WEBVTT

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

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

s1
00:00:00.920 --> 00:00:07.000
Okay, I want to tell you a story about a factory that taught itself how to remember.

s2
00:00:07.000 --> 00:00:08.680
Hi, I'm Rushabh.

s3
00:00:08.680 --> 00:00:11.920
I run Machinecraft, a 100 people factory in India.

s4
00:00:11.920 --> 00:00:15.280
No data science team, no ML budget, none of that.

s5
00:00:15.280 --> 00:00:21.520
And somehow we ended up building a 36 AI agent that runs our entire go-to-market.

s6
00:00:21.520 --> 00:00:24.920
I think that's still a little ridiculous.

s7
00:00:24.920 --> 00:00:30.080
Let me show you how it happened and why you can do the same thing.

s8
00:00:30.080 --> 00:00:32.040
So, here's the thing about our company.

s9
00:00:32.040 --> 00:00:34.960
From the outside it looks like machines and metal.

s10
00:00:34.960 --> 00:00:40.240
But the actual company, the part that matters, isn't the machines, it's the knowledge.

s11
00:00:40.240 --> 00:00:47.560
Who the customer is, what we quoted them in 2019, why that one machine needed that weird custom tweak.

s12
00:00:47.560 --> 00:00:51.600
And for three generations, all of that lived in exactly three brains.

s13
00:00:51.600 --> 00:00:56.680
Initially my grandfather's, then my father's, and now mine.

s14
00:00:57.640 --> 00:01:00.720
Which is a genuinely terrifying way to run a company.

s15
00:01:00.720 --> 00:01:02.800
When you sit with it.

s16
00:01:02.800 --> 00:01:08.640
A lot of people have joined us, people have left us, the revolving door never stopped.

s17
00:01:08.640 --> 00:01:15.280
And every single time someone walked out, a chunk of our brain walked out with them.

s18
00:01:15.280 --> 00:01:18.800
We weren't scared of the competitors, we were scared of forgetting.

s19
00:01:18.800 --> 00:01:27.280
Or waking up one day and realizing the whole company only existed inside two increasingly tired heads.

s20
00:01:27.400 --> 00:01:29.640
So, I had an idea.

s21
00:01:29.640 --> 00:01:31.080
I'll be honest.

s22
00:01:31.080 --> 00:01:32.800
Sounded insane first.

s23
00:01:32.800 --> 00:01:42.240
But what if instead of writing the knowledge down in some document nobody ever reads, what if we grew a brain that just held it?

s24
00:01:42.240 --> 00:01:46.680
Not a chatbot you poke at, a twin of the company.

s25
00:01:46.680 --> 00:01:48.600
I didn't hire a sales team.

s26
00:01:48.600 --> 00:01:50.960
I tried to build one.

s27
00:01:50.960 --> 00:01:55.760
A quick detour because you need to know how messy this is.

s28
00:01:55.760 --> 00:01:57.880
We make thermoforming machines.

s29
00:01:57.880 --> 00:02:01.120
They heat up a plastic sheet and shape it.

s30
00:02:01.120 --> 00:02:10.840
Same core machine, but it ends up making hydroponic farm trays, spa bathtubs, EV car panels, medical casings, and even packaging.

s31
00:02:10.840 --> 00:02:14.959
Seven totally different worlds, seven totally different buyers.

s32
00:02:14.959 --> 00:02:18.520
So, this brain couldn't just memorize a brochure.

s33
00:02:18.520 --> 00:02:24.560
It had to know which universe a given customer lives in.

s34
00:02:24.560 --> 00:02:27.400
Step one was almost boringly simple.

s35
00:02:27.400 --> 00:02:30.400
Feed it everything, and I mean everything.

s36
00:02:30.400 --> 00:02:38.720
Years of quotes, drawings, payment schedules, timelines, email threads, hundreds of gigabytes of our own private history.

s37
00:02:38.720 --> 00:02:42.760
Not the public internet, our internet.

s38
00:02:42.880 --> 00:02:44.640
And here's the plot twist.

s39
00:02:44.640 --> 00:02:48.360
The part that surprises every engineer I tell this to.

s40
00:02:48.360 --> 00:02:50.120
We never trained a model.

s41
00:02:50.120 --> 00:02:53.360
No GPUs humming in the basement, no fine-tuning.

s42
00:02:53.360 --> 00:03:02.959
We just looked at all the history, chopped it into bite-size chunks, and let off-the-shelf models read it and pull out the facts.

s43
00:03:02.959 --> 00:03:07.840
We stored the meaning of each chunk as vectors and relationships.

s44
00:03:07.840 --> 00:03:10.200
Who's connected to what as a graph.

s45
00:03:10.200 --> 00:03:12.720
The brain isn't a smarter model.

s46
00:03:12.720 --> 00:03:17.760
It's actually a really, really well-organized memory.

s47
00:03:17.760 --> 00:03:22.000
Now, this is where it gets a little weird in a good way.

s48
00:03:22.000 --> 00:03:27.440
We stopped thinking of Era as a software and started thinking of it as something we were raising.

s49
00:03:27.440 --> 00:03:30.920
So, we gave it a body modeled on biology.

s50
00:03:30.920 --> 00:03:42.360
Senses to figure out who it's talking to, a gut to digest the documents into facts, a memory, a dream cycle, an immune system to fight off bad information.

s51
00:03:42.360 --> 00:03:43.560
Why biology?

s52
00:03:43.560 --> 00:03:50.000
Well, because evolution already spent a billion years solving how do you stay coherent over time?

s53
00:03:50.000 --> 00:03:52.519
We just copied the homework.

s54
00:03:52.519 --> 00:03:54.440
Okay, so the big question.

s55
00:03:54.440 --> 00:03:58.080
Why 36 agents instead of one genius mega prompt?

s56
00:03:58.080 --> 00:04:07.200
Because, and you already know this if you've ever tried it, one prompt that's supposed to do everything ends up doing everything badly.

s57
00:04:07.200 --> 00:04:10.480
So, Ira isn't one mind, it's a pantheon.

s58
00:04:10.480 --> 00:04:13.560
A whole cast of specialists.

s59
00:04:13.560 --> 00:04:16.200
Each one has exactly one job.

s60
00:04:16.200 --> 00:04:18.359
Athena runs the room.

s61
00:04:18.359 --> 00:04:21.239
Prometheus owns the sale.

s62
00:04:21.239 --> 00:04:23.640
Plutus does pricing.

s63
00:04:23.640 --> 00:04:27.480
Hephaestus knows every machine spec cold.

s64
00:04:27.480 --> 00:04:33.160
Vera fact checks everything, and Memnon, my favorite, guards corrections.

s65
00:04:33.160 --> 00:04:38.320
So, the second a human fixes something, it stays fixed forever.

s66
00:04:38.320 --> 00:04:40.480
One agent, one job.

s67
00:04:40.480 --> 00:04:43.800
It's a team, not a hero.

s68
00:04:43.800 --> 00:04:45.480
And here's the cool part.

s69
00:04:45.480 --> 00:04:46.919
They hold meetings.

s70
00:04:46.919 --> 00:04:49.200
Athena pulls in specialists.

s71
00:04:49.200 --> 00:04:53.880
They actually argue, and a single answer comes out the other side.

s72
00:04:53.880 --> 00:05:00.320
It's like having a board room that never sleeps, never gets tired, and somehow has no ego.

s73
00:05:00.320 --> 00:05:03.320
So, what does all this actually run?

s74
00:05:03.320 --> 00:05:06.240
Honestly, the whole front business.

s75
00:05:06.240 --> 00:05:12.919
Everything between a stranger exists somewhere, and now they're a customer.

s76
00:05:13.040 --> 00:05:15.600
Nine concrete jobs every single day.

s77
00:05:15.600 --> 00:05:19.600
Outbound emails that actually reference my real world.

s78
00:05:19.600 --> 00:05:24.120
Account briefs built from cross-checked truths before a call.

s79
00:05:24.120 --> 00:05:25.440
Quotations.

s80
00:05:25.440 --> 00:05:29.480
A swipe left, swipe right mode for outreach.

s81
00:05:29.480 --> 00:05:34.560
Reviving dead leads, which I call blast from the past.

s82
00:05:34.560 --> 00:05:40.840
Inbound replies, and figuring out before we waste an hour whether a company is even a fit.

s83
00:05:40.840 --> 00:05:45.480
Nine jobs, one operator who never sleeps.

s84
00:05:45.480 --> 00:05:47.480
Where does all this live?

s85
00:05:47.480 --> 00:05:48.960
One cursor tab.

s86
00:05:48.960 --> 00:05:50.400
That's genuinely it.

s87
00:05:50.400 --> 00:05:53.840
You type and Eira reaches out with a dozen hands.

s88
00:05:53.840 --> 00:06:03.240
Searches the knowledge base, reads the inbox, drafts the email, builds the code, and then shows you before anything actually goes out.

s89
00:06:03.240 --> 00:06:06.480
Under the hood, it's genuinely a real stack.

s90
00:06:06.480 --> 00:06:08.960
Not a demo held together with duct tape.

s91
00:06:08.960 --> 00:06:13.480
Databases for vectors, for relationship graph, for the CRM.

s92
00:06:13.480 --> 00:06:17.520
Three different model providers, each picked for the job it's actually best for.

s93
00:06:17.520 --> 00:06:23.360
Tools for Google, for swallowing documents, for every communication channel, plus monitoring.

s94
00:06:23.360 --> 00:06:26.400
So, we can see what it's thinking.

s95
00:06:26.400 --> 00:06:32.960
All of it, every capability exposed as 213 tools over one protocol.

s96
00:06:32.960 --> 00:06:38.720
And the golden rule, the one we never break, Eira drafts, human sends.

s97
00:06:38.720 --> 00:06:40.480
Now, memory.

s98
00:06:40.480 --> 00:06:46.760
And this is the part where most AI quietly lies to you, because a raw language model is basically a goldfish.

s99
00:06:46.760 --> 00:06:52.520
Brilliant for about 30 seconds, and then you close the tab and forgets you ever existed.

s100
00:06:52.520 --> 00:06:55.800
So, we engineered memory on purpose, in layers.

s101
00:06:55.800 --> 00:06:58.560
Working memory for the last few minutes.

s102
00:06:58.560 --> 00:07:02.280
Pinned facts about someone who who is.

s103
00:07:02.280 --> 00:07:05.760
Episodes, whole conversations as little stories.

s104
00:07:05.760 --> 00:07:10.760
Relationships with warmth that grows from stranger to trusted.

s105
00:07:10.760 --> 00:07:12.919
And a bouncer at the door.

s106
00:07:12.919 --> 00:07:19.240
A salience gate that decides what's even worth remembering, so the brain doesn't fill up with junk.

s107
00:07:19.240 --> 00:07:22.600
When two facts disagree, corrections win.

s108
00:07:22.600 --> 00:07:26.080
Continuity without making things up.

s109
00:07:26.080 --> 00:07:28.760
And then, I genuinely love this part.

s110
00:07:28.760 --> 00:07:32.240
At night, it dreams.

s111
00:07:32.320 --> 00:07:35.200
Every night, Eira runs a sleep cycle.

s112
00:07:35.200 --> 00:07:45.840
It replays the day, locks in useful stuff, hunts for contradictions, gently forgets the stale junk, and turns the day's work into reusable skills.

s113
00:07:45.840 --> 00:07:49.920
In the morning, there's a little dream report waiting for me to read.

s114
00:07:49.920 --> 00:07:51.400
Here's what I consolidated.

s115
00:07:51.400 --> 00:07:54.240
Here's what I Here's what I let go of.

s116
00:07:54.240 --> 00:07:56.600
Here's what I figured out while you were asleep.

s117
00:07:56.600 --> 00:08:01.040
The thing literally gets smarter overnight.

s118
00:08:01.040 --> 00:08:03.280
And here's the part I care about the most.

s119
00:08:03.280 --> 00:08:05.880
Every agent has a conscience.

s120
00:08:05.880 --> 00:08:12.040
And it is emphatically not to be helpful, be harmless.

s121
00:08:12.040 --> 00:08:20.040
It's a soul file written from the principles of a Jain family business that's been doing this for the last three generations.

s122
00:08:20.040 --> 00:08:23.080
Five old ideas turn into engineering rules.

s123
00:08:23.080 --> 00:08:25.520
No single source has the whole truth.

s124
00:08:25.520 --> 00:08:28.280
So, cross-check before you speak.

s125
00:08:28.280 --> 00:08:30.760
Never say things absolutely.

s126
00:08:30.760 --> 00:08:33.200
Cite the document and the date.

s127
00:08:33.200 --> 00:08:35.680
Do your own job, not someone else's.

s128
00:08:35.680 --> 00:08:38.360
Report the truth even when the truth is ugly.

s129
00:08:38.360 --> 00:08:40.120
And nobody works alone.

s130
00:08:40.120 --> 00:08:44.520
Ancient philosophy running as guardrails in production.

s131
00:08:44.520 --> 00:08:46.000
Now, let's talk money.

s132
00:08:46.000 --> 00:08:50.480
Because this is the part that should make the whole industry a little uncomfortable.

s133
00:08:50.480 --> 00:08:51.960
There was no training bill.

s134
00:08:51.960 --> 00:08:53.000
Zero.

s135
00:08:53.000 --> 00:08:54.839
The expensive part was never compute.

s136
00:08:54.839 --> 00:08:57.600
It was teaching a company to remember itself.

s137
00:08:57.600 --> 00:09:01.400
An agency quoted us 230 grand to build this.

s138
00:09:01.400 --> 00:09:03.560
We built it for around 30.

s139
00:09:03.560 --> 00:09:05.880
That's cheaper than a nice watch.

s140
00:09:05.880 --> 00:09:09.839
And it runs on a couple of thousand dollars a month.

s141
00:09:09.839 --> 00:09:11.360
So, here's the move.

s142
00:09:11.360 --> 00:09:14.839
We pulled the whole architecture out and made it forkable.

s143
00:09:14.839 --> 00:09:16.600
We call it Brain OS.

s144
00:09:16.600 --> 00:09:18.920
It ships as an empty nervous system.

s145
00:09:18.920 --> 00:09:22.160
The agents, the memory, the dream cycle, the soul file.

s146
00:09:22.160 --> 00:09:24.320
All there, completely blank.

s147
00:09:24.320 --> 00:09:28.560
You pour your own company's truth into it and from inside out.

s148
00:09:28.560 --> 00:09:32.800
Because here's the thing nobody can outsource for you.

s149
00:09:32.800 --> 00:09:36.800
Only you can build your company's brain.

s150
00:09:37.040 --> 00:09:40.120
We are a 100 people factory with no data scientists.

s151
00:09:40.120 --> 00:09:42.840
If we can grow a brain, you can too.

s152
00:09:42.840 --> 00:09:44.360
We are not selling ours to you.

s153
00:09:44.360 --> 00:09:46.680
We are helping you build your own.

s154
00:09:46.680 --> 00:09:51.560
forkmybrain.org Go build something that remembers.

s155
00:09:51.560 --> 00:09:53.800
Thank you.