Episode 28: Adversarial Agents from London

I've been some
I'm coming for you from a bus stop. And
the advertising Singapore here um
Funny enough, that's not my bus.
And um
Yeah, this is London. This is what
London looks like.
Can you hear me clearly? Yes, I can hear
you. They are advertising Singapore and
this is London.
Yeah.
>> or what? I thought like you had some
super deluxe bus stop area where you
could sit with a laptop but you're
calling from the phone. No, no. To be
To be honest, the fact that I have my
phone out and the cyclist bugging me by
it shows you I have nerves of steel.
Because it's quite common to get your
phone nicked here.
Yeah, well, that that was Saigon 5 years
ago. I don't know. I mean, you guys
>> What did they do? Stop the crime in
Saigon?
Cut people's hands off?
Uh well,
I don't know. I'm not in District 1
anymore, but like
it's surprising to hear that London is
like that. Belgium wasn't like that in
2006 when I left.
To hear Europe being in a situation that
Saigon, Vietnam, a third world country
is in Yeah.
>> 6 years ago. That's the thing that my
wife and I joke about because
like for example, to
our experience with building stuff here
in England is like third world.
But mentality like
you have to pay someone, you have to pay
someone, you have to pay someone. It's
just pay someone for that. You can't
actually build.
Yeah, so
Yeah. And that's true, actually. All
right, so I just wanted to say that the
conference yesterday was great and I'm
going to It's only 2 more days of it.
It's incredible.
So I'm looking forward to that. And uh 4
days? Yeah.
I went to an
I went to an anthropic talk and I went
to
They just made a whole bunch of
announcements at Anthropic.
>> Yeah. And
It's insane. Um
The It was a workshop, but it
It was more like a presentation. They
would like to They were basically going
through that that blog that I think we
we kind of
might have mentioned to each other.
The one about
the one about the harness, the one about
the
the agents, adversarial agents, and the
evaluators, and things like that. Oh,
yeah. Yeah, I adversarial agents is the
keyword now. I think I called it
adversarial, but it's adversarial, of
course, like I would
The other word that they kept on
dropping like it was hot was rubric
rubric, which I think I had to look it
up,
which means
instruction. They put it rubric
for some reason. Have you ever heard
this expression or term?
No. What's What is it?
Rubric.
Rubric?
Mhm.
Not sure what that is.
Um
But I
I thought
I I have a I have a good idea how to you
know, create my own harness. If I have I
have like a good start about that.
So, I I have a good idea about creating
something that can spend $100 and come
up with a you know,
a richer game maker or some other some
other thing. So,
so they they equipped us how to spend a
chunk of change, and then of course,
I mentioned like, "Oh, what happens if
it's going wrong?" It's like, "Oh, you
have trust issues."
It sounds like your harness or your
evaluators aren't good enough or
something.
Okay, this is my boss. And you said,
"No, it's Claude Code that gave me trust
issues."
Well, I'm going to mute myself.
>> Yeah, that's
They're in the bus now?
I'm in the double decker
Kikikik double decker.
But busy. I don't know what the empty is
down over there.
God, I'm going to be that guy on the bus
talking.
Hate those people.
Uh
And the the other talk that was really
interesting was
by Pedantic. I don't know if you've ever
used Pedantic.
Pedantic is a Python unit testing
library or what?
Or is it Um
behavior driven behavior testing I know
it is like a schema manager, like a
Oh. like a model manager.
Why is my face like this? Like they're
super
It's like somebody smashed me in the
face and there's
Yeah, so so that's pretty cool. I mean
I thought you were going to Yeah, so you
did the workshop. What else did you, you
know, pick up from the hallway track or
what they call it?
I can hear you, but
it dropped.
Well, then
while
we're waiting for Kai to get back, let's
talk about
amazing Project Glass Wing.
I mispronounced it.
Became embarrassing.
Glass being a project that Anthropic
announced with their latest Is it
Mythos?
model
uh basically
identifying so many vulnerabilities.
Yeah, I'm back. I'm back. I was just
talking about project Glass wing and
Mythos and is it Mythos, right? The new
uh
Mhm.
model that Anthropic does not want to
release until the internet has had the
time to prepare.
Yeah, there's there's definitely some
hype about it and and the uh
And during the Anthropic
Well,
the Anthropic um
Say that Like the crazy thing about the
Anthropic talk was that he was like
talking about the
the harnesses and the evaluator and the
and the prompts that you would do. But
then he warned that when Mythos comes
out, the
your harness would have to be updated
because
because you know, Mythos is a lot
better.
So, your harness would have to like
focus
on other things.
You know what I mean?
>> Yeah. It's like it's like, hey, we're
doing all this work but like once this
new model comes out, we'll be able to do
things better. So, you just have to
like, you know,
push yourself down the line a bit.
Yeah, I mean, there's this interesting
talk I think Is it Simon Wilson on his
interview but I think somebody else as
well. They talked like our role right
now is to
Simon Wilson Willison Wilson I don't
know how to pronounce his last name. He
definitely talks about keeping uh
a backlog of projects that you you try
and then you see the AI cannot do and
you put it on the backlog. Like you say
this one day we'll be able to do it and
just keep going back and trying out
those projects. So, do keep feeling out
like the boundaries as they are moving
as models are getting better, like where
are we now? Where How far can we go? And
to keep living on that like they call it
the border. There was I saw two
different talks YouTubes YouTube
presentations about um
what we can do because I think the the
biggest problem is
which I think we mentioned before is
not realizing
what the model can already do and still
doing things that we should just hand
off to the model. And I
think
Cuz like Anthropic announced two things,
right? They announced this um
three things. They announced the project
last week and the impact of of that cap-
capabilities detecting decade-olds um
zero days inside foundational
infrastructure in the internet would,
you know, create huge impact on the
banking systems and the trust that we
can have on on on I mean, even if it's
marketing from Anthropic, which I don't
think it is, but you may assume it is.
Um it is a fact that that these these
models like Opus, they're able to detect
bugs and then
pra- practically zero days, right? And
if the bug bounty is not being awarded
properly, um these get sold on the black
market or state actors and hackers abuse
these to, you know, have real
significant impact.
>> Yeah. I was I was had a thought that
like
AI agents
you can you can effectively
organize botnets of agents nowadays cuz
that's all the rage, right? I think it
was quite expensive, but that's what
Cloud Botnet is, right? I mean, the
whole Cloud Book was basically a botnet,
um you know, directed by people to
simulate
um all kinds of stories. And then people
are like, "Oh my god, these agents are
planning a revolution. They're designing
their own language." Yeah, yeah, yeah,
okay. Um but but that was one of the
announcements, right? Basically, we're
reaching a
a level of of of intelligence that that
it's rare amongst humans, and humans
have that have that expert level
capabilities are the ones that are able
to basically hack into
software, find zero-days,
and I mean
I really enjoyed reading Phineas
Hacker's breakdown of how he hacks the
hacking team and how he built like a
zero-day on a router that he knew uh
like a a different an um
type of router device that they used,
and he he tried it his exploit against
several other routers on the internet
before he actually
attacked the target. So, this type of
intelligence
there are humans that have it, and they
are able to do it clearly.
And if you put that in the wrong hands
It makes me think
Cloud
Cloudflare is going to do well because
they have
all these security products to
stop agents to the moment.
It's not a bad observation, yeah. I
mean, any any organization that
that is I mean, Cloudflare is known for
their extremely good
DDoS protection and so on. Don't get
started Don't get me started on their UI
skills. They're the
worst.
Um
but they anyway, the second announcement
was the a product for agent like I don't
know, I haven't looked too much into it,
but it sounds like hosted agent teams.
So, you can run or build agent teams.
Now,
um or you can just have an
Anthropic hosted agent teams, which is
something really super interesting.
I think he I think I think the guy
mentioned it yesterday, and they
coordinate using the file system, I
think he was saying.
Um
There's There's a lot going on with
Anthropic. May-
L- I mean, there's quite a few startups
that are going and saying like, "Okay,
you know, Stripe published a markdown
saying that this is our agent minion
agent system. Uh we're not going to open
source this, but if you take this
markdown, you can build your own.
Um and so I think there's a lot of
startups that that were built around
that saying like, oh, you know, Stripe
can build it, but not a lot of like
medium-sized companies or small startups
cannot build it. So there's definitely a
need for a SaaS that provides this. And
I believe that like if you're going to
pay Anthropic for the module, then I
might as well use Anthropic for the
agent team if they provide a service,
right? Why would you need to pay some
third-party startup
um to pay to to do agent orchestration.
Uh but there's
discussion that I I had a lot with my
friend because I'm talking about hey,
this spectrum design is really good.
Um
but they're just going like
now you just need a you just need a
single skill. You just need a brainstorm
skill and Opus and um you don't need to
go through this whole spectrum design
workflow anymore. Um
you get like an interview experience
where it asks you, what about this? What
about that? Just a few key points that
you want to clarify and then it goes off
and builds it. Like you don't need all
of that additional stuff anymore and it
I think we're definitely going towards
that.
>> Have you Do you Do you know what I talk
Do you know what I mean when I have to
share the slides with you? But do you
know what I mean by by um
an evaluator and an adversarial agent
because
And I know you since you do spec dev or
and spec dev
spectrum development
the the way that they explained it is
that the plan is very high level
and then the emphasis goes in the
evaluator. So basically in the evaluator
stage
you put the effort
I mean, this is my interpretation.
But my interpretation is that like
your evaluator is basically where your
spec now lives and because
you're you're basically at the role
where
the product owner is basically or the
the client or something is is basically
evaluating what was built.
So, now you have to
uh make sure the evaluator has I guess
your requirements and your expectations
all there
so that it can evaluate
the the build uh stage.
Does that make sense? Yeah, makes sense.
I mean
I didn't I I hadn't thought that far
because because I always I use
adversarial agents a lot now. I really
love them.
Um but it does mean I I can do it very
effectively because I spent so much time
planning. So, because I spent so much
time planning
I can tell the agent now verify that
whatever we build matches and every
single time there's something. There's
like it most most of the time it just
skips the tests. It just says, "Oh, this
no needed for this test. Uh that one
would take too long." Like last time it
told me I'm not writing I'm not writing
the test because it wouldn't be suitable
for a unit test because it would take
too long to run because it run it forks
out to get clone. I'm as like, "Yeah,
but I would have liked to known that,
you know, that you are not writing those
tests because of that." Cuz they're all
just missing. So, the adversarial re uh
review always surfaces these stuff. It's
always super interesting. So,
but it does require like you say an
evaluator to be very much aware about
exactly what do you want. Otherwise
what can an like
like
uh what can an adversarial review give
you if you don't have a clear definition
of what you expect versus what you got,
right?
Yeah, exactly. Exactly. So,
oh, the the other tip that Ash dropped
which I think had a chuckle in the crowd
was uh
he was saying that when you write your
um
write your what do you call it? Your
specs or what you want or your tips
or your breadcrumbs
in JSON because for some reason he was
saying that the
that the the agent will
or less likely override the JSON.
And
I thought that was a weird but That is
weird cuz like I don't know with me I
want to ignore the JSON.
>> I want the agent and me to own the the
markdown together. I like the agent is
constantly rewriting it and I'm always
really careful looking at how it
rewrites it based on my instructions. So
for me markdown is pretty good um
but maybe
>> Yeah yeah like nothing I could
catch with a uh Yeah I think it would be
much harder for me to review the diff if
it was JSON.
Yeah.
So I don't know that's a bit weird step.
Let me let me talk about the Pedantic
one which I thought was really
interesting. So there's a CEO of
Pedantic Samuel I can't remember his
surname.
I love to see the guys like this. He was
very
kind of inspiring. He's like
the alpha geek in my opinion you know.
He's the CEO of a company.
I think Pedantic company's it's got
I don't know how if it's doing well but
like
it's great to see the CEO like
obviously hands-on coding the
the demo
and giving the presentation
and
anyway I just love to see a CEO who
knows his stuff.
And then
he presented a framework
an AI observability framework where
where essentially
you
uh the demo I thought was pretty cool.
It basically
uses a scan technique to sort of refine
the prompt.
So basically you can
you can go
it picks the best things of the prompt
and
and and so that it has like a
an agent to sort of
find a
a better or try different other prompts.
And then has like, you know, 100 cycles.
And then it And then you have like a
like a golden
uh test set. You have a test set that
says like what the right answers were.
So then the agent's able to to change
the prompt so that the prompt gives
better and better and better results
with high accuracy.
I thought it was I was super cool, but
of course I asked the question like, you
know, if there's a new model
isn't this all this money spent going to
going to waste?
And he said yes, but like in certain in
certain use cases like uh you know,
where things
uh
where things need to be a bit more
uh accurate or it's
deterministic is the wrong word, but
like when you when you want your prompt
to be accurate
if you want your prompt to be, you know,
instead of 90% accurate 99% accurate
this is the way to do it or or
this is one approach to consider.
And I thought it was fascinating because
I did You you do start from like a
stupid prompt like, you know, find the
family relations of these politicians.
And then
and then once you throw throw it into
this GAN
um
uh
harness
you have a much more involved um
prompt
which has all those words like, you
know, carefully select and and and has
all the steps.
I thought it was fascinating how things.
I heard also about industry jargon. Ah,
I remember when when um there was I saw
some Twitter
about design because a lot of people
like complain about AI design looks the
same. You can load these um
you know, Anthropic web design
things, but um
ultimately the the generated designs are
not very good. And so, there was this um
I think it's from the Google Stitch
announcements when they talk about
because Stitch is really good at
creating web designs now.
And one of the UX designers at Google
went through the the procedures that he
took
um to go from a mediocre web design to
like a real expert level design where he
says like, "As a designer, we talk about
these terminology and this is a
terminology you need to be a familiar
with if you want to manage the design
created by a model to look professional
and not look like slop." And um and and
I think it was also related to industry
jargon like being able to be precise by
using the right terminology because the
models they know, right? They've been
trained all of it. It's just up to us
like if if we know the terms, we can
then trigger the the the the level of of
um
of what do you say, quality that that is
that we will accept if we are if we are
familiar with it. But if we're not, we
will just accept some low-level quality
web design because we're like, "Okay, we
know very basic stuff and this looks
good enough, right?" But if you're an
expert web designer, you know all the
all the trigger terms that basically the
model knows like, "All right, we're
talking about some serious professional
web design here. We're going to do
We're going to make something that has
the proper, you know, what do you call
it, highlights and
dark tones and whatever. This this tweet
this tweet was super interesting because
I really like
I used to really like doing like, you
know, graphics and and stuff when I was
playing Counter-Strike and I made my my
my team's web page and
um I want a lot of websites, so I really
used to love this, but I was just not to
not that good at
>> Yeah, it is it is fascinating to think
that you can drop a word and then it
will
the the will behave differently. Well,
absolutely.
Well, it makes total sense, right? I
mean, I was just having a call earlier
with a um
This is actually an interesting story.
Um we identified that in our Terraform
workspace, we have quite a few
dependencies, and sometimes cycles were
created where something depends on
something, and then that thing ends up
depending on the previous thing.
So, we wanted to visualize our our
Terraform, you know, uh with like remote
state lookups, how how this whole thing
is connected. Mhm.
>> which is what I built the grid for, if
you remember, like the grid
>> Yeah, yeah, yeah, I was going to I was
going to say that, like isn't that what
Takos is all about? Uh this is what the
grid uh a separate Takos is is for. It's
allows you to define your dependencies.
But, um we have those dependencies in
our Terraform config, and we're not
rolling out the grid at the moment due
to the fact that I'm, you know,
part-time on that project, so
Um so, I decided to to just
I I have somebody new that came in, and
um
I'm onboarding him in a very AI-native
way, asking him to do everything with
AI. And uh he built something. Um
And it's very funny, because
it the model knows an open-source tool
that maybe has like, I don't know, 100
people, 100 stars, super low,
and it has a website. And
it has a feature that we don't use,
because I forked it.
And when it builds something, it says
it start talking about this feature. And
I went to through the whole log, like is
there any web search? Like how did the
model
come to the conclusion that the feature
is there without a web search?
It can only be because it's somewhere in
its corpus, and it's inside its weights.
I don't know how many billion
parameter model this is, that it, you
know, knows that this particular open
source project has that particular
feature which you know, we don't use.
But how did it know?
You know, like you can ask it crazy
things, like do people throw cats off
towers in Belgium? And then they will go
like, yes, there's this tradition in
this city where people throw cats off
off off a tower and it's because after
the winter they had too many cats, so
they had to throw them down the tower.
It's like, how the hell does it know
this?
You know?
I think we talked about this before.
Well, this cat thing I've heard about
it. Um Yes.
Uh
Well
I
So, I
Um I'm approaching my destination soon.
The the the last um session I went to
was one by Google DeepMind.
And the one thing that I found just you
you were mentioning like an AI native
onboarding.
So, the way that they did it was like
um
okay, we're going to run through this
this example.
I for this voice API Lirium or something
like this. And they they asked the the
workshop attendees just to install a
skill.
And it worked pretty well. Like there
were people using
there were there were people using
anti-gravity, there's people using like
myself using Claude Code.
There was a surprising a lot of people
using code.
But but at the end of the day almost
everyone managed to
to create
um
a proof of concept or whatever, like a a
demo using these Google APIs just via a
skill. And I thought to myself
wow, it's amazing cuz no one wrote any
code
uh during this workshop.
But everyone managed to produce
um an app
in a way. So,
but crazy times.
This is But it's just amazing that like
the skills are really You don't need to
type anything into an editor.
Oh, I'm passing Westminster Abbey.
This is where the king gets
>> Now, there's also this um
people are talking about how amazing it
is that uh and um what's Karpathy?
He he says he talks about uh something
that he uses, a process.
And he described it and people really
liked it the idea. And then he said,
"Look, if you want to build it, here's a
gist. Give this to your coding agent and
it will build the platform for you."
And then people started raving like, "Oh
my god, he just did software
distribution through a gist.
Ah!"
Yeah, all through a all through a
prompt. Okay, I need to get off this bus
now.
Um
So, you're going to have to excuse me. I
need
>> one, right? I mean
I got to put me on mute.
>> the same thing.
Although Yeah.
The the conference venue is right next
to the Houses of Parliament. I can show
you
um Big Ben in a bit.
Got up at the wrong time. Big Ben from
the Edge of Tomorrow.
You know Edge of Tomorrow?
And dropped.
So, Edge of Tomorrow is a film with Tom
Cruise where he is fighting an alien
invasion and they are actually
at the part of the war where the humans
lose and they lose by taking London. So,
there's a quite a few
scenes of aliens invading London.
And the plot of the movie is that
every time one of the Well, the agents
have to keep The aliens have a
capability of certain aliens resetting
the time if they get killed, allowing
a replay. It's kind of like a game.
And for some reason
one the main actor
gets that skill. So, every time he dies,
he resets
to the previous day
whenever he acquired the skill of being
able to reset.
And why I watched it six times is
because it's exactly like a computer
game, you know, when you
keep getting keep dying and then keep
getting reset
to until you execute the level perfectly
because you can predict things that are
going to happen.
Uh wait.
Are you just talking by yourself?
Yeah.
You're you're like
I That's another thing that that's a bit
of a trend.
Oh, oh, there's there's a married couple
here.
Uh you mean
>> Uh we're we're we're basically like
before before a recorded AI talk, like
you can go like
you saw that tweet. I think you shared
it to me like
and yeah
and with my
the sales figures went 3x and then
or something crazy like you make some
crazy statements
and then people join the call and then
when it summarizes the headline is
something insane like And then
you talked you
you talked about how the aliens landed
and uh
and then then he switched to uh you
know, some financials. Oh, yeah, you can
you see Big Ben over there? Yeah, I see
it. Nice. So so
Yeah, this is my start.
So, I'm assuming that when you were
when I
when I got disconnected there, you were
just just talking about some wild
things.
No, it was not too wild. It was related
to the fact that I only saw Big Well, I
don't I've not Maybe I've seen Big Ben
in real life, but I was I most memorably
for me is the scenes in the movie Edge
of Tomorrow where Tom Cruise is um
fighting
>> a good invasion. Is that the one he dies
all the time, right?
>> Yeah, he dies and resets. So, I was just
explaining that that aliens are
fighting.
The female lead was great, too. She's uh
South African, isn't she? Or is she also
No, no, that that No, she's not. So, not
that one.
Yeah, I really liked the the cuz I was
doing yoga at the time and then there's
a scene in the in the movie when she's
doing yoga and I thought, "Yeah, yoga,
baby."
She's doing exercises in this training
hall where um it's like
a simulation of They're in these huge
suits to fight, right? And so, have to
get used to it.
>> it's super cool movie.
It's super cool movie. I love it. Yeah,
great movie.
Um okay, well
>> about something else.
I'm talking something about Ah, yeah, uh
software distribution through a gist.
But, it's it's not Karpathy who did this
first, right? I mean, Minions from
Stripe was also like just point them at
this blog and they'll build it
um type of distribution.
Yeah.
It's pretty cool. Yeah.
But
I could I could
I couldn't help but think that like I
think we had a blip with the internet or
pulled um Anthropic was was uh having
some disruptions Yeah, they do they
>> in the day yesterday. Yeah, yeah. And I
couldn't help but think to myself like
you know
we we we we we think how cool it is that
you can build something with uh a
prompt.
But, of course, if you don't have that
infrastructure in place
then
But, that's why
>> don't have
I I like just imagine the future where
this mythos is amazing and it's able to
do our work
and I'm like happily using it like
probably most people are
but the minute it it's it gets taken
away from me
it's like
it's going to be crazy.
Yeah, but but at the same time I mean we
talked about this before, right? That's
why people that are like doom prepping
for the time that those AI inference
providers whatever due to war or
whatever are no longer able to provide
the service cuz I've been in countries
where the service is not active and it's
really painful.
Um, that's why and now with Gemma from
from Google people are I mean I'm
reading very very conflicting
evaluations of it. People are like, "Oh
my god, Gemma is really worse than
Gwen."
Uh, and some people are like, "Once you
fine-tune it and you you know how to set
it up properly, Gemma is amazing." What
about Gemma? Like is it like an open
Open weights model, yeah. Like it's on
on hugging face.
Um, so you can
run it.
Uh, I'm waiting for a a big paycheck and
and buying like an M5. You know, Mac Mac
is people always say like Apple was
dumb, they don't have Apple intelligent
like their Apple intelligence projects
failed, but they really built great
hardware where they they make the
bandwidth between the memory and the CPU
so good that Yeah, yeah. We were joking
there was some there was some ex-Apple
people in the in our telling to over
lunch.
And uh, they were sort of joking that
like Apple is so bad at AI, but they
have the best hardware.
Yeah. And uh,
and right now their strategy seems to be
uh, you know, pulling pulling along
either Google or uh, open AI
to get the best
open um, to get the best model without
them actually having to do anything.
Cuz I guess they have the cash or
something.
Cuz they they haven't actually fully
committed to using the Google
model or or open AI model in their in
their products
allegedly.
So people say it's because Apple has
such a high reputation to uphold. But I
mean that that if like because the AI
can fail so spectacularly um because of
it being non-deterministic and it could
result in the product being in a very
bad
daylight like where it fails.
Yeah, that's that's why they haven't
integrated it but He he mentioned it uh
one of the ex-Apple people said like,
you know, they own the the interface or
they own the platform so they are like
in the best position, right?
And we we just observed inside
the lunch hall, you know, most people
are using
most people are using iPhones, right?
Okay, I got to get This is This is the
thumbnail here. This is the AI engineer
Europe
thumbnails.
They had like a huge lobster out here
yesterday.
Yeah, you sent me a picture.
We put it on I don't know why I I don't
know why it isn't here today.
Possibly something
Possibly. I don't even like
I'm like, how do they get these huge
lobsters? Is someone just traveling
around with a [ __ ] with a [ __ ]
But
But how did how did they get a lobster
here?
These buildings all around here are just
magnificent. I love them.
It looks like Ghent. I mean, if you go
to city center Ghent, you have these
massive um
like cathedrals and and and wide market
places very similar to this.
So hey, do you have a Riverside account
because ideally I I could end this
recording and you could you could upload
the
or finish editing these three clips.
Cuz I'm going to have an absolutely no
time today.
I have a Riverside account. I even have
pro for 2 months for free.
And uh
Maybe I'll try invite you. Or something.
I don't know. I don't know if it works
if two people have different accounts.
You mentioned that if I want to edit it,
then it needs to be with a Can you just
send me your email or your login? Cuz I
tried it last time and it said like
this.
Mhm. Okay, I'm going to go in here and
and volunteer.
Uh I'm going to be doing crowd control
this morning. This is the way I scored
entrance.
Anyway,
it's going to be a great day. Bye. All
right. Bye.
I