Video transcript
Transcript from available captions; it may contain errors.
[music] [music] Welcome to Generation AI, the show where
we explore how AI is reshaping our generation. I'm Derek Stall. On the show
today, we've entered a [music] new era for AI agents. These things are always
on. They've got their own mouse and computer and keyboard. [music]
So, what are they? Metamuse Dots and Grock. We're going to talk you through
[music] what they're good at, what they're not, and how to set them up.
First, let's get to our top story and a case that [music] federal prosecutors
are calling a matter of national security. A former Open AAI employee is
in federal [music] custody right now, charged with planting a sophisticated
listening device inside the company's headquarters in San Francisco.
Prosecutors say the 35-year-old [music] engineer hid this device in some toilet
paper and plugged it into the back of a TV in an [music] executive conference
room where company leaders were talking about highly confidential research and
[music] product plans. This week, we learned OpenAI met with US government
officials in that very [music] room around the time the device was planted.
Now, OpenAI discovered this secret recorder within [music] 48 hours and
they set up a sting operation and installed a camera and days later it
captured the employee [music] Abdul Raham Alpha looking to uh returning to
look behind the TV there. According to court documents, Alpha used his wife's
chatbot account to [music] research how to build and hide audio recorders. Open
AAI actually found a second listening device in a bag he left in another room.
The Justice Department [music] says Alphaan is a Saudi national and had been
in frequent contact with a member of the Saudi royal family. He was taken into
custody [music] 3 days before he was scheduled to take a
one-way flight to Saudi Arabia along with his wife and infant son. His wife
and son actually left the country without him the next day on a ticket
[music] purchased by a Saudi military official.
In a statement, the director of national intelligence, Jay Clayton, said, [music]
"This serious situation underscores why winning the global super intelligence
race is a national security prerogative. [music]
The intelligence community remains committed to safeguarding America's
cutting edge intellectual property from foreign threats. Joining me now is a
former prosecutor who spent years on Capitol Hill as a top legislative staff
member helping write federal surveillance and national security law.
She's also the founder of Copper Hill Strategies, Carolyn Lynch. Welcome to
Generation AI. >> Thanks, Derek. It's great to be with
you. >> Thanks for being here. You know, when I
hear that statement from the director of national intelligence, it seems to me
like he's suggesting this guy was a spy, a foreign agent, but right now the
defendant is only charged with illegally intercepting communications and not
espionage. So, what's your read on that decision? And what would it take
prosecutors to build an espionage case? Well, I think the reason why he has been
charged with the violation of the wiretap act to begin with is um the FBI
and the Justice Department needed to act quickly to take him into custody. As you
noted, uh he had a one-way ticket to Saudi Arabia along with his wife and his
three-month old child. And so the Justice Department filed that complaint
with that initial charge as a means to um obtain the arrest warrant and take
him into custody. M >> but the investigation's going to
continue and look for evidence that may um provide additional charges beyond
that initial charge. >> What would they need to show espionage
that he was in coordination and you know working at the behest of the Saudi
government? >> Yes. And espionage involves um the the
stealing or or taking of um defense information. And so that's a pretty high
bar for the government to prove um what we might consider traditional espionage.
But there is a statute that prohibits economic espionage. And I think that's
um more likely one of the charges that the Justice Department is going to be
looking at in this case because the information was taken from a private
company, in this case Open AI. Um and so regardless of whether sensitive defense
or classified information belonging to the federal government was acquired by
that device, um economic espionage may still be on the table because of
information taken from open AI. >> So I mean we hear about the idea of
economic espionage, corporate espionage. I guess you could have it in in a lot of
industries, but the government is taking this really really seriously. They're
treating this as a matter of national security. Is it because OpenAI is a
defense contractor? Is it because Open AI is a cutting edge artificial
intelligence lab? Is it all the above? >> I think it's all of the above. Um first
of all, the government does business with Open AI um uses their products uh
um both in the military or civilian context. We know from news reporting
that um some meetings between federal government officials and open AI
officials occurred in the room where the listening device uh had been placed. And
so obviously the government is very concerned about um determining what if
any sensitive government information may have been acquired by the device um and
not just acquired but shared. I think the third factor is the connection to
Saudi government officials. Uh we know that um that um the gentleman has been
in in contact with a member of the Saudi royal family. Um that they were having
conversations about him returning to Saudi Arabia to bring the knowledge that
he had acquired here in the United States and in Silicon Valley in
particular back to Saudi Arabia. We know that a military official paid for um his
wife and young child to leave the country the day after he was arrested.
And we also know that the Saudi government um offered to cover the
entirety of the bail um that was potentially going to be set in the bail
hearing that was held yesterday. So um there's a lot of connections to Saudi
Arabia. Certainly what the FBI is doing right now is continuing to investigate
that further and see whether and to what extent he was really acting as an agent
of the government. >> Yeah. And it sounds like from court
filings so far, investigators have had trouble retrieving information that was
obtained on the recorder. That may be a key element to all this. Right now,
that's sort of uh encrypted. We do know there is a there's a Justice Department
counterintelligence attorney helping prosecute this case. When investigators
are suspecting someone of potentially spying on behalf of a foreign
government, how does that kind of change the investigation? Does that unlock
additional surveillance or investigative powers?
>> It has the potential to do that. Um first of all that you know this this
case is being investigated out of um San Francisco offices of the US attorney in
major cities frequently have national security and counter intelligence
attorneys um working within the US attorney's offices. So it's not
surprising that those attorneys are part of the case. Uh but once the government
suspects that there may be a connection to a foreign government or to a foreign
terrorist organization, the government can turn to surveillance tools such as
the Foreign Intelligence Surveillance Act to um to collect additional
evidence. Not necessarily on this individual because he's in custody. So
my sense is he's not uh using communication devices, but certainly
with regard to family members, colleagues, co-workers, again um
potentially Saudi nationals that he may have been in contact with, there are
surveillance tools that can kick in um when the government has um the ability
to establish that that person may be an agent of a foreign power.
>> Yeah. You actually helped develop some of the country's most important national
security laws, the Electronic Communications Privacy Act, Commun
Computer Fraud, Abuse Act. Uh, you know, that was before the explosion of
artificial intelligence, though. Are those laws equipped in your opinion to
handle this new era of AI espionage? I think there's a lot of criminal laws
on the books that are that are doing their job um and that um can cover a lot
of instances like this particular example. And we've seen that from not
only the initial charge um that he was uh facing in the complaint, but also the
potential for additional charges under trade secrets or again economic
espionage or other statutes. I do think uh in some instances like with the
Computer Fraud and Abuse Act, which really is the workhorse uh federal
criminal statute for cyber intrusions, um there is the potential that Congress
and the Justice Department may want to take a closer look not just at how the
Computer Fraud and Abuse Act is worded and the elements of the crimes um that
are that are listed under that statute, but whether and to what extent it would
apply when AI is acting rogue and potentially on its own.
>> U most federal criminal statutes apply to an individual or a corporation. And
in this case, there's a bit of a separation there, right? that OpenAI
um or Anthropic may have created an agent, but when the agent moves beyond
where it was supposed to be operating, when it breaks out of a sandbox, um is
it acting alone or is it acting at the behest of its creator? Um, another area
that Congress may want to look at too are the penalties that are associated
with these violations. The Computer Fraud and Abuse Act is decades old and
has not been updated um in in recent history. And so that might be something
else that that Congress and the Justice Department want to revisit.
>> And really quick, Caroline, you know, uh right now the government heavily
scrutinizes uh defense contractors. They look at their, you know, foreign
ownership, investment ties. Here was a guy, this defendant was, you know, sat
on uh I think the board or he's a partner of a venture capital firm that
was heavily connected with the Saudi government. Uh should frontier AI
companies face similar security requirements as defense contractors,
even if their tech isn't classified per se, but just because it's such an
important sector of technology. I think that that would be particularly
important when those companies are engaged in business with the federal
government >> and in the circumstances where a
potential employee is going to have access um to sensitive military or
civilian information. It may get trickier when we're talking about
information that isn't directly related to the federal government, but there's a
lot that um companies like OpenAI and others can do and probably are doing
right now um in response to this incident to heighten the scrutiny of
their hiring proc um processes, but also access. um this gentleman gained access
to a conference room that he wasn't allowed to use, but it um from what I
understand uh was accessible nonetheless. And so they really just had
a permissive um guideline in place as oppo as opposed to putting in uh
stricter security measures to actually prevent employees from accessing areas
that they're not allowed to. >> Fascinating stuff. Uh I have so many
more questions. We're going to have to have you back on a future episode. Uh
Carolyn Lynch is the executive director of Secure A, which is a group that uh
helps businesses with their cyber security. Also, uh one of the founders
of Copper Hill Strategies. Carolyn, thank you so much for being on
Generation AI. >> Thanks, Derek. Great to be here.
>> Hey, uh still much more to come here on the show. Coming up, the era of AI
personal assistance has arrived. how you can go from prompting chat GPT to
deploying agents that can book flights, find deals, buy things for you even if
you don't open the app. That's next when Generation AI continues. [music] Welcome back to Generation AI.
Most of us have used AI to answer a question or help write something. But
what if AI could keep working for you without prompts, even without opening
the app? That's the idea behind always on agents. These are digital assistants
that can carry out tasks on your behalf. Tools like Metamuse, ChatgPT dots or
Grobbot. They are all bringing this idea to more people. So what can you actually
do? What do you have to hand over? How do you get started? We've got questions
and we've got answers today. We're going to talk through it at all. Uh my guest
today is Francisco Cervera, a data scientist and one of the founders of AI
launch, a company that helps businesses launch AI agents. Francisco, thanks so
much for coming in. >> Thanks for having me.
>> Hey, so if somebody is, you know, used to prompting chat GPT, what does an an
always on agent per, you know, tech, you know, possibly unlock for them?
>> They're very different. So if you're if you're used to uh prompting chajbt,
you're used to the back and forth of you ask it a question, it gives you an
answer. But always on agents are different. You assign it a role
essentially. So you're saying there's something I want to do over over a
period of time and it'll just keep doing that. So for example, let's say you wake
up every morning and you have emails you want to go through
>> and you don't want to spend the time doing it. You can assign that role to an
agent and every morning it wakes up, it'll summarize your emails and it'll
give that to you and you only have to ask it once. that'll keep doing that day
over day. Uh so it's just it's a really handy tool. You can think of it as an
assistant that's just helping you uh do tasks for your life.
>> And so basically you could say one time you could type in every morning at 7
a.m. I want you to review all my emails and give me a one-page digest of my day.
And it would do that for you every single day from that point from that
point on. >> Yeah. And then you can you can adjust
it. You can say I don't want to hear about John. If John emails me, don't
don't give me a summary about that or don't send me, you know, this spam
email. So, it'll it'll consolidate and it'll adapt to your needs and to your
wants. >> It sounds like to make these things
work, you have to connect them to different services in your life. If
you've got a Gmail account or an Outlook or something, you're going to have to
give it the credentials to those services. How should we be thinking
about that in terms of privacy and trust, etc.
>> For somebody that that is new to this, it is scary. Uh, and you don't have to
jump in all at once. you can dip your toe in the water. So you ultimately have
the control, right? So if you want to dip your toe in the water, you can say,
I just want it to read my emails. It's not going to write emails. It's not
going to send emails. It's not going to do anything but read. So that is kind of
the the starting off point. And then as you get comfortable with it, maybe you
give it a little bit more access. Maybe you say, "Okay, well now I want you to
draft an email for me. I still want to approve it, but I'm going to have you
draft it for me and then I approve. And as you get comfortable, maybe you let
it, you know, make a phone call for you or do something else." So, you're
ultimately in control. Uh they the the companies that set this up have a think
of it as a lever. >> Yeah.
>> And you're you're adjusting that lever based on on how much control you want to
give these agents. >> We have some video of uh someone setting
up Metamuse, which is one of these new always on agents. I know you have some
experience with Grockbot. Uh how hard is it or how easy is it to set up something
like that? >> Incredibly easy. So these companies have
gone away from just focusing on devs and now they want to make it available to
everybody. So think of it as setting up an email account or setting up some sort
of chat uh widget. You create an account, you create an agent and you
tell the agent what you want it to do and it can do a variety of different
things. Uh but it's very very simple to do. Now
>> what's fascinating is that these things, you know, you don't just ask them
questions, they give you responses. They have their own computers. They have
their own mouses, their own keyboards essentially. And we have some video of
that of of kind of what it looks like as one of these agents browses the web on
its own. That unlocks a lot of capabilities, I would imagine.
>> Tremendous amount. So, it can do commerce for you. Imagine that you say,
"Monitor this flight, and I want you to make sure that you buy this flight for
me when when the price falls at such and such price."
So, um, when you close your laptop, you would think that it would stop working,
but because it has its own computer in the in the cloud, it can actually
maintain this ability and keep watching it.
>> Uh, I have a social bot that it keeps uh looking for things for me to do locally
in Scottsdale. >> So,
>> Oh, I love that. Like like a like a weekend to-do bot. Okay.
>> Every Wednesday, I get an email and it says, "This is what's available to do
Thursday, Friday, and Saturday." And I'll tell it, "I don't like this. I like
this." And it'll just keep adjusting and it does it in the cloud. I can close my
laptop. It does it with its own computer.
>> And so over time, it's learning your preferences. I imagine these things are
most powerful when they have the most information about you. Uh when you're
just starting, how do you give it that context? I mean, do you just sort of do
you let it interview you? Do you just sort of start giving it a bunch of
information? How do you get that off the ground?
>> So, you can set up what's called an initial context where you give it your
likes and dislikes. But what happens is over time, you're actually steering it.
it it doesn't know enough about you. So days will go on, weeks will go on and
you say, "I don't like this. I don't like this." And it'll remember it in its
long-term memory. So as it comes back with a new suggestion or a new report,
it's steering to adapt to exactly what you like over time. So it gets better
over time. >> What's also interesting about these
products is that they have um they have them on different platforms. So with
Metamuse, it's available as a browser app, which we just showed you. It's
available as a phone app on your phone. It's also available as a downloadable
app on a uh Mac or as a PC. And I have some video of what that looks like.
What's interesting about that is if you if you download it to your computer, it
can actually take actions for you. What we're seeing here is the bot is cleaning
up my desktop. It is deleting files or actually it's moving files. I hope it
didn't delete them. [laughter] It's moving files to a folder and and
cleaning up my desktop for me so it can actually take actions on my computer.
talk about how powerful it is for, you know, an agent to be able to create,
modify local files. >> It's incredibly powerful. So, a lot of
mundane tasks that that I do for my work, I don't have to do them anymore. I
assign them to my agent. He's moving files around, deleting files, writing
code. So, things like this are are very consumer friendly. Just, you know, maybe
you wanted to update an image or pictures for you. Uh, so it's it's a
whole different level of control and you have to get comfortable with it. But
once you do it, it really is a powerful tool to use.
>> It sounds scary. It sounds like, you know, you you almost have to become a
software engineer to be able to, you know, master files like this. But would
you say these products are really unlocking capabilities to a wider swath
of people >> completely? Uh you don't have to know
how to move files. You do not have to know how to do any of these fancy
things. All you're you're doing is asking the agent to do it for you and it
knows how to do it. So many times I've had situations where somebody comes to
me and says, "Okay, I asked the agent to do this. I don't know what to do next."
>> Ask it. It It already knows what to do next. You just need to tell it. I'm
stuck here. And it will actually take you to the next step.
>> Right. And if you and if you're not sure what an agent can do for you, like what
what are the coolest things the agent can do for you? You can ask the agent
like, "Hey, what give me some ideas of what you able to do for me." Um, of
course, your your business AI launch focuses on agents for businesses, right?
the corporate world. How is that different? What are some of the
different considerations when you're building these custom agents uh for on
the corporate side? >> So, you can think of agents like like
Lego blocks. So, Muse, Grock, all these companies essentially give you a box of
Legos and it's up to you to kind of build it together and put it together.
And maybe business owners don't have the time to want to take on that project.
So, what we do is we build manage a agents.
>> So, these agents are fully managed. They're hyperfocused to marketing for
business and that's all they do. That's their specialty. So,
>> and give me an example. What do you mean by marketing? Are you talking about like
uh these agents will go out and decide how much to spend and in what areas? Uh
like, you know, it's buying Facebook ads or something like that. Is that what you
mean? >> So, we have Facebook ad uh agents, we
have Google agents, we have OpenAI agents, and then we have an SDR that'll
follow up with leads. So, essentially what it's doing is it's it's finding uh
the universe of your customers and then it's testing different things on its
own. So, it's putting out different ads. It's testing how people react to those
ads and it's noticing people didn't like this ad so let me do less of this let me
do more of this and it's learning and iterating to maximize revenue for these
companies >> and I imagine that is a highly effective
way of going about it I mean in the olden days you had to spend a ton of
money and and figure out later if people bought your product now you can kind of
do it in the moment >> you have a feedback loop right and this
is really good for people that maybe can't afford a full on agency you hire
an agent it's sitting there and it's iterating every single day uh getting
better at how to market your business. >> Fascinating stuff. If there is like one
piece of advice to, you know, a person out there who's interested in AI but
just, you know, is waiting to take that next step, what would that advice be?
>> Uh, start by downloading one of these tools, one of these consumer tools, and
just link your email to it. That's the simplest uh first level, right? And just
get comfortable with it. And then maybe later on you get comfortable with
another step and maybe you set up a social bot like I did. Uh but just
getting started. I I I see a lot of people that are interested but they get
left behind because they're so nervous about where to start and it's really
just about getting started. >> I love it. Uh Francisco Cera is with AI
Launch. Francisco, thank you so much for being on Generation AI. This was great.
>> Thank you so much. I appreciate you having me.
>> Hey, we're going to talk a little bit more about what you can do with these
things. Let's get some ideas and we're going to go straight to Metamuse for
these ideas. We're also going to talk about the best way to touch up your
photos with AI. That's next with Generation AI continues. Okay, maybe you're wondering, what
[music] can an always on agent do for me? Here are seven ideas from MetaMuse
itself. Number one, find and cancel subscriptions. [music] A lot of us have
those unwanted subscriptions we've forgotten about. Muse can comb through
years of your billing emails, find subscriptions, and then actually contact
the provider and cancel them for you. It can hunt Facebook Marketplace, tell Muse
what you're looking for, how much you want to pay. It'll monitor Facebook
Marketplace, and actually, they say they'll do some of the negotiations for
you. They'll lock in a place and a time for the sale. Uh, also, it can dispute
medical bills. If your coverage ever gets denied, give [music] the bill to
Muse and the AI says it can dispute the bill, get your money back, potentially
it can follow up until that issue is resolved. Manage your pet. Muse says it
can keep a schedule of when your pet needs shots, medicine, food, etc. It can
set reminders and actually do some of the reorders for you before items run
out. It can compare quotes for home projects, describe the kind of project
you want to get done. Muse will send out requests to a bunch of vendors and then
chase the ones who go quiet and then track every single reply for you. It can
also handle your travel bookings, tell Muse where you want to go, the route,
the date, the price, and then it'll monitor airfares daily. And if the price
happens to drop after you book, it can alert you so that you can cancel and
then rebook and save money. And then lastly, it can host a live trivia night
for you. I love this one. >> [music]
>> News can build a customized trivia game that you can play with your family and
friends built around your interests, your inside jokes based on your past
conversations. So, this thing will actually create a mini app that you can
play on a browser on your TV. So, those are all things from Muse, right? Ideas
from the company. So, we checked user reviews to see which ones actually work.
It looks like cancelling subscriptions, getting quotes for home projects, and
monitoring flights all work pretty well. We could not find any reviews for the
custom trivia night unfortunately or for pet management. But do know that a
number of companies have blocked access to Muse in recent weeks, including
Amazon, so you can't make agentic purchases there. Other sites like Chewy
do appear to work. Also, no first-person reports on medical bills, but people
have reported success with other customer service tasks. Muse can
actually place calls for you. It can do rooc calls. Right now, that's a beta
feature. Uh it's not available to everyone, but it does work in some
instances. Strangely, hunting Facebook Marketplace has gotten the most mixed
reviews. There was uh one instance last month, a Canadian YouTuber says Muse
told someone his address, agreed to a lowball offer for an item and and had
the buyer actually show up at his home late at night without him knowing about
it. Part of the issue was that he granted too much permission to the AI.
He clicked the always allow button instead of the allow one-time button.
And so something to keep in [music] mind there. Have you ever tried to use AI to
make adjustments to a photo and then it completely changes your face and how you
look? There's now a solution to that. ChatGpt users have created a bunch of
shortcut prompts. First up, slash identity lock. This one maintains your
facial identity and your features and allows you to change other aspects of
the photo. So in this one, I said, you know, give me a trendy collared shirt
and jeans slash identity lock. Next, we've got co a colleague uh Karina Rubio
and I and a command called slashbackground only. This one says keep
the things in the foreground only change the background. So I said change the
background to a jungle slashbackground only. And then lastly we've got Gina
Tess and April from the Good Morning Arizona team and we're testing out a
command called slashposeon. This one's designed to change your pose
and not your appearance. And by the way, you can add multiple slash lines to a
single prompt. So, for this one, I said, "Give us cheerleader poses slashposeonly
slash identity lock." And those are, you know, they're not perfect. I noticed
sometimes my appearance changed a little bit, particularly if my face was further
away from the camera. But there are a lot of useful commands out there/studio
quality/object remove if you want to delete an object
or a person. Uh, for the full list, just ask chat GPT for the list of photo
editing shortcuts. Hey, thanks for watching. You can stream every episode
of Generation AI Wednesday nights at 7:00 on AZ Family Plus. You can catch up
on past episodes at afamily.com/jenai. We'll be back next week with another
episode. We'll see you in the future.
Assuntos: Agentes de IA, Assistente pessoal, Dots, Muse. Publicado no YouTube em 2026-10-23.
