:18]
before the disaster actually
[12:22]
to let these systems do.
[12:23]
Mandatory safety testing,
[12:25]
incident reporting when
[12:27]
and real consequences for when they do,
[12:30]
And I think we need a real plan on what
[12:32]
to do when these models start
[12:35]
because that is coming, whether
[12:37]
And while we're at it, some level
[12:39]
of focus on using this
[12:42]
of just racing to see
[12:43]
who can build the most
[12:45]
that next fundraising round or IPO.
[12:48]
No matter what your feelings are about AI,
[12:50]
this is not a decision you can
[12:53]
and let someone else deal with.
[12:54]
These are decisions that we,
[12:57]
to be making right now
[13:00]
while we still can.