WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=JRTAtZ5iBkU

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/JRTAtZ5iBkU.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:07.860 --> 00:00:09.860
[music]

s2
00:00:10.400 --> 00:00:14.410
Hi, I'm Amol, CEO of Nori Atentic.

s3
00:00:14.410 --> 00:00:14.880
[music]

s4
00:00:14.880 --> 00:00:20.960
We deploy an AI employee that understands your company, your code, docs, Slack,

s5
00:00:20.960 --> 00:00:21.440
[music]

s6
00:00:21.440 --> 00:00:23.160
and other kinds of data.

s7
00:00:23.160 --> 00:00:28.000
We spend a lot of time thinking about how coding agents really work.

s8
00:00:28.000 --> 00:00:31.200
Most people think coding agents only write code.

s9
00:00:31.200 --> 00:00:33.405
But if you ask me, that's just bad marketing.

s10
00:00:33.405 --> 00:00:33.720
[music]

s11
00:00:33.720 --> 00:00:35.320
Forget the name for a second.

s12
00:00:35.320 --> 00:00:38.240
Coding agents can do almost anything.

s13
00:00:38.240 --> 00:00:39.720
There's just one trick.

s14
00:00:39.720 --> 00:00:43.885
You have to be able to think like an agent to get it to do what you want it to do.

s15
00:00:43.885 --> 00:00:44.560
[music]

s16
00:00:44.560 --> 00:00:53.710
Today, we're going to talk about how we use coding agents to do something most people think agents are terrible at, make visual artifacts, like slides, docs,

s17
00:00:53.710 --> 00:00:53.960
[music]

s18
00:00:53.960 --> 00:00:57.320
and yeah, even video.

s19
00:00:58.295 --> 00:00:59.280
[music]

s20
00:00:59.280 --> 00:01:05.680
Every day, the world pours something like 34,000 human years into making slide decks.

s21
00:01:05.680 --> 00:01:08.775
Most of that time isn't the thinking, it's the fiddling.

s22
00:01:08.775 --> 00:01:09.840
[music]

s23
00:01:09.840 --> 00:01:18.680
A deck that takes 10 hours should really take about 25 minutes once you remove all the formatting and the branding and the moving things around.

s24
00:01:18.680 --> 00:01:20.520
Say you need to make a slide.

s25
00:01:20.520 --> 00:01:21.920
What do you do?

s26
00:01:21.920 --> 00:01:29.000
You open a tool, PowerPoint, Slides, Figma, Canva, and then you start manipulating a canvas.

s27
00:01:29.000 --> 00:01:32.880
Every one of these tools is built for human hands and human eyes.

s28
00:01:32.880 --> 00:01:41.480
Click, drag, drop, resize, snap to grid, all motions and patterns that make sense for our geospatial view of the world.

s29
00:01:41.480 --> 00:01:47.040
There is a data structure underneath, but it's in a format that only the application can read.

s30
00:01:47.040 --> 00:01:49.640
What happens when you hand these tools to an agent?

s31
00:01:49.640 --> 00:01:52.560
Well, the output comes out all wrong.

s32
00:01:52.560 --> 00:01:58.720
Things overlap in weird ways, you can't see the text, there's no alignment, It's just garbage.

s33
00:01:58.720 --> 00:02:01.560
AI skeptics say that it's not just the tools.

s34
00:02:01.560 --> 00:02:04.560
Agents fundamentally can't reason about space.

s35
00:02:04.560 --> 00:02:10.479
And there are whole benchmarks like Arc AGI that are built exactly around that premise.

s36
00:02:10.479 --> 00:02:14.200
There's a famous little test for this from developer Simon Willison.

s37
00:02:14.200 --> 00:02:17.320
He asks every new model the same thing.

s38
00:02:17.320 --> 00:02:20.520
Can you draw a pelican riding a bicycle?

s39
00:02:20.520 --> 00:02:21.800
But there's a trick.

s40
00:02:21.800 --> 00:02:25.120
The agent is only allowed to use SVG.

s41
00:02:25.120 --> 00:02:29.000
It's a quick gut check for whether a model can reason about space at all.

s42
00:02:29.000 --> 00:02:32.320
Here's some examples of what the models actually give you on this test.

s43
00:02:32.320 --> 00:02:32.616
And

s44
00:02:32.616 --> 00:02:33.120
[music]

s45
00:02:33.120 --> 00:02:35.760
yeah, these are pretty bad.

s46
00:02:35.760 --> 00:02:38.511
Like genuinely deeply really bad.

s47
00:02:38.511 --> 00:02:39.400
[music]

s48
00:02:39.400 --> 00:02:41.000
So, does that mean it's hopeless?

s49
00:02:41.000 --> 00:02:43.720
Agents are just doomed to be bad at graphics?

s50
00:02:43.720 --> 00:02:45.959
No, I don't think so.

s51
00:02:45.959 --> 00:02:49.680
If you ask me, it's not the model, it's the medium.

s52
00:02:49.680 --> 00:02:57.506
If I asked you, someone who is presumably human, to handwrite an SVG of a pelican, you wouldn't be able to do that either.

s53
00:02:57.506 --> 00:02:57.920
[music]

s54
00:02:57.920 --> 00:03:00.560
SVGs are just a wall of numbers.

s55
00:03:00.560 --> 00:03:05.120
You can't go from a wall of numbers to a pelican, you just can't see that way.

s56
00:03:05.120 --> 00:03:07.560
That's just not how people think.

s57
00:03:07.560 --> 00:03:09.160
We think graphically.

s58
00:03:09.160 --> 00:03:12.200
So, we build tools that let us draw on a canvas.

s59
00:03:12.200 --> 00:03:19.800
Figma MCPs, PowerPoint CLIs, screenshot and replace loops, what do all of these agent tools have in common?

s60
00:03:19.800 --> 00:03:23.160
They all approach the problem like a human.

s61
00:03:23.160 --> 00:03:25.280
But an AI is not a human.

s62
00:03:25.280 --> 00:03:29.959
Asking an AI to use a canvas is like asking a human to write SVG by hand.

s63
00:03:29.959 --> 00:03:31.800
It doesn't really make sense.

s64
00:03:31.800 --> 00:03:37.760
You need to give the AI tools based on how it thinks, not in pixels, in language.

s65
00:03:37.760 --> 00:03:42.800
Words, tokens, structure, that is its native medium.

s66
00:03:42.800 --> 00:03:46.600
Imagine a language that's incredible at describing layout.

s67
00:03:46.600 --> 00:03:50.440
That models have seen and trained on billions of examples of.

s68
00:03:50.440 --> 00:03:52.600
That they understand intuitively.

s69
00:03:52.600 --> 00:03:57.320
that renders to pixels and can run everywhere.

s70
00:03:57.640 --> 00:03:59.240
Oh, right.

s71
00:03:59.240 --> 00:04:02.320
HTML lets the model think in structure.

s72
00:04:02.320 --> 00:04:05.080
HTML tags have meanings built into the language.

s73
00:04:05.080 --> 00:04:09.440
A heading, a chart, a grid, and the browser turns it all into pixels.

s74
00:04:09.440 --> 00:04:19.480
So, the model never actually places a coordinate, and you can get all sorts of visual effects, charts and layouts, fonts and motion, all of it for free.

s75
00:04:19.480 --> 00:04:21.440
Remember that pelican from earlier?

s76
00:04:21.440 --> 00:04:25.640
Now, ask it to do the same exact task, but in HTML.

s77
00:04:25.640 --> 00:04:29.919
Same bird, but now it's in a structure that the model can reason about.

s78
00:04:29.919 --> 00:04:34.600
And you can read and theme and edit every single line of it.

s79
00:04:34.600 --> 00:04:37.600
I spent my whole life building slide decks with PowerPoint.

s80
00:04:37.600 --> 00:04:41.653
So, I always thought that those two things, slide decks and PowerPoint, were synonyms.

s81
00:04:41.653 --> 00:04:42.600
[music]

s82
00:04:42.600 --> 00:04:44.800
But that's just not really true, is it?

s83
00:04:44.800 --> 00:04:49.280
PowerPoint is a tool that you use to make slide decks.

s84
00:04:49.280 --> 00:04:52.800
The deck itself, that's just the presentation mode.

s85
00:04:52.800 --> 00:04:57.880
And as it turns out, no one in your audience is going to care how you got to the presentation mode.

s86
00:04:57.880 --> 00:05:00.648
The editing format is totally arbitrary.

s87
00:05:00.648 --> 00:05:00.919
[music]

s88
00:05:00.919 --> 00:05:11.000
So, you can just pick the editing format that the agents are already good at, HTML, and if you need to render to a different format like PDF later on.

s89
00:05:11.000 --> 00:05:16.368
We use this HTML trick to build all of our slide decks, our board decks, and our sales decks.

s90
00:05:16.368 --> 00:05:16.560
[music]

s91
00:05:16.560 --> 00:05:20.560
These are real things that we actually present and send out constantly.

s92
00:05:20.560 --> 00:05:22.400
We use it for our docs, too.

s93
00:05:22.400 --> 00:05:26.760
It gives our docs color and vibrancy, all while following our brand.

s94
00:05:26.760 --> 00:05:31.600
And of course, we also use it to make videos like this one.

s95
00:05:31.600 --> 00:05:34.640
What you're watching is just HTML and CSS.

s96
00:05:34.640 --> 00:05:38.600
It's literally just divs all the way down.

s97
00:05:40.720 --> 00:05:44.800
Almost everything is better with a little structure and a little bit of color.

s98
00:05:44.800 --> 00:05:53.000
Plain text is a choice, generally a choice of convenience, but it's usually the wrong one if you're actually trying to create something of use.

s99
00:05:53.000 --> 00:06:00.280
Now, I do want to take a quick beat here and point out that a beautiful deck on its own is generally not worth anything.

s100
00:06:00.280 --> 00:06:05.880
You still have to go and get all of that content, all of the things that actually populate that deck, right?

s101
00:06:05.880 --> 00:06:09.000
Well, again, we can think like the model.

s102
00:06:09.000 --> 00:06:17.240
If you just give the model access to your data, say your call transcripts or your emails, you can have the model build the deck end to end.

s103
00:06:17.240 --> 00:06:22.080
Let your agents do all the grunt work while you focus on vision and story.

s104
00:06:22.080 --> 00:06:24.440
That's what Nori Sessions lets you do.

s105
00:06:24.440 --> 00:06:29.120
I've built entire board decks for my phone on the subway during my commute.

s106
00:06:29.120 --> 00:06:30.000
Why?

s107
00:06:30.000 --> 00:06:33.400
Because our Nori bot lives in the fabric of our company.

s108
00:06:33.400 --> 00:06:39.760
Of course, Nori ships with everything you need to make this all work, so don't bother reinventing the wheel.

s109
00:06:39.760 --> 00:06:40.800
That's my little spiel.

s110
00:06:40.800 --> 00:06:42.200
Thanks for listening.

s111
00:06:42.200 --> 00:06:45.440
If you have just one takeaway, it's this.

s112
00:06:45.440 --> 00:06:47.440
Stop thinking like a user.

s113
00:06:47.440 --> 00:06:49.040
Think like the model.

s114
00:06:49.040 --> 00:06:54.619
Give it the right language, and for graphics, all you need is HTML.

s115
00:06:54.619 --> 00:06:56.619
[music]