WEBVTT

00:00:01.240 --> 00:00:02.720
I've been some

00:00:02.720 --> 00:00:04.760
I'm coming for you from a bus stop. And

00:00:04.760 --> 00:00:08.520
the advertising Singapore here um

00:00:08.520 --> 00:00:12.120
Funny enough, that's not my bus.

00:00:12.120 --> 00:00:14.120
And um

00:00:14.120 --> 00:00:15.760
Yeah, this is London. This is what

00:00:15.760 --> 00:00:17.520
London looks like.

00:00:17.520 --> 00:00:19.480
Can you hear me clearly? Yes, I can hear

00:00:19.480 --> 00:00:21.000
you. They are advertising Singapore and

00:00:21.000 --> 00:00:22.400
this is London.

00:00:22.400 --> 00:00:23.240
Yeah.

00:00:23.240 --> 00:00:25.040
>> or what? I thought like you had some

00:00:25.040 --> 00:00:27.360
super deluxe bus stop area where you

00:00:27.360 --> 00:00:28.520
could sit with a laptop but you're

00:00:28.520 --> 00:00:31.440
calling from the phone. No, no. To be

00:00:31.440 --> 00:00:32.599
To be honest, the fact that I have my

00:00:32.599 --> 00:00:36.040
phone out and the cyclist bugging me by

00:00:36.040 --> 00:00:38.600
it shows you I have nerves of steel.

00:00:38.600 --> 00:00:41.080
Because it's quite common to get your

00:00:41.080 --> 00:00:42.680
phone nicked here.

00:00:42.680 --> 00:00:45.360
Yeah, well, that that was Saigon 5 years

00:00:45.360 --> 00:00:47.000
ago. I don't know. I mean, you guys

00:00:47.000 --> 00:00:49.080
>> What did they do? Stop the crime in

00:00:49.080 --> 00:00:50.640
Saigon?

00:00:50.640 --> 00:00:52.120
Cut people's hands off?

00:00:52.120 --> 00:00:54.040
Uh well,

00:00:54.040 --> 00:00:55.560
I don't know. I'm not in District 1

00:00:55.560 --> 00:00:57.440
anymore, but like

00:00:57.440 --> 00:00:59.160
it's surprising to hear that London is

00:00:59.160 --> 00:01:01.320
like that. Belgium wasn't like that in

00:01:01.320 --> 00:01:03.760
2006 when I left.

00:01:03.760 --> 00:01:06.280
To hear Europe being in a situation that

00:01:06.280 --> 00:01:08.240
Saigon, Vietnam, a third world country

00:01:08.240 --> 00:01:09.600
is in Yeah.

00:01:09.600 --> 00:01:13.040
>> 6 years ago. That's the thing that my

00:01:13.040 --> 00:01:15.440
wife and I joke about because

00:01:15.440 --> 00:01:17.160
like for example, to

00:01:17.160 --> 00:01:18.840
our experience with building stuff here

00:01:18.840 --> 00:01:21.320
in England is like third world.

00:01:21.320 --> 00:01:24.320
But mentality like

00:01:24.320 --> 00:01:25.760
you have to pay someone, you have to pay

00:01:25.760 --> 00:01:27.560
someone, you have to pay someone. It's

00:01:27.560 --> 00:01:29.120
just pay someone for that. You can't

00:01:29.120 --> 00:01:31.920
actually build.

00:01:31.920 --> 00:01:34.200
Yeah, so

00:01:34.200 --> 00:01:36.800
Yeah. And that's true, actually. All

00:01:36.800 --> 00:01:38.400
right, so I just wanted to say that the

00:01:38.400 --> 00:01:39.920
conference yesterday was great and I'm

00:01:39.920 --> 00:01:42.080
going to It's only 2 more days of it.

00:01:42.080 --> 00:01:44.080
It's incredible.

00:01:44.080 --> 00:01:46.640
So I'm looking forward to that. And uh 4

00:01:46.640 --> 00:01:48.840
days? Yeah.

00:01:48.840 --> 00:01:50.040
I went to an

00:01:50.040 --> 00:01:51.760
I went to an anthropic talk and I went

00:01:51.760 --> 00:01:52.840
to

00:01:52.840 --> 00:01:53.880
They just made a whole bunch of

00:01:53.880 --> 00:01:55.360
announcements at Anthropic.

00:01:55.360 --> 00:01:57.840
>> Yeah. And

00:01:57.840 --> 00:02:00.000
It's insane. Um

00:02:00.000 --> 00:02:01.920
The It was a workshop, but it

00:02:01.920 --> 00:02:03.400
It was more like a presentation. They

00:02:03.400 --> 00:02:05.080
would like to They were basically going

00:02:05.080 --> 00:02:07.080
through that that blog that I think we

00:02:07.080 --> 00:02:08.960
we kind of

00:02:08.960 --> 00:02:10.759
might have mentioned to each other.

00:02:10.759 --> 00:02:12.800
The one about

00:02:12.800 --> 00:02:14.320
the one about the harness, the one about

00:02:14.320 --> 00:02:15.680
the

00:02:15.680 --> 00:02:19.520
the agents, adversarial agents, and the

00:02:19.520 --> 00:02:21.360
evaluators, and things like that. Oh,

00:02:21.360 --> 00:02:24.720
yeah. Yeah, I adversarial agents is the

00:02:24.720 --> 00:02:26.640
keyword now. I think I called it

00:02:26.640 --> 00:02:29.080
adversarial, but it's adversarial, of

00:02:29.080 --> 00:02:31.200
course, like I would

00:02:31.200 --> 00:02:32.320
The other word that they kept on

00:02:32.320 --> 00:02:34.840
dropping like it was hot was rubric

00:02:34.840 --> 00:02:37.120
rubric, which I think I had to look it

00:02:37.120 --> 00:02:38.200
up,

00:02:38.200 --> 00:02:40.440
which means

00:02:40.440 --> 00:02:43.400
instruction. They put it rubric

00:02:43.400 --> 00:02:44.760
for some reason. Have you ever heard

00:02:44.760 --> 00:02:49.640
this expression or term?

00:02:49.640 --> 00:02:52.360
No. What's What is it?

00:02:52.360 --> 00:02:53.959
Rubric.

00:02:53.959 --> 00:02:55.440
Rubric?

00:02:55.440 --> 00:02:57.200
Mhm.

00:02:57.200 --> 00:02:59.720
Not sure what that is.

00:02:59.720 --> 00:03:02.200
Um

00:03:02.200 --> 00:03:03.239
But I

00:03:03.239 --> 00:03:04.920
I thought

00:03:04.920 --> 00:03:07.440
I I have a I have a good idea how to you

00:03:07.440 --> 00:03:09.560
know, create my own harness. If I have I

00:03:09.560 --> 00:03:12.000
have like a good start about that.

00:03:12.000 --> 00:03:15.040
So, I I have a good idea about creating

00:03:15.040 --> 00:03:17.280
something that can spend $100 and come

00:03:17.280 --> 00:03:19.200
up with a you know,

00:03:19.200 --> 00:03:21.400
a richer game maker or some other some

00:03:21.400 --> 00:03:23.360
other thing. So,

00:03:23.360 --> 00:03:25.400
so they they equipped us how to spend a

00:03:25.400 --> 00:03:28.280
chunk of change, and then of course,

00:03:28.280 --> 00:03:29.920
I mentioned like, "Oh, what happens if

00:03:29.920 --> 00:03:31.040
it's going wrong?" It's like, "Oh, you

00:03:31.040 --> 00:03:34.320
have trust issues."

00:03:34.320 --> 00:03:36.360
It sounds like your harness or your

00:03:36.360 --> 00:03:37.680
evaluators aren't good enough or

00:03:37.680 --> 00:03:38.920
something.

00:03:38.920 --> 00:03:40.880
Okay, this is my boss. And you said,

00:03:40.880 --> 00:03:42.640
"No, it's Claude Code that gave me trust

00:03:42.640 --> 00:03:46.600
issues."

00:03:46.600 --> 00:04:07.840
Well, I'm going to mute myself.

00:04:07.840 --> 00:04:09.440
>> Yeah, that's

00:04:09.440 --> 00:04:10.840
They're in the bus now?

00:04:10.840 --> 00:04:12.560
I'm in the double decker

00:04:12.560 --> 00:04:16.799
Kikikik double decker.

00:04:16.799 --> 00:04:18.600
But busy. I don't know what the empty is

00:04:18.600 --> 00:04:25.400
down over there.

00:04:25.400 --> 00:04:26.640
God, I'm going to be that guy on the bus

00:04:26.640 --> 00:04:28.080
talking.

00:04:28.080 --> 00:04:32.560
Hate those people.

00:04:32.560 --> 00:04:35.440
Uh

00:04:35.440 --> 00:04:37.080
And the the other talk that was really

00:04:37.080 --> 00:04:38.480
interesting was

00:04:38.480 --> 00:04:40.360
by Pedantic. I don't know if you've ever

00:04:40.360 --> 00:04:45.160
used Pedantic.

00:04:45.160 --> 00:04:47.800
Pedantic is a Python unit testing

00:04:47.800 --> 00:04:49.680
library or what?

00:04:49.680 --> 00:04:51.440
Or is it Um

00:04:51.440 --> 00:04:53.720
behavior driven behavior testing I know

00:04:53.720 --> 00:04:56.600
it is like a schema manager, like a

00:04:56.600 --> 00:04:59.440
Oh. like a model manager.

00:04:59.440 --> 00:05:01.000
Why is my face like this? Like they're

00:05:01.000 --> 00:05:03.680
super

00:05:03.680 --> 00:05:04.960
It's like somebody smashed me in the

00:05:04.960 --> 00:05:08.760
face and there's

00:05:08.760 --> 00:05:11.840
Yeah, so so that's pretty cool. I mean

00:05:11.840 --> 00:05:13.800
I thought you were going to Yeah, so you

00:05:13.800 --> 00:05:16.160
did the workshop. What else did you, you

00:05:16.160 --> 00:05:18.680
know, pick up from the hallway track or

00:05:18.680 --> 00:05:26.520
what they call it?

00:05:26.520 --> 00:05:28.320
I can hear you, but

00:05:28.320 --> 00:05:30.440
it dropped.

00:05:30.440 --> 00:05:34.080
Well, then

00:05:34.080 --> 00:05:35.640
while

00:05:35.640 --> 00:05:37.880
we're waiting for Kai to get back, let's

00:05:37.880 --> 00:05:39.400
talk about

00:05:39.400 --> 00:05:43.240
amazing Project Glass Wing.

00:05:43.240 --> 00:05:45.200
I mispronounced it.

00:05:45.200 --> 00:05:47.680
Became embarrassing.

00:05:47.680 --> 00:05:50.360
Glass being a project that Anthropic

00:05:50.360 --> 00:05:53.120
announced with their latest Is it

00:05:53.120 --> 00:05:55.880
Mythos?

00:05:55.880 --> 00:05:57.760
model

00:05:57.760 --> 00:06:00.280
uh basically

00:06:00.280 --> 00:06:03.760
identifying so many vulnerabilities.

00:06:03.760 --> 00:06:06.080
Yeah, I'm back. I'm back. I was just

00:06:06.080 --> 00:06:07.640
talking about project Glass wing and

00:06:07.640 --> 00:06:10.720
Mythos and is it Mythos, right? The new

00:06:10.720 --> 00:06:11.560
uh

00:06:11.560 --> 00:06:12.720
Mhm.

00:06:12.720 --> 00:06:14.840
model that Anthropic does not want to

00:06:14.840 --> 00:06:17.280
release until the internet has had the

00:06:17.280 --> 00:06:20.440
time to prepare.

00:06:20.440 --> 00:06:21.960
Yeah, there's there's definitely some

00:06:21.960 --> 00:06:24.600
hype about it and and the uh

00:06:24.600 --> 00:06:26.760
And during the Anthropic

00:06:26.760 --> 00:06:27.919
Well,

00:06:27.919 --> 00:06:31.880
the Anthropic um

00:06:31.880 --> 00:06:33.919
Say that Like the crazy thing about the

00:06:33.919 --> 00:06:35.680
Anthropic talk was that he was like

00:06:35.680 --> 00:06:37.160
talking about the

00:06:37.160 --> 00:06:39.760
the harnesses and the evaluator and the

00:06:39.760 --> 00:06:41.680
and the prompts that you would do. But

00:06:41.680 --> 00:06:43.640
then he warned that when Mythos comes

00:06:43.640 --> 00:06:44.960
out, the

00:06:44.960 --> 00:06:47.120
your harness would have to be updated

00:06:47.120 --> 00:06:48.840
because

00:06:48.840 --> 00:06:50.760
because you know, Mythos is a lot

00:06:50.760 --> 00:06:52.360
better.

00:06:52.360 --> 00:06:53.960
So, your harness would have to like

00:06:53.960 --> 00:06:55.360
focus

00:06:55.360 --> 00:06:58.400
on other things.

00:06:58.400 --> 00:06:59.240
You know what I mean?

00:06:59.240 --> 00:07:01.080
>> Yeah. It's like it's like, hey, we're

00:07:01.080 --> 00:07:03.360
doing all this work but like once this

00:07:03.360 --> 00:07:05.040
new model comes out, we'll be able to do

00:07:05.040 --> 00:07:06.760
things better. So, you just have to

00:07:06.760 --> 00:07:08.040
like, you know,

00:07:08.040 --> 00:07:10.080
push yourself down the line a bit.

00:07:10.080 --> 00:07:11.880
Yeah, I mean, there's this interesting

00:07:11.880 --> 00:07:14.720
talk I think Is it Simon Wilson on his

00:07:14.720 --> 00:07:16.320
interview but I think somebody else as

00:07:16.320 --> 00:07:18.880
well. They talked like our role right

00:07:18.880 --> 00:07:21.200
now is to

00:07:21.200 --> 00:07:23.919
Simon Wilson Willison Wilson I don't

00:07:23.919 --> 00:07:25.520
know how to pronounce his last name. He

00:07:25.520 --> 00:07:29.600
definitely talks about keeping uh

00:07:29.600 --> 00:07:33.440
a backlog of projects that you you try

00:07:33.440 --> 00:07:35.560
and then you see the AI cannot do and

00:07:35.560 --> 00:07:37.080
you put it on the backlog. Like you say

00:07:37.080 --> 00:07:39.000
this one day we'll be able to do it and

00:07:39.000 --> 00:07:41.360
just keep going back and trying out

00:07:41.360 --> 00:07:44.480
those projects. So, do keep feeling out

00:07:44.480 --> 00:07:46.520
like the boundaries as they are moving

00:07:46.520 --> 00:07:48.280
as models are getting better, like where

00:07:48.280 --> 00:07:50.720
are we now? Where How far can we go? And

00:07:50.720 --> 00:07:52.560
to keep living on that like they call it

00:07:52.560 --> 00:07:54.000
the border. There was I saw two

00:07:54.000 --> 00:07:56.520
different talks YouTubes YouTube

00:07:56.520 --> 00:07:58.600
presentations about um

00:07:58.600 --> 00:08:00.680
what we can do because I think the the

00:08:00.680 --> 00:08:02.480
biggest problem is

00:08:02.480 --> 00:08:04.640
which I think we mentioned before is

00:08:04.640 --> 00:08:06.200
not realizing

00:08:06.200 --> 00:08:08.480
what the model can already do and still

00:08:08.480 --> 00:08:09.760
doing things that we should just hand

00:08:09.760 --> 00:08:11.600
off to the model. And I

00:08:11.600 --> 00:08:13.720
think

00:08:13.720 --> 00:08:16.040
Cuz like Anthropic announced two things,

00:08:16.040 --> 00:08:18.360
right? They announced this um

00:08:18.360 --> 00:08:19.800
three things. They announced the project

00:08:19.800 --> 00:08:22.680
last week and the impact of of that cap-

00:08:22.680 --> 00:08:25.840
capabilities detecting decade-olds um

00:08:25.840 --> 00:08:27.760
zero days inside foundational

00:08:27.760 --> 00:08:29.840
infrastructure in the internet would,

00:08:29.840 --> 00:08:32.120
you know, create huge impact on the

00:08:32.120 --> 00:08:34.440
banking systems and the trust that we

00:08:34.440 --> 00:08:37.400
can have on on on I mean, even if it's

00:08:37.400 --> 00:08:39.280
marketing from Anthropic, which I don't

00:08:39.280 --> 00:08:41.320
think it is, but you may assume it is.

00:08:41.320 --> 00:08:43.800
Um it is a fact that that these these

00:08:43.800 --> 00:08:46.000
models like Opus, they're able to detect

00:08:46.000 --> 00:08:47.839
bugs and then

00:08:47.839 --> 00:08:50.720
pra- practically zero days, right? And

00:08:50.720 --> 00:08:53.160
if the bug bounty is not being awarded

00:08:53.160 --> 00:08:55.800
properly, um these get sold on the black

00:08:55.800 --> 00:08:58.800
market or state actors and hackers abuse

00:08:58.800 --> 00:09:01.480
these to, you know, have real

00:09:01.480 --> 00:09:02.600
significant impact.

00:09:02.600 --> 00:09:04.720
>> Yeah. I was I was had a thought that

00:09:04.720 --> 00:09:05.720
like

00:09:05.720 --> 00:09:07.920
AI agents

00:09:07.920 --> 00:09:10.640
you can you can effectively

00:09:10.640 --> 00:09:13.720
organize botnets of agents nowadays cuz

00:09:13.720 --> 00:09:15.680
that's all the rage, right? I think it

00:09:15.680 --> 00:09:17.000
was quite expensive, but that's what

00:09:17.000 --> 00:09:18.280
Cloud Botnet is, right? I mean, the

00:09:18.280 --> 00:09:22.200
whole Cloud Book was basically a botnet,

00:09:22.200 --> 00:09:24.600
um you know, directed by people to

00:09:24.600 --> 00:09:25.880
simulate

00:09:25.880 --> 00:09:28.080
um all kinds of stories. And then people

00:09:28.080 --> 00:09:29.320
are like, "Oh my god, these agents are

00:09:29.320 --> 00:09:31.480
planning a revolution. They're designing

00:09:31.480 --> 00:09:33.080
their own language." Yeah, yeah, yeah,

00:09:33.080 --> 00:09:35.400
okay. Um but but that was one of the

00:09:35.400 --> 00:09:37.240
announcements, right? Basically, we're

00:09:37.240 --> 00:09:38.240
reaching a

00:09:38.240 --> 00:09:41.680
a level of of of intelligence that that

00:09:41.680 --> 00:09:44.800
it's rare amongst humans, and humans

00:09:44.800 --> 00:09:47.680
have that have that expert level

00:09:47.680 --> 00:09:49.760
capabilities are the ones that are able

00:09:49.760 --> 00:09:51.560
to basically hack into

00:09:51.560 --> 00:09:54.080
software, find zero-days,

00:09:54.080 --> 00:09:56.280
and I mean

00:09:56.280 --> 00:09:58.000
I really enjoyed reading Phineas

00:09:58.000 --> 00:10:00.400
Hacker's breakdown of how he hacks the

00:10:00.400 --> 00:10:02.720
hacking team and how he built like a

00:10:02.720 --> 00:10:05.520
zero-day on a router that he knew uh

00:10:05.520 --> 00:10:08.280
like a a different an um

00:10:08.280 --> 00:10:09.880
type of router device that they used,

00:10:09.880 --> 00:10:12.640
and he he tried it his exploit against

00:10:12.640 --> 00:10:14.320
several other routers on the internet

00:10:14.320 --> 00:10:16.160
before he actually

00:10:16.160 --> 00:10:18.080
attacked the target. So, this type of

00:10:18.080 --> 00:10:19.520
intelligence

00:10:19.520 --> 00:10:21.120
there are humans that have it, and they

00:10:21.120 --> 00:10:24.000
are able to do it clearly.

00:10:24.000 --> 00:10:26.080
And if you put that in the wrong hands

00:10:26.080 --> 00:10:27.600
It makes me think

00:10:27.600 --> 00:10:28.720
Cloud

00:10:28.720 --> 00:10:30.280
Cloudflare is going to do well because

00:10:30.280 --> 00:10:31.560
they have

00:10:31.560 --> 00:10:33.800
all these security products to

00:10:33.800 --> 00:10:36.320
stop agents to the moment.

00:10:36.320 --> 00:10:37.560
It's not a bad observation, yeah. I

00:10:37.560 --> 00:10:40.480
mean, any any organization that

00:10:40.480 --> 00:10:42.520
that is I mean, Cloudflare is known for

00:10:42.520 --> 00:10:45.000
their extremely good

00:10:45.000 --> 00:10:46.960
DDoS protection and so on. Don't get

00:10:46.960 --> 00:10:49.160
started Don't get me started on their UI

00:10:49.160 --> 00:10:50.560
skills. They're the

00:10:50.560 --> 00:10:51.960
worst.

00:10:51.960 --> 00:10:52.640
Um

00:10:52.640 --> 00:10:54.760
but they anyway, the second announcement

00:10:54.760 --> 00:10:58.360
was the a product for agent like I don't

00:10:58.360 --> 00:10:59.600
know, I haven't looked too much into it,

00:10:59.600 --> 00:11:02.560
but it sounds like hosted agent teams.

00:11:02.560 --> 00:11:06.000
So, you can run or build agent teams.

00:11:06.000 --> 00:11:07.120
Now,

00:11:07.120 --> 00:11:09.920
um or you can just have an

00:11:09.920 --> 00:11:12.360
Anthropic hosted agent teams, which is

00:11:12.360 --> 00:11:14.960
something really super interesting.

00:11:14.960 --> 00:11:16.800
I think he I think I think the guy

00:11:16.800 --> 00:11:18.280
mentioned it yesterday, and they

00:11:18.280 --> 00:11:20.280
coordinate using the file system, I

00:11:20.280 --> 00:11:22.280
think he was saying.

00:11:22.280 --> 00:11:23.680
Um

00:11:23.680 --> 00:11:25.320
There's There's a lot going on with

00:11:25.320 --> 00:11:27.160
Anthropic. May-

00:11:27.160 --> 00:11:29.080
L- I mean, there's quite a few startups

00:11:29.080 --> 00:11:31.000
that are going and saying like, "Okay,

00:11:31.000 --> 00:11:33.240
you know, Stripe published a markdown

00:11:33.240 --> 00:11:34.840
saying that this is our agent minion

00:11:34.840 --> 00:11:36.760
agent system. Uh we're not going to open

00:11:36.760 --> 00:11:37.840
source this, but if you take this

00:11:37.840 --> 00:11:39.720
markdown, you can build your own.

00:11:39.720 --> 00:11:40.800
Um and so I think there's a lot of

00:11:40.800 --> 00:11:42.760
startups that that were built around

00:11:42.760 --> 00:11:44.800
that saying like, oh, you know, Stripe

00:11:44.800 --> 00:11:46.640
can build it, but not a lot of like

00:11:46.640 --> 00:11:48.400
medium-sized companies or small startups

00:11:48.400 --> 00:11:49.840
cannot build it. So there's definitely a

00:11:49.840 --> 00:11:52.040
need for a SaaS that provides this. And

00:11:52.040 --> 00:11:53.600
I believe that like if you're going to

00:11:53.600 --> 00:11:55.640
pay Anthropic for the module, then I

00:11:55.640 --> 00:11:57.240
might as well use Anthropic for the

00:11:57.240 --> 00:11:58.920
agent team if they provide a service,

00:11:58.920 --> 00:12:00.880
right? Why would you need to pay some

00:12:00.880 --> 00:12:03.360
third-party startup

00:12:03.360 --> 00:12:06.600
um to pay to to do agent orchestration.

00:12:06.600 --> 00:12:07.960
Uh but there's

00:12:07.960 --> 00:12:09.720
discussion that I I had a lot with my

00:12:09.720 --> 00:12:11.240
friend because I'm talking about hey,

00:12:11.240 --> 00:12:14.000
this spectrum design is really good.

00:12:14.000 --> 00:12:14.600
Um

00:12:14.600 --> 00:12:17.160
but they're just going like

00:12:17.160 --> 00:12:18.960
now you just need a you just need a

00:12:18.960 --> 00:12:20.960
single skill. You just need a brainstorm

00:12:20.960 --> 00:12:24.480
skill and Opus and um you don't need to

00:12:24.480 --> 00:12:26.120
go through this whole spectrum design

00:12:26.120 --> 00:12:28.720
workflow anymore. Um

00:12:28.720 --> 00:12:31.000
you get like an interview experience

00:12:31.000 --> 00:12:32.800
where it asks you, what about this? What

00:12:32.800 --> 00:12:34.680
about that? Just a few key points that

00:12:34.680 --> 00:12:36.600
you want to clarify and then it goes off

00:12:36.600 --> 00:12:38.360
and builds it. Like you don't need all

00:12:38.360 --> 00:12:40.160
of that additional stuff anymore and it

00:12:40.160 --> 00:12:41.480
I think we're definitely going towards

00:12:41.480 --> 00:12:41.800
that.

00:12:41.800 --> 00:12:44.160
>> Have you Do you Do you know what I talk

00:12:44.160 --> 00:12:45.320
Do you know what I mean when I have to

00:12:45.320 --> 00:12:47.640
share the slides with you? But do you

00:12:47.640 --> 00:12:51.600
know what I mean by by um

00:12:51.600 --> 00:12:54.160
an evaluator and an adversarial agent

00:12:54.160 --> 00:12:55.839
because

00:12:55.839 --> 00:12:59.200
And I know you since you do spec dev or

00:12:59.200 --> 00:13:00.360
and spec dev

00:13:00.360 --> 00:13:02.000
spectrum development

00:13:02.000 --> 00:13:03.360
the the way that they explained it is

00:13:03.360 --> 00:13:05.680
that the plan is very high level

00:13:05.680 --> 00:13:07.400
and then the emphasis goes in the

00:13:07.400 --> 00:13:09.480
evaluator. So basically in the evaluator

00:13:09.480 --> 00:13:10.640
stage

00:13:10.640 --> 00:13:12.960
you put the effort

00:13:12.960 --> 00:13:14.960
I mean, this is my interpretation.

00:13:14.960 --> 00:13:17.440
But my interpretation is that like

00:13:17.440 --> 00:13:19.760
your evaluator is basically where your

00:13:19.760 --> 00:13:22.560
spec now lives and because

00:13:22.560 --> 00:13:25.200
you're you're basically at the role

00:13:25.200 --> 00:13:27.000
where

00:13:27.000 --> 00:13:29.720
the product owner is basically or the

00:13:29.720 --> 00:13:33.840
the client or something is is basically

00:13:33.840 --> 00:13:36.440
evaluating what was built.

00:13:36.440 --> 00:13:39.080
So, now you have to

00:13:39.080 --> 00:13:41.240
uh make sure the evaluator has I guess

00:13:41.240 --> 00:13:43.520
your requirements and your expectations

00:13:43.520 --> 00:13:44.800
all there

00:13:44.800 --> 00:13:47.400
so that it can evaluate

00:13:47.400 --> 00:13:50.440
the the build uh stage.

00:13:50.440 --> 00:13:52.120
Does that make sense? Yeah, makes sense.

00:13:52.120 --> 00:13:53.240
I mean

00:13:53.240 --> 00:13:54.680
I didn't I I hadn't thought that far

00:13:54.680 --> 00:13:56.960
because because I always I use

00:13:56.960 --> 00:13:59.080
adversarial agents a lot now. I really

00:13:59.080 --> 00:14:00.120
love them.

00:14:00.120 --> 00:14:04.000
Um but it does mean I I can do it very

00:14:04.000 --> 00:14:05.760
effectively because I spent so much time

00:14:05.760 --> 00:14:08.080
planning. So, because I spent so much

00:14:08.080 --> 00:14:09.520
time planning

00:14:09.520 --> 00:14:11.760
I can tell the agent now verify that

00:14:11.760 --> 00:14:14.280
whatever we build matches and every

00:14:14.280 --> 00:14:16.120
single time there's something. There's

00:14:16.120 --> 00:14:18.040
like it most most of the time it just

00:14:18.040 --> 00:14:20.200
skips the tests. It just says, "Oh, this

00:14:20.200 --> 00:14:22.040
no needed for this test. Uh that one

00:14:22.040 --> 00:14:24.160
would take too long." Like last time it

00:14:24.160 --> 00:14:25.880
told me I'm not writing I'm not writing

00:14:25.880 --> 00:14:27.600
the test because it wouldn't be suitable

00:14:27.600 --> 00:14:28.960
for a unit test because it would take

00:14:28.960 --> 00:14:30.800
too long to run because it run it forks

00:14:30.800 --> 00:14:33.040
out to get clone. I'm as like, "Yeah,

00:14:33.040 --> 00:14:34.440
but I would have liked to known that,

00:14:34.440 --> 00:14:35.880
you know, that you are not writing those

00:14:35.880 --> 00:14:37.360
tests because of that." Cuz they're all

00:14:37.360 --> 00:14:39.680
just missing. So, the adversarial re uh

00:14:39.680 --> 00:14:42.160
review always surfaces these stuff. It's

00:14:42.160 --> 00:14:44.400
always super interesting. So,

00:14:44.400 --> 00:14:45.880
but it does require like you say an

00:14:45.880 --> 00:14:48.560
evaluator to be very much aware about

00:14:48.560 --> 00:14:51.080
exactly what do you want. Otherwise

00:14:51.080 --> 00:14:52.680
what can an like

00:14:52.680 --> 00:14:53.800
like

00:14:53.800 --> 00:14:57.080
uh what can an adversarial review give

00:14:57.080 --> 00:14:59.160
you if you don't have a clear definition

00:14:59.160 --> 00:15:00.800
of what you expect versus what you got,

00:15:00.800 --> 00:15:01.360
right?

00:15:01.360 --> 00:15:03.920
Yeah, exactly. Exactly. So,

00:15:03.920 --> 00:15:06.560
oh, the the other tip that Ash dropped

00:15:06.560 --> 00:15:08.720
which I think had a chuckle in the crowd

00:15:08.720 --> 00:15:10.080
was uh

00:15:10.080 --> 00:15:13.560
he was saying that when you write your

00:15:13.560 --> 00:15:14.440
um

00:15:14.440 --> 00:15:16.200
write your what do you call it? Your

00:15:16.200 --> 00:15:19.440
specs or what you want or your tips

00:15:19.440 --> 00:15:21.839
or your breadcrumbs

00:15:21.839 --> 00:15:24.480
in JSON because for some reason he was

00:15:24.480 --> 00:15:25.600
saying that the

00:15:25.600 --> 00:15:27.960
that the the agent will

00:15:27.960 --> 00:15:31.400
or less likely override the JSON.

00:15:31.400 --> 00:15:34.320
And

00:15:34.320 --> 00:15:36.080
I thought that was a weird but That is

00:15:36.080 --> 00:15:39.000
weird cuz like I don't know with me I

00:15:39.000 --> 00:15:41.320
want to ignore the JSON.

00:15:41.320 --> 00:15:43.280
>> I want the agent and me to own the the

00:15:43.280 --> 00:15:45.440
markdown together. I like the agent is

00:15:45.440 --> 00:15:47.160
constantly rewriting it and I'm always

00:15:47.160 --> 00:15:48.640
really careful looking at how it

00:15:48.640 --> 00:15:51.240
rewrites it based on my instructions. So

00:15:51.240 --> 00:15:53.680
for me markdown is pretty good um

00:15:53.680 --> 00:15:54.080
but maybe

00:15:54.080 --> 00:15:56.280
>> Yeah yeah like nothing I could

00:15:56.280 --> 00:15:58.600
catch with a uh Yeah I think it would be

00:15:58.600 --> 00:16:00.280
much harder for me to review the diff if

00:16:00.280 --> 00:16:02.200
it was JSON.

00:16:02.200 --> 00:16:03.760
Yeah.

00:16:03.760 --> 00:16:05.960
So I don't know that's a bit weird step.

00:16:05.960 --> 00:16:07.440
Let me let me talk about the Pedantic

00:16:07.440 --> 00:16:08.320
one which I thought was really

00:16:08.320 --> 00:16:10.560
interesting. So there's a CEO of

00:16:10.560 --> 00:16:13.640
Pedantic Samuel I can't remember his

00:16:13.640 --> 00:16:15.440
surname.

00:16:15.440 --> 00:16:17.360
I love to see the guys like this. He was

00:16:17.360 --> 00:16:19.400
very

00:16:19.400 --> 00:16:21.320
kind of inspiring. He's like

00:16:21.320 --> 00:16:25.120
the alpha geek in my opinion you know.

00:16:25.120 --> 00:16:27.720
He's the CEO of a company.

00:16:27.720 --> 00:16:31.440
I think Pedantic company's it's got

00:16:31.440 --> 00:16:32.960
I don't know how if it's doing well but

00:16:32.960 --> 00:16:33.920
like

00:16:33.920 --> 00:16:36.880
it's great to see the CEO like

00:16:36.880 --> 00:16:39.000
obviously hands-on coding the

00:16:39.000 --> 00:16:40.680
the demo

00:16:40.680 --> 00:16:43.760
and giving the presentation

00:16:43.760 --> 00:16:45.240
and

00:16:45.240 --> 00:16:46.680
anyway I just love to see a CEO who

00:16:46.680 --> 00:16:51.400
knows his stuff.

00:16:51.400 --> 00:16:52.520
And then

00:16:52.520 --> 00:16:54.600
he presented a framework

00:16:54.600 --> 00:16:56.880
an AI observability framework where

00:16:56.880 --> 00:16:58.360
where essentially

00:16:58.360 --> 00:16:59.640
you

00:16:59.640 --> 00:17:01.320
uh the demo I thought was pretty cool.

00:17:01.320 --> 00:17:03.280
It basically

00:17:03.280 --> 00:17:05.959
uses a scan technique to sort of refine

00:17:05.959 --> 00:17:07.240
the prompt.

00:17:07.240 --> 00:17:09.400
So basically you can

00:17:09.400 --> 00:17:10.800
you can go

00:17:10.800 --> 00:17:12.280
it picks the best things of the prompt

00:17:12.280 --> 00:17:13.120
and

00:17:13.120 --> 00:17:15.439
and and so that it has like a

00:17:15.439 --> 00:17:17.600
an agent to sort of

00:17:17.600 --> 00:17:18.560
find a

00:17:18.560 --> 00:17:21.360
a better or try different other prompts.

00:17:21.360 --> 00:17:24.720
And then has like, you know, 100 cycles.

00:17:24.720 --> 00:17:26.760
And then it And then you have like a

00:17:26.760 --> 00:17:28.439
like a golden

00:17:28.439 --> 00:17:29.880
uh test set. You have a test set that

00:17:29.880 --> 00:17:32.040
says like what the right answers were.

00:17:32.040 --> 00:17:34.880
So then the agent's able to to change

00:17:34.880 --> 00:17:38.280
the prompt so that the prompt gives

00:17:38.280 --> 00:17:40.520
better and better and better results

00:17:40.520 --> 00:17:43.040
with high accuracy.

00:17:43.040 --> 00:17:44.560
I thought it was I was super cool, but

00:17:44.560 --> 00:17:46.080
of course I asked the question like, you

00:17:46.080 --> 00:17:48.320
know, if there's a new model

00:17:48.320 --> 00:17:51.160
isn't this all this money spent going to

00:17:51.160 --> 00:17:52.800
going to waste?

00:17:52.800 --> 00:17:55.520
And he said yes, but like in certain in

00:17:55.520 --> 00:17:57.600
certain use cases like uh you know,

00:17:57.600 --> 00:17:59.560
where things

00:17:59.560 --> 00:18:02.120
uh

00:18:02.120 --> 00:18:04.560
where things need to be a bit more

00:18:04.560 --> 00:18:06.360
uh accurate or it's

00:18:06.360 --> 00:18:07.960
deterministic is the wrong word, but

00:18:07.960 --> 00:18:09.760
like when you when you want your prompt

00:18:09.760 --> 00:18:11.280
to be accurate

00:18:11.280 --> 00:18:13.480
if you want your prompt to be, you know,

00:18:13.480 --> 00:18:17.000
instead of 90% accurate 99% accurate

00:18:17.000 --> 00:18:19.280
this is the way to do it or or

00:18:19.280 --> 00:18:22.000
this is one approach to consider.

00:18:22.000 --> 00:18:23.520
And I thought it was fascinating because

00:18:23.520 --> 00:18:25.520
I did You you do start from like a

00:18:25.520 --> 00:18:27.480
stupid prompt like, you know, find the

00:18:27.480 --> 00:18:31.080
family relations of these politicians.

00:18:31.080 --> 00:18:32.160
And then

00:18:32.160 --> 00:18:33.960
and then once you throw throw it into

00:18:33.960 --> 00:18:35.880
this GAN

00:18:35.880 --> 00:18:37.720
um

00:18:37.720 --> 00:18:38.360
uh

00:18:38.360 --> 00:18:40.280
harness

00:18:40.280 --> 00:18:44.760
you have a much more involved um

00:18:44.760 --> 00:18:46.080
prompt

00:18:46.080 --> 00:18:47.840
which has all those words like, you

00:18:47.840 --> 00:18:50.520
know, carefully select and and and has

00:18:50.520 --> 00:18:52.040
all the steps.

00:18:52.040 --> 00:18:54.480
I thought it was fascinating how things.

00:18:54.480 --> 00:18:57.800
I heard also about industry jargon. Ah,

00:18:57.800 --> 00:19:00.680
I remember when when um there was I saw

00:19:00.680 --> 00:19:02.760
some Twitter

00:19:02.760 --> 00:19:05.320
about design because a lot of people

00:19:05.320 --> 00:19:07.080
like complain about AI design looks the

00:19:07.080 --> 00:19:09.640
same. You can load these um

00:19:09.640 --> 00:19:12.040
you know, Anthropic web design

00:19:12.040 --> 00:19:13.960
things, but um

00:19:13.960 --> 00:19:17.080
ultimately the the generated designs are

00:19:17.080 --> 00:19:20.240
not very good. And so, there was this um

00:19:20.240 --> 00:19:22.600
I think it's from the Google Stitch

00:19:22.600 --> 00:19:24.640
announcements when they talk about

00:19:24.640 --> 00:19:26.960
because Stitch is really good at

00:19:26.960 --> 00:19:29.560
creating web designs now.

00:19:29.560 --> 00:19:33.040
And one of the UX designers at Google

00:19:33.040 --> 00:19:35.320
went through the the procedures that he

00:19:35.320 --> 00:19:36.440
took

00:19:36.440 --> 00:19:39.760
um to go from a mediocre web design to

00:19:39.760 --> 00:19:42.640
like a real expert level design where he

00:19:42.640 --> 00:19:44.800
says like, "As a designer, we talk about

00:19:44.800 --> 00:19:46.120
these terminology and this is a

00:19:46.120 --> 00:19:47.720
terminology you need to be a familiar

00:19:47.720 --> 00:19:49.960
with if you want to manage the design

00:19:49.960 --> 00:19:51.880
created by a model to look professional

00:19:51.880 --> 00:19:54.600
and not look like slop." And um and and

00:19:54.600 --> 00:19:57.080
I think it was also related to industry

00:19:57.080 --> 00:20:00.120
jargon like being able to be precise by

00:20:00.120 --> 00:20:01.960
using the right terminology because the

00:20:01.960 --> 00:20:03.280
models they know, right? They've been

00:20:03.280 --> 00:20:05.360
trained all of it. It's just up to us

00:20:05.360 --> 00:20:07.720
like if if we know the terms, we can

00:20:07.720 --> 00:20:11.160
then trigger the the the the level of of

00:20:11.160 --> 00:20:12.280
um

00:20:12.280 --> 00:20:15.840
of what do you say, quality that that is

00:20:15.840 --> 00:20:17.800
that we will accept if we are if we are

00:20:17.800 --> 00:20:19.680
familiar with it. But if we're not, we

00:20:19.680 --> 00:20:21.920
will just accept some low-level quality

00:20:21.920 --> 00:20:23.600
web design because we're like, "Okay, we

00:20:23.600 --> 00:20:24.960
know very basic stuff and this looks

00:20:24.960 --> 00:20:26.360
good enough, right?" But if you're an

00:20:26.360 --> 00:20:28.200
expert web designer, you know all the

00:20:28.200 --> 00:20:30.200
all the trigger terms that basically the

00:20:30.200 --> 00:20:31.600
model knows like, "All right, we're

00:20:31.600 --> 00:20:33.120
talking about some serious professional

00:20:33.120 --> 00:20:35.120
web design here. We're going to do

00:20:35.120 --> 00:20:36.400
We're going to make something that has

00:20:36.400 --> 00:20:38.400
the proper, you know, what do you call

00:20:38.400 --> 00:20:40.240
it, highlights and

00:20:40.240 --> 00:20:42.440
dark tones and whatever. This this tweet

00:20:42.440 --> 00:20:44.080
this tweet was super interesting because

00:20:44.080 --> 00:20:45.600
I really like

00:20:45.600 --> 00:20:47.920
I used to really like doing like, you

00:20:47.920 --> 00:20:50.040
know, graphics and and stuff when I was

00:20:50.040 --> 00:20:52.240
playing Counter-Strike and I made my my

00:20:52.240 --> 00:20:54.800
my team's web page and

00:20:54.800 --> 00:20:57.720
um I want a lot of websites, so I really

00:20:57.720 --> 00:21:00.280
used to love this, but I was just not to

00:21:00.280 --> 00:21:01.000
not that good at

00:21:01.000 --> 00:21:02.600
>> Yeah, it is it is fascinating to think

00:21:02.600 --> 00:21:04.440
that you can drop a word and then it

00:21:04.440 --> 00:21:05.280
will

00:21:05.280 --> 00:21:07.720
the the will behave differently. Well,

00:21:07.720 --> 00:21:08.600
absolutely.

00:21:08.600 --> 00:21:10.160
Well, it makes total sense, right? I

00:21:10.160 --> 00:21:12.200
mean, I was just having a call earlier

00:21:12.200 --> 00:21:13.680
with a um

00:21:13.680 --> 00:21:15.640
This is actually an interesting story.

00:21:15.640 --> 00:21:19.240
Um we identified that in our Terraform

00:21:19.240 --> 00:21:20.840
workspace, we have quite a few

00:21:20.840 --> 00:21:23.240
dependencies, and sometimes cycles were

00:21:23.240 --> 00:21:25.400
created where something depends on

00:21:25.400 --> 00:21:27.600
something, and then that thing ends up

00:21:27.600 --> 00:21:29.520
depending on the previous thing.

00:21:29.520 --> 00:21:32.320
So, we wanted to visualize our our

00:21:32.320 --> 00:21:34.800
Terraform, you know, uh with like remote

00:21:34.800 --> 00:21:36.920
state lookups, how how this whole thing

00:21:36.920 --> 00:21:38.640
is connected. Mhm.

00:21:38.640 --> 00:21:40.160
>> which is what I built the grid for, if

00:21:40.160 --> 00:21:41.240
you remember, like the grid

00:21:41.240 --> 00:21:42.760
>> Yeah, yeah, yeah, I was going to I was

00:21:42.760 --> 00:21:44.000
going to say that, like isn't that what

00:21:44.000 --> 00:21:46.200
Takos is all about? Uh this is what the

00:21:46.200 --> 00:21:49.640
grid uh a separate Takos is is for. It's

00:21:49.640 --> 00:21:51.760
allows you to define your dependencies.

00:21:51.760 --> 00:21:53.600
But, um we have those dependencies in

00:21:53.600 --> 00:21:55.040
our Terraform config, and we're not

00:21:55.040 --> 00:21:56.360
rolling out the grid at the moment due

00:21:56.360 --> 00:21:57.640
to the fact that I'm, you know,

00:21:57.640 --> 00:22:00.080
part-time on that project, so

00:22:00.080 --> 00:22:04.200
Um so, I decided to to just

00:22:04.200 --> 00:22:07.000
I I have somebody new that came in, and

00:22:07.000 --> 00:22:08.120
um

00:22:08.120 --> 00:22:10.920
I'm onboarding him in a very AI-native

00:22:10.920 --> 00:22:12.760
way, asking him to do everything with

00:22:12.760 --> 00:22:17.040
AI. And uh he built something. Um

00:22:17.040 --> 00:22:22.640
And it's very funny, because

00:22:22.640 --> 00:22:26.840
it the model knows an open-source tool

00:22:26.840 --> 00:22:29.280
that maybe has like, I don't know, 100

00:22:29.280 --> 00:22:32.480
people, 100 stars, super low,

00:22:32.480 --> 00:22:35.480
and it has a website. And

00:22:35.480 --> 00:22:37.480
it has a feature that we don't use,

00:22:37.480 --> 00:22:39.040
because I forked it.

00:22:39.040 --> 00:22:42.200
And when it builds something, it says

00:22:42.200 --> 00:22:44.280
it start talking about this feature. And

00:22:44.280 --> 00:22:46.600
I went to through the whole log, like is

00:22:46.600 --> 00:22:48.680
there any web search? Like how did the

00:22:48.680 --> 00:22:49.720
model

00:22:49.720 --> 00:22:51.360
come to the conclusion that the feature

00:22:51.360 --> 00:22:53.840
is there without a web search?

00:22:53.840 --> 00:22:55.760
It can only be because it's somewhere in

00:22:55.760 --> 00:22:58.360
its corpus, and it's inside its weights.

00:22:58.360 --> 00:23:00.680
I don't know how many billion

00:23:00.680 --> 00:23:03.760
parameter model this is, that it, you

00:23:03.760 --> 00:23:05.440
know, knows that this particular open

00:23:05.440 --> 00:23:06.880
source project has that particular

00:23:06.880 --> 00:23:09.000
feature which you know, we don't use.

00:23:09.000 --> 00:23:10.520
But how did it know?

00:23:10.520 --> 00:23:12.520
You know, like you can ask it crazy

00:23:12.520 --> 00:23:14.640
things, like do people throw cats off

00:23:14.640 --> 00:23:16.920
towers in Belgium? And then they will go

00:23:16.920 --> 00:23:18.320
like, yes, there's this tradition in

00:23:18.320 --> 00:23:20.040
this city where people throw cats off

00:23:20.040 --> 00:23:22.360
off off a tower and it's because after

00:23:22.360 --> 00:23:24.440
the winter they had too many cats, so

00:23:24.440 --> 00:23:26.760
they had to throw them down the tower.

00:23:26.760 --> 00:23:27.960
It's like, how the hell does it know

00:23:27.960 --> 00:23:29.640
this?

00:23:29.640 --> 00:23:31.560
You know?

00:23:31.560 --> 00:23:34.520
I think we talked about this before.

00:23:34.520 --> 00:23:36.480
Well, this cat thing I've heard about

00:23:36.480 --> 00:23:40.080
it. Um Yes.

00:23:40.080 --> 00:23:41.560
Uh

00:23:41.560 --> 00:23:43.280
Well

00:23:43.280 --> 00:23:44.520
I

00:23:44.520 --> 00:23:46.480
So, I

00:23:46.480 --> 00:23:49.280
Um I'm approaching my destination soon.

00:23:49.280 --> 00:23:51.720
The the the last um session I went to

00:23:51.720 --> 00:23:55.400
was one by Google DeepMind.

00:23:55.400 --> 00:23:58.520
And the one thing that I found just you

00:23:58.520 --> 00:24:00.400
you were mentioning like an AI native

00:24:00.400 --> 00:24:01.760
onboarding.

00:24:01.760 --> 00:24:04.800
So, the way that they did it was like

00:24:04.800 --> 00:24:06.560
um

00:24:06.560 --> 00:24:08.120
okay, we're going to run through this

00:24:08.120 --> 00:24:10.000
this example.

00:24:10.000 --> 00:24:13.280
I for this voice API Lirium or something

00:24:13.280 --> 00:24:16.120
like this. And they they asked the the

00:24:16.120 --> 00:24:17.680
workshop attendees just to install a

00:24:17.680 --> 00:24:19.480
skill.

00:24:19.480 --> 00:24:22.040
And it worked pretty well. Like there

00:24:22.040 --> 00:24:23.680
were people using

00:24:23.680 --> 00:24:25.040
there were there were people using

00:24:25.040 --> 00:24:26.800
anti-gravity, there's people using like

00:24:26.800 --> 00:24:28.920
myself using Claude Code.

00:24:28.920 --> 00:24:31.000
There was a surprising a lot of people

00:24:31.000 --> 00:24:33.240
using code.

00:24:33.240 --> 00:24:35.160
But but at the end of the day almost

00:24:35.160 --> 00:24:37.600
everyone managed to

00:24:37.600 --> 00:24:39.400
to create

00:24:39.400 --> 00:24:41.280
um

00:24:41.280 --> 00:24:44.040
a proof of concept or whatever, like a a

00:24:44.040 --> 00:24:47.240
demo using these Google APIs just via a

00:24:47.240 --> 00:24:49.200
skill. And I thought to myself

00:24:49.200 --> 00:24:51.120
wow, it's amazing cuz no one wrote any

00:24:51.120 --> 00:24:52.760
code

00:24:52.760 --> 00:24:54.800
uh during this workshop.

00:24:54.800 --> 00:24:57.320
But everyone managed to produce

00:24:57.320 --> 00:25:00.040
um an app

00:25:00.040 --> 00:25:02.840
in a way. So,

00:25:02.840 --> 00:25:05.920
but crazy times.

00:25:05.920 --> 00:25:08.120
This is But it's just amazing that like

00:25:08.120 --> 00:25:10.160
the skills are really You don't need to

00:25:10.160 --> 00:25:15.320
type anything into an editor.

00:25:15.320 --> 00:25:18.000
Oh, I'm passing Westminster Abbey.

00:25:18.000 --> 00:25:19.080
This is where the king gets

00:25:19.080 --> 00:25:21.080
>> Now, there's also this um

00:25:21.080 --> 00:25:23.040
people are talking about how amazing it

00:25:23.040 --> 00:25:26.720
is that uh and um what's Karpathy?

00:25:26.720 --> 00:25:30.240
He he says he talks about uh something

00:25:30.240 --> 00:25:32.400
that he uses, a process.

00:25:32.400 --> 00:25:34.560
And he described it and people really

00:25:34.560 --> 00:25:36.960
liked it the idea. And then he said,

00:25:36.960 --> 00:25:38.880
"Look, if you want to build it, here's a

00:25:38.880 --> 00:25:41.640
gist. Give this to your coding agent and

00:25:41.640 --> 00:25:43.880
it will build the platform for you."

00:25:43.880 --> 00:25:45.640
And then people started raving like, "Oh

00:25:45.640 --> 00:25:47.080
my god, he just did software

00:25:47.080 --> 00:25:49.400
distribution through a gist.

00:25:49.400 --> 00:25:50.520
Ah!"

00:25:50.520 --> 00:25:51.920
Yeah, all through a all through a

00:25:51.920 --> 00:25:53.600
prompt. Okay, I need to get off this bus

00:25:53.600 --> 00:25:54.600
now.

00:25:54.600 --> 00:25:55.600
Um

00:25:55.600 --> 00:25:56.880
So, you're going to have to excuse me. I

00:25:56.880 --> 00:25:57.240
need

00:25:57.240 --> 00:25:58.760
>> one, right? I mean

00:25:58.760 --> 00:26:00.440
I got to put me on mute.

00:26:00.440 --> 00:26:03.480
>> the same thing.

00:26:03.480 --> 00:26:04.760
Although Yeah.

00:26:04.760 --> 00:26:06.160
The the conference venue is right next

00:26:06.160 --> 00:26:07.880
to the Houses of Parliament. I can show

00:26:07.880 --> 00:26:09.320
you

00:26:09.320 --> 00:26:16.280
um Big Ben in a bit.

00:26:16.280 --> 00:26:18.120
Got up at the wrong time. Big Ben from

00:26:18.120 --> 00:26:25.160
the Edge of Tomorrow.

00:26:25.160 --> 00:26:29.120
You know Edge of Tomorrow?

00:26:29.120 --> 00:26:30.800
And dropped.

00:26:30.800 --> 00:26:32.440
So, Edge of Tomorrow is a film with Tom

00:26:32.440 --> 00:26:34.760
Cruise where he is fighting an alien

00:26:34.760 --> 00:26:38.480
invasion and they are actually

00:26:38.480 --> 00:26:41.000
at the part of the war where the humans

00:26:41.000 --> 00:26:44.840
lose and they lose by taking London. So,

00:26:44.840 --> 00:26:46.600
there's a quite a few

00:26:46.600 --> 00:26:49.720
scenes of aliens invading London.

00:26:49.720 --> 00:26:51.920
And the plot of the movie is that

00:26:51.920 --> 00:26:55.000
every time one of the Well, the agents

00:26:55.000 --> 00:26:56.400
have to keep The aliens have a

00:26:56.400 --> 00:26:59.560
capability of certain aliens resetting

00:26:59.560 --> 00:27:02.760
the time if they get killed, allowing

00:27:02.760 --> 00:27:05.040
a replay. It's kind of like a game.

00:27:05.040 --> 00:27:06.920
And for some reason

00:27:06.920 --> 00:27:08.920
one the main actor

00:27:08.920 --> 00:27:10.600
gets that skill. So, every time he dies,

00:27:10.600 --> 00:27:12.000
he resets

00:27:12.000 --> 00:27:13.560
to the previous day

00:27:13.560 --> 00:27:15.480
whenever he acquired the skill of being

00:27:15.480 --> 00:27:17.200
able to reset.

00:27:17.200 --> 00:27:18.560
And why I watched it six times is

00:27:18.560 --> 00:27:20.040
because it's exactly like a computer

00:27:20.040 --> 00:27:21.560
game, you know, when you

00:27:21.560 --> 00:27:23.640
keep getting keep dying and then keep

00:27:23.640 --> 00:27:25.440
getting reset

00:27:25.440 --> 00:27:28.240
to until you execute the level perfectly

00:27:28.240 --> 00:27:30.400
because you can predict things that are

00:27:30.400 --> 00:27:32.040
going to happen.

00:27:32.040 --> 00:27:34.000
Uh wait.

00:27:34.000 --> 00:27:36.360
Are you just talking by yourself?

00:27:36.360 --> 00:27:38.840
Yeah.

00:27:38.840 --> 00:27:40.560
You're you're like

00:27:40.560 --> 00:27:42.840
I That's another thing that that's a bit

00:27:42.840 --> 00:27:44.480
of a trend.

00:27:44.480 --> 00:27:46.360
Oh, oh, there's there's a married couple

00:27:46.360 --> 00:27:48.160
here.

00:27:48.160 --> 00:27:48.800
Uh you mean

00:27:48.800 --> 00:27:51.280
>> Uh we're we're we're basically like

00:27:51.280 --> 00:27:54.280
before before a recorded AI talk, like

00:27:54.280 --> 00:27:55.640
you can go like

00:27:55.640 --> 00:27:56.880
you saw that tweet. I think you shared

00:27:56.880 --> 00:27:58.040
it to me like

00:27:58.040 --> 00:27:59.080
and yeah

00:27:59.080 --> 00:28:00.080
and with my

00:28:00.080 --> 00:28:02.840
the sales figures went 3x and then

00:28:02.840 --> 00:28:04.440
or something crazy like you make some

00:28:04.440 --> 00:28:06.120
crazy statements

00:28:06.120 --> 00:28:07.920
and then people join the call and then

00:28:07.920 --> 00:28:09.800
when it summarizes the headline is

00:28:09.800 --> 00:28:12.480
something insane like And then

00:28:12.480 --> 00:28:13.720
you talked you

00:28:13.720 --> 00:28:17.640
you talked about how the aliens landed

00:28:17.640 --> 00:28:19.440
and uh

00:28:19.440 --> 00:28:21.720
and then then he switched to uh you

00:28:21.720 --> 00:28:23.480
know, some financials. Oh, yeah, you can

00:28:23.480 --> 00:28:26.080
you see Big Ben over there? Yeah, I see

00:28:26.080 --> 00:28:28.800
it. Nice. So so

00:28:28.800 --> 00:28:32.120
Yeah, this is my start.

00:28:32.120 --> 00:28:35.800
So, I'm assuming that when you were

00:28:35.800 --> 00:28:38.040
when I

00:28:38.040 --> 00:28:39.520
when I got disconnected there, you were

00:28:39.520 --> 00:28:41.120
just just talking about some wild

00:28:41.120 --> 00:28:42.400
things.

00:28:42.400 --> 00:28:44.720
No, it was not too wild. It was related

00:28:44.720 --> 00:28:46.600
to the fact that I only saw Big Well, I

00:28:46.600 --> 00:28:48.760
don't I've not Maybe I've seen Big Ben

00:28:48.760 --> 00:28:51.400
in real life, but I was I most memorably

00:28:51.400 --> 00:28:54.120
for me is the scenes in the movie Edge

00:28:54.120 --> 00:28:56.680
of Tomorrow where Tom Cruise is um

00:28:56.680 --> 00:28:57.160
fighting

00:28:57.160 --> 00:28:59.320
>> a good invasion. Is that the one he dies

00:28:59.320 --> 00:29:00.280
all the time, right?

00:29:00.280 --> 00:29:02.120
>> Yeah, he dies and resets. So, I was just

00:29:02.120 --> 00:29:03.800
explaining that that aliens are

00:29:03.800 --> 00:29:04.680
fighting.

00:29:04.680 --> 00:29:07.680
The female lead was great, too. She's uh

00:29:07.680 --> 00:29:09.440
South African, isn't she? Or is she also

00:29:09.440 --> 00:29:12.240
No, no, that that No, she's not. So, not

00:29:12.240 --> 00:29:14.160
that one.

00:29:14.160 --> 00:29:16.640
Yeah, I really liked the the cuz I was

00:29:16.640 --> 00:29:18.440
doing yoga at the time and then there's

00:29:18.440 --> 00:29:20.160
a scene in the in the movie when she's

00:29:20.160 --> 00:29:22.120
doing yoga and I thought, "Yeah, yoga,

00:29:22.120 --> 00:29:24.000
baby."

00:29:24.000 --> 00:29:26.240
She's doing exercises in this training

00:29:26.240 --> 00:29:28.960
hall where um it's like

00:29:28.960 --> 00:29:31.320
a simulation of They're in these huge

00:29:31.320 --> 00:29:33.120
suits to fight, right? And so, have to

00:29:33.120 --> 00:29:33.920
get used to it.

00:29:33.920 --> 00:29:35.880
>> it's super cool movie.

00:29:35.880 --> 00:29:37.720
It's super cool movie. I love it. Yeah,

00:29:37.720 --> 00:29:39.400
great movie.

00:29:39.400 --> 00:29:40.400
Um okay, well

00:29:40.400 --> 00:29:42.160
>> about something else.

00:29:42.160 --> 00:29:44.920
I'm talking something about Ah, yeah, uh

00:29:44.920 --> 00:29:47.040
software distribution through a gist.

00:29:47.040 --> 00:29:48.800
But, it's it's not Karpathy who did this

00:29:48.800 --> 00:29:51.240
first, right? I mean, Minions from

00:29:51.240 --> 00:29:53.160
Stripe was also like just point them at

00:29:53.160 --> 00:29:56.560
this blog and they'll build it

00:29:56.560 --> 00:29:59.120
um type of distribution.

00:29:59.120 --> 00:30:01.080
Yeah.

00:30:01.080 --> 00:30:03.200
It's pretty cool. Yeah.

00:30:03.200 --> 00:30:04.760
But

00:30:04.760 --> 00:30:05.680
I could I could

00:30:05.680 --> 00:30:07.520
I couldn't help but think that like I

00:30:07.520 --> 00:30:10.120
think we had a blip with the internet or

00:30:10.120 --> 00:30:13.800
pulled um Anthropic was was uh having

00:30:13.800 --> 00:30:16.680
some disruptions Yeah, they do they

00:30:16.680 --> 00:30:19.120
>> in the day yesterday. Yeah, yeah. And I

00:30:19.120 --> 00:30:22.200
couldn't help but think to myself like

00:30:22.200 --> 00:30:23.640
you know

00:30:23.640 --> 00:30:26.120
we we we we we think how cool it is that

00:30:26.120 --> 00:30:28.800
you can build something with uh a

00:30:28.800 --> 00:30:30.440
prompt.

00:30:30.440 --> 00:30:32.520
But, of course, if you don't have that

00:30:32.520 --> 00:30:35.160
infrastructure in place

00:30:35.160 --> 00:30:37.880
then

00:30:37.880 --> 00:30:38.680
But, that's why

00:30:38.680 --> 00:30:39.920
>> don't have

00:30:39.920 --> 00:30:42.480
I I like just imagine the future where

00:30:42.480 --> 00:30:45.280
this mythos is amazing and it's able to

00:30:45.280 --> 00:30:46.880
do our work

00:30:46.880 --> 00:30:49.920
and I'm like happily using it like

00:30:49.920 --> 00:30:52.040
probably most people are

00:30:52.040 --> 00:30:53.800
but the minute it it's it gets taken

00:30:53.800 --> 00:30:56.080
away from me

00:30:56.080 --> 00:30:57.520
it's like

00:30:57.520 --> 00:30:59.480
it's going to be crazy.

00:30:59.480 --> 00:31:01.600
Yeah, but but at the same time I mean we

00:31:01.600 --> 00:31:02.960
talked about this before, right? That's

00:31:02.960 --> 00:31:05.440
why people that are like doom prepping

00:31:05.440 --> 00:31:08.000
for the time that those AI inference

00:31:08.000 --> 00:31:11.080
providers whatever due to war or

00:31:11.080 --> 00:31:13.160
whatever are no longer able to provide

00:31:13.160 --> 00:31:15.000
the service cuz I've been in countries

00:31:15.000 --> 00:31:16.680
where the service is not active and it's

00:31:16.680 --> 00:31:18.200
really painful.

00:31:18.200 --> 00:31:21.520
Um, that's why and now with Gemma from

00:31:21.520 --> 00:31:23.760
from Google people are I mean I'm

00:31:23.760 --> 00:31:26.960
reading very very conflicting

00:31:26.960 --> 00:31:28.400
evaluations of it. People are like, "Oh

00:31:28.400 --> 00:31:29.920
my god, Gemma is really worse than

00:31:29.920 --> 00:31:31.000
Gwen."

00:31:31.000 --> 00:31:32.240
Uh, and some people are like, "Once you

00:31:32.240 --> 00:31:34.360
fine-tune it and you you know how to set

00:31:34.360 --> 00:31:37.840
it up properly, Gemma is amazing." What

00:31:37.840 --> 00:31:40.160
about Gemma? Like is it like an open

00:31:40.160 --> 00:31:42.560
Open weights model, yeah. Like it's on

00:31:42.560 --> 00:31:44.520
on hugging face.

00:31:44.520 --> 00:31:45.840
Um, so you can

00:31:45.840 --> 00:31:46.960
run it.

00:31:46.960 --> 00:31:49.480
Uh, I'm waiting for a a big paycheck and

00:31:49.480 --> 00:31:52.200
and buying like an M5. You know, Mac Mac

00:31:52.200 --> 00:31:53.960
is people always say like Apple was

00:31:53.960 --> 00:31:55.520
dumb, they don't have Apple intelligent

00:31:55.520 --> 00:31:57.080
like their Apple intelligence projects

00:31:57.080 --> 00:31:58.840
failed, but they really built great

00:31:58.840 --> 00:32:01.040
hardware where they they make the

00:32:01.040 --> 00:32:03.160
bandwidth between the memory and the CPU

00:32:03.160 --> 00:32:06.720
so good that Yeah, yeah. We were joking

00:32:06.720 --> 00:32:08.640
there was some there was some ex-Apple

00:32:08.640 --> 00:32:10.560
people in the in our telling to over

00:32:10.560 --> 00:32:12.280
lunch.

00:32:12.280 --> 00:32:14.240
And uh, they were sort of joking that

00:32:14.240 --> 00:32:16.680
like Apple is so bad at AI, but they

00:32:16.680 --> 00:32:18.280
have the best hardware.

00:32:18.280 --> 00:32:20.600
Yeah. And uh,

00:32:20.600 --> 00:32:23.640
and right now their strategy seems to be

00:32:23.640 --> 00:32:26.200
uh, you know, pulling pulling along

00:32:26.200 --> 00:32:29.920
either Google or uh, open AI

00:32:29.920 --> 00:32:32.000
to get the best

00:32:32.000 --> 00:32:34.400
open um, to get the best model without

00:32:34.400 --> 00:32:37.040
them actually having to do anything.

00:32:37.040 --> 00:32:38.240
Cuz I guess they have the cash or

00:32:38.240 --> 00:32:40.680
something.

00:32:40.680 --> 00:32:42.040
Cuz they they haven't actually fully

00:32:42.040 --> 00:32:44.360
committed to using the Google

00:32:44.360 --> 00:32:46.960
model or or open AI model in their in

00:32:46.960 --> 00:32:48.520
their products

00:32:48.520 --> 00:32:50.280
allegedly.

00:32:50.280 --> 00:32:51.760
So people say it's because Apple has

00:32:51.760 --> 00:32:54.720
such a high reputation to uphold. But I

00:32:54.720 --> 00:32:57.120
mean that that if like because the AI

00:32:57.120 --> 00:33:00.920
can fail so spectacularly um because of

00:33:00.920 --> 00:33:03.080
it being non-deterministic and it could

00:33:03.080 --> 00:33:04.920
result in the product being in a very

00:33:04.920 --> 00:33:05.800
bad

00:33:05.800 --> 00:33:08.200
daylight like where it fails.

00:33:08.200 --> 00:33:09.480
Yeah, that's that's why they haven't

00:33:09.480 --> 00:33:12.760
integrated it but He he mentioned it uh

00:33:12.760 --> 00:33:15.160
one of the ex-Apple people said like,

00:33:15.160 --> 00:33:17.160
you know, they own the the interface or

00:33:17.160 --> 00:33:20.000
they own the platform so they are like

00:33:20.000 --> 00:33:22.040
in the best position, right?

00:33:22.040 --> 00:33:24.680
And we we just observed inside

00:33:24.680 --> 00:33:26.680
the lunch hall, you know, most people

00:33:26.680 --> 00:33:29.120
are using

00:33:29.120 --> 00:33:31.040
most people are using iPhones, right?

00:33:31.040 --> 00:33:32.920
Okay, I got to get This is This is the

00:33:32.920 --> 00:33:35.000
thumbnail here. This is the AI engineer

00:33:35.000 --> 00:33:36.120
Europe

00:33:36.120 --> 00:33:39.320
thumbnails.

00:33:39.320 --> 00:33:40.840
They had like a huge lobster out here

00:33:40.840 --> 00:33:42.040
yesterday.

00:33:42.040 --> 00:33:44.000
Yeah, you sent me a picture.

00:33:44.000 --> 00:33:45.480
We put it on I don't know why I I don't

00:33:45.480 --> 00:33:47.840
know why it isn't here today.

00:33:47.840 --> 00:33:51.000
Possibly something

00:33:51.000 --> 00:33:54.440
Possibly. I don't even like

00:33:54.440 --> 00:33:56.240
I'm like, how do they get these huge

00:33:56.240 --> 00:33:58.320
lobsters? Is someone just traveling

00:33:58.320 --> 00:34:01.800
around with a [ __ ] with a [ __ ]

00:34:01.800 --> 00:34:02.720
But

00:34:02.720 --> 00:34:04.400
But how did how did they get a lobster

00:34:04.400 --> 00:34:06.400
here?

00:34:06.400 --> 00:34:08.159
These buildings all around here are just

00:34:08.159 --> 00:34:10.440
magnificent. I love them.

00:34:10.440 --> 00:34:12.159
It looks like Ghent. I mean, if you go

00:34:12.159 --> 00:34:13.800
to city center Ghent, you have these

00:34:13.800 --> 00:34:16.040
massive um

00:34:16.040 --> 00:34:19.560
like cathedrals and and and wide market

00:34:19.560 --> 00:34:23.040
places very similar to this.

00:34:23.040 --> 00:34:24.840
So hey, do you have a Riverside account

00:34:24.840 --> 00:34:26.800
because ideally I I could end this

00:34:26.800 --> 00:34:28.800
recording and you could you could upload

00:34:28.800 --> 00:34:29.440
the

00:34:29.440 --> 00:34:32.360
or finish editing these three clips.

00:34:32.360 --> 00:34:33.919
Cuz I'm going to have an absolutely no

00:34:33.919 --> 00:34:35.879
time today.

00:34:35.879 --> 00:34:37.800
I have a Riverside account. I even have

00:34:37.800 --> 00:34:40.879
pro for 2 months for free.

00:34:40.879 --> 00:34:42.480
And uh

00:34:42.480 --> 00:34:44.679
Maybe I'll try invite you. Or something.

00:34:44.679 --> 00:34:46.280
I don't know. I don't know if it works

00:34:46.280 --> 00:34:47.679
if two people have different accounts.

00:34:47.679 --> 00:34:49.360
You mentioned that if I want to edit it,

00:34:49.360 --> 00:34:52.520
then it needs to be with a Can you just

00:34:52.520 --> 00:34:54.840
send me your email or your login? Cuz I

00:34:54.840 --> 00:34:56.240
tried it last time and it said like

00:34:56.240 --> 00:34:58.200
this.

00:34:58.200 --> 00:35:00.760
Mhm. Okay, I'm going to go in here and

00:35:00.760 --> 00:35:02.480
and volunteer.

00:35:02.480 --> 00:35:04.000
Uh I'm going to be doing crowd control

00:35:04.000 --> 00:35:06.840
this morning. This is the way I scored

00:35:06.840 --> 00:35:08.720
entrance.

00:35:08.720 --> 00:35:09.920
Anyway,

00:35:09.920 --> 00:35:11.800
it's going to be a great day. Bye. All

00:35:11.800 --> 00:35:20.720
right. Bye.

00:35:20.720 --> 00:35:22.880
I