A New Layer of Engineering Is Emerging | Hermes Labs Field Notes Ep. 1

These systems are good at synthesizing

what they know, but they're not good at

telling you what they don't know. Right?

And that's really, really important

because a lot of decisions aren't just

made on what you know. They're also made

on what you don't know. And you

sometimes have to learn to not make a

decision based on what you don't know.

Right. This is going to be a fun video.

Hello everybody. Early Boss from Hermes

Labs here. And this is Hermes Labs

Diaries episode 1. I decided to start

this series after a long time not making

videos because I've been watching a lot

of these startup videos. Startup

something episode one and episode two

and day in the life of and they're

really inspiring. At least to me they're

really inspiring. But I found that to me

they're not relatable. So I figured that

being a solo guy who went from not being

technical at all to being what a few of

my engineer friends have called

basically top down engineering. So

instead of approaching engineering from

the syntax up, just learning from the

from the top down, from the systems and

the abstractions, probably way cooler

story to some people or at least a good

complimentary story to the other side.

So let's do this. So if you see my

previous videos, you've probably seen

that around a year ago, I started making

AI and philosophy, AI and anthropology,

and AI and language videos. At the time,

I was not technical at all. Honestly, I

started using Replet around that time

and that's as technical as I felt. But I

was really really obsessed and I

probably spent way more money than I

should have on Replet, but in hindsight,

it was part of the learning curve and

that money was an investment that

probably people in four years of college

don't don't get the the same return. So,

how do you go from that to at this point

having a a basically epistemic

engineering company, right? where you're

confident and comfortable going up to

senior engineers or going up to

companies. I'm not really going to say

names. I'm in San Francisco, but I've

got to meet a lot of people who don't

really know how their AI outputs are

generated and they're using these

systems where they kind of retroactively

fix them. And it it's become really,

really, really cool to me to be able to

see my progress. And I think other

people have like seen it when I spent

too much time off YouTube and I should

have been showcasing it. So, I think

it's time. So, let's get to it. What is

Hermes Labs? what I say that it's one of

the first cyber companies out there. And

what do I mean by I went from

nontechnical to technical? Mike, it it

it adds it adds a human element that I

think as much as we're turning to a

digital age and I'm one of the first

ones that says if I let these agents go

go wild with the with the runtime

assurance that somebody like like Hermes

Labs and like Roly Boss can provide, we

definitely need the human element in

there. So, minimum editing, just me

talking about the hardships because it's

definitely not easy when you're doing

something without credentials. It really

doesn't matter if you can really show

value to people, that's all they care

about. So, let's just drop the

pretensions. Let's drop the M dashes.

Let's drop the the scripts that

everybody's playing and just talk about

reality, right? How did I go from making

videos about rock carvings being the

original AI around a year ago that if

you go and talk to Gemini and mention

Roelly Bosch, that's one of the things

that it might bring up even though it

has like 200 views. It's amazing how

training data works. How do I go from

that to now having semi-autonomous

system that's basically creating

inrogate software based on based on

harmonics and based on epistemology,

right? And and why does hermeneutics

matter? Why does epistemology matter?

what even are these things? Because I

talk to people and nobody knows even

though you use both you understand these

concepts at a very basic level for sure.

So I think we're at a point where when

you talk to people about the AI outputs

that they get, why they see what they

see, why they talk to the thing like

they talk to the thing, you're going to

get a lot of projection. You're going to

get a lot of I don't know. You're

getting a lot wishful thinking as well,

but you get very very very little

feedback that lets you know that people

are making a response guaranteeing that

the output of these systems is like well

chiseled and well fundamented and well

based. And that's how you go from

talking about ancient rock carvings

being AI to then talking about AI and

philosophy and and and Jeffrey Hinton

and and the issues there and and the

importance of language because you're

parting from a basis of what does this

thing even mean to begin with and then

once you part from the basis of what

does this thing even mean to begin with

then when you see an output then there's

a question of like well what does this

output mean and and and where does it

come from and how is it generated not

from a mechanistic level but again from

what is true and what does it mean

level, right? And and and that matters,

right? Because at the end of the day, if

and I see this all the time when I when

I go out to events and hackathons and

stuff and and demo days, if you're

basically doing I mean engineering from

the like from the bottom up, if if you

spend a a couple months just working on

on a little bit of the fundamentals, you

could do a lot of stuff now with AI.

Basically, a lot of stuff is just

plugging when it comes to to to using

APIs or hackathons, they're just

plugging API keys, right, into a pretty

interface and and often creating like a

good user flow that like that can scale

and that's hardened. But when you're

working with these systems that have so

much potential, because they do, but

also such a huge range and possibilities

of just screwing up and causing huge

liability, knowing how to get what you

want is part of the engineering, right?

And I I often talk to people and they

say, "Well, we got the best context. We

got the best data and and it's like

they're talking about, you know, well,

the model's going to get smarter and and

it's already really smart." But I mean,

it's it's a sarcastic parrot, right?

Where basically it doesn't even know

what it regurgitates and it regurgitates

based on a median, right? So it, you

know, I I these people who are often

like top of their class in like Stanford

or Berkeley or, you know, top top of the

line schools in like India, China or or

the East Coast, they're talking about

these things. And I tell them, look,

you're not the median student, right?

And you've been in classes where you've

had bad students and median students,

and you all share the same context and

the same data, but you all have

different intent, you have different

framing. um and you leverage that

context differently, right? And

different level of attention to it, etc.

It's the same thing just dumping all the

context and and and even if you do

context engineering, well, I mean, it's

the same thing, right? Um just like for

me to be able to get here, what matter

wasn't just the context. Everybody has

the same context on AI. It was the

intent and the attention and the

precision and um I mean the intent the

intent I can't highlight it enough to be

able to chisel that with as much

precision as you can. It's not easy and

you have to find new ways. But as

somebody who before I considered myself

an engineer, I thought of this stuff

like isn't engineering kind of problem

solving in a way and isn't not

understanding the output like one of the

biggest problems. So why are we not all