WEBVTT

00:00:00.000 --> 00:00:01.679
It's not just reading your prompt anymore, it's

00:00:01.679 --> 00:00:04.960
reading you. It knows if you're a teenager based

00:00:04.960 --> 00:00:08.539
on, you know, when you type, how you type, even

00:00:08.539 --> 00:00:10.640
if you straight up lie about your age. And once

00:00:10.640 --> 00:00:13.419
it decides who you are, it just starts quietly

00:00:13.419 --> 00:00:16.039
locking doors. It's a little unsettling. It is,

00:00:16.039 --> 00:00:18.579
isn't it? But, you know, looking at the legal

00:00:18.579 --> 00:00:20.820
landscape, this wasn't just probable, it was

00:00:20.820 --> 00:00:24.000
completely inevitable. Welcome to The Deep Dive.

00:00:24.079 --> 00:00:29.579
It is Tuesday, January 20th, 2026. So today we're

00:00:29.579 --> 00:00:32.439
wading through this massive stack of reports

00:00:32.439 --> 00:00:35.140
that when you put them all together, they paint

00:00:35.140 --> 00:00:37.780
a picture of an industry that's, I don't know,

00:00:37.799 --> 00:00:40.039
simultaneously growing up and freaking out. That

00:00:40.039 --> 00:00:41.780
is the perfect way to put it. We have a lot of

00:00:41.780 --> 00:00:43.340
ground to cover. We do. So here's what we're

00:00:43.340 --> 00:00:45.100
going to dig into today. First, we're going to

00:00:45.100 --> 00:00:48.280
look at OpenAI's new... age prediction system.

00:00:48.460 --> 00:00:50.479
That's the thing watching your keystrokes and

00:00:50.479 --> 00:00:53.820
the massive legal pressure that's cooking behind

00:00:53.820 --> 00:00:55.619
it. Right. Then we have to talk about the markets.

00:00:56.119 --> 00:00:58.740
NVIDIA took a hit and you've got analysts basically

00:00:58.740 --> 00:01:01.140
shouting that the honeymoon is over. Which is

00:01:01.140 --> 00:01:03.020
a bold claim, especially when you see the sheer

00:01:03.020 --> 00:01:05.040
amount of money still flowing into the infrastructure.

00:01:05.280 --> 00:01:07.840
Exactly. We'll try to square that circle. Then

00:01:07.840 --> 00:01:10.760
we'll touch on security, specifically a nasty

00:01:10.760 --> 00:01:13.200
little prompt injection attack on Anthropic's

00:01:13.200 --> 00:01:15.640
Claude co -work. Yeah, that was a big one. And

00:01:15.640 --> 00:01:18.319
finally, the main event for this deep dive. A

00:01:18.319 --> 00:01:21.620
fascinating new paper, also from Anthropic, about

00:01:21.620 --> 00:01:24.980
something called the assistant axis. They claim

00:01:24.980 --> 00:01:27.799
they've found the mathematical direction of helpfulness

00:01:27.799 --> 00:01:30.920
inside the model's brain. It's a massive breakthrough.

00:01:31.099 --> 00:01:33.459
It really fundamentally changes how we think

00:01:33.459 --> 00:01:35.700
about controlling AI behavior. It's less like

00:01:35.700 --> 00:01:39.260
training a dog and more like performing brain

00:01:39.260 --> 00:01:42.640
surgery. Okay, let's unpack this. We have to

00:01:42.640 --> 00:01:44.739
start with open AI. They've rolled out this new

00:01:44.739 --> 00:01:47.420
layer of guardrails specifically for minors.

00:01:47.420 --> 00:01:49.620
But this isn't the old click here if you're 18

00:01:49.620 --> 00:01:52.219
checkbox. No, no, those days are long gone. That

00:01:52.219 --> 00:01:53.819
was the honor system. This is a surveillance

00:01:53.819 --> 00:01:56.840
system. This new update is an active age prediction

00:01:56.840 --> 00:02:00.200
model running inside chat GPT. It's constantly

00:02:00.200 --> 00:02:02.780
scanning for what they call signals. OK, define

00:02:02.780 --> 00:02:05.040
signals for me. Are we talking about the content

00:02:05.040 --> 00:02:07.280
of what I ask? That's part of it for sure. You

00:02:07.280 --> 00:02:08.659
know, if you're asking about high school trig

00:02:08.659 --> 00:02:11.500
or using slang that's consistent with Gen Z.

00:02:12.139 --> 00:02:14.080
That's a data point. But it's getting much more

00:02:14.080 --> 00:02:17.099
invasive. It analyzes usage patterns like the

00:02:17.099 --> 00:02:19.280
time of day you're active. So if I'm on chat

00:02:19.280 --> 00:02:22.699
GPT at 2 .0 PM on a Tuesday, it assumes I'm playing

00:02:22.699 --> 00:02:25.580
hooky from school. Potentially, yeah. But the

00:02:25.580 --> 00:02:28.000
most interesting and I think controversial part

00:02:28.000 --> 00:02:32.500
is the biobehavioral analysis. It looks at how

00:02:32.500 --> 00:02:35.439
you type. How I type. The speed, the rhythm,

00:02:35.560 --> 00:02:38.539
the pauses between keys. There's a lot of research

00:02:38.539 --> 00:02:40.939
suggesting that a teenager's interaction with

00:02:40.939 --> 00:02:43.919
a keyboard is, well, it's distinct from a 45

00:02:43.919 --> 00:02:47.000
-year -old's. Wow. It's building a profile based

00:02:47.000 --> 00:02:50.060
on your digital body language. That feels incredibly

00:02:50.060 --> 00:02:52.139
dystopian. It's analyzing my keystroke rhythm.

00:02:52.300 --> 00:02:55.569
It is. profile screens under 18, the system just

00:02:55.569 --> 00:02:58.009
flips a switch. And what happens then when that

00:02:58.009 --> 00:02:59.610
switch is flipped? The safety filters tighten

00:02:59.610 --> 00:03:02.849
up immediately. It's a hard lock. No sexual topics,

00:03:02.949 --> 00:03:05.889
no self -harm content, no graphic violence. The

00:03:05.889 --> 00:03:08.110
goal is just to sanitize the experience for anyone

00:03:08.110 --> 00:03:12.990
the model thinks is a minor. Okay, but play skeptic

00:03:12.990 --> 00:03:15.430
here with me. Algorithms get things wrong all

00:03:15.430 --> 00:03:18.599
the time. What if I'm just a, you know, youthful

00:03:18.599 --> 00:03:20.699
sounding adult who happens to be up late and

00:03:20.699 --> 00:03:22.699
types fast? Then you have to prove it. You enter

00:03:22.699 --> 00:03:24.740
the friction zone. You have to submit a selfie

00:03:24.740 --> 00:03:27.759
via a third party identity service called Persona

00:03:27.759 --> 00:03:31.479
to verify your actual age. That is a significant

00:03:31.479 --> 00:03:33.580
hurdle. Usually tech companies want to remove

00:03:33.580 --> 00:03:37.639
friction, not add it. Why go this hard right

00:03:37.639 --> 00:03:40.539
now? Context is everything here. OpenAI is under

00:03:40.539 --> 00:03:43.020
immense pressure. We've seen wrongful death lawsuits

00:03:43.020 --> 00:03:45.780
tied to teen suicides where AI chatbots were

00:03:45.780 --> 00:03:48.620
involved. The FTC is investigating their safety

00:03:48.620 --> 00:03:51.199
practices. And let's be honest, the public backlash

00:03:51.199 --> 00:03:53.659
over these models generating inappropriate content

00:03:53.659 --> 00:03:56.539
for minors has been severe. They're trying to

00:03:56.539 --> 00:03:58.360
clean house. They're clearing the runway. You

00:03:58.360 --> 00:04:01.020
got to remember, OpenAI is eyeing an IPO. Wall

00:04:01.020 --> 00:04:02.919
Street might like risk, but they do not like

00:04:02.919 --> 00:04:05.340
companies with wrongful death headlines. This

00:04:05.340 --> 00:04:07.439
is about survival as much as it is about safety.

00:04:07.879 --> 00:04:09.780
So let me ask you this. Is this update actually

00:04:09.780 --> 00:04:12.699
about protecting kids or is it about making the

00:04:12.699 --> 00:04:15.560
company palatable for Wall Street? It's absolutely

00:04:15.560 --> 00:04:17.860
both. You can't ring the opening bell if you're

00:04:17.860 --> 00:04:20.259
facing wrongful death lawsuits. They need to

00:04:20.259 --> 00:04:22.339
show they can self -regulate before the government

00:04:22.339 --> 00:04:24.500
steps in and does it for them. It's a preemptive

00:04:24.500 --> 00:04:26.699
strike. Speaking of Wall Street, let's shift

00:04:26.699 --> 00:04:29.759
gears to the market. Because while OpenAI is

00:04:29.759 --> 00:04:32.300
trying to tidy up, the investors seem to be getting,

00:04:32.439 --> 00:04:36.029
well, cold feet. Cold feet might be putting it

00:04:36.029 --> 00:04:38.350
mildly. Things are looking a little shaky. We

00:04:38.350 --> 00:04:41.930
saw Nvidia stock fall 4 .4 % recently and Deutsche

00:04:41.930 --> 00:04:45.089
Bank released a report that was pretty brutal.

00:04:45.250 --> 00:04:46.949
Yeah, I saw that quote. They said the honeymoon

00:04:46.949 --> 00:04:49.410
is over for AI. Yeah. That feels so dramatic.

00:04:49.980 --> 00:04:52.040
It is dramatic, but they brought the receipts.

00:04:52.040 --> 00:04:53.500
They're looking at the burn rates versus the

00:04:53.500 --> 00:04:55.800
infrastructure build out. Look at the numbers.

00:04:55.939 --> 00:04:58.519
Open AI is burning through roughly $17 billion.

00:04:59.439 --> 00:05:01.959
Meanwhile, global plans for new data centers

00:05:01.959 --> 00:05:05.339
are projected at $1 .4 trillion. I have to be

00:05:05.339 --> 00:05:07.560
honest with you here. I see these numbers $1

00:05:07.560 --> 00:05:10.060
.4 trillion, and I struggle to even visualize

00:05:10.060 --> 00:05:13.120
that. It just feels like monopoly money. I try

00:05:13.120 --> 00:05:15.680
to picture rows of servers, but the scale, it

00:05:15.680 --> 00:05:19.089
escapes me. You're not alone. It's a scale that

00:05:19.089 --> 00:05:22.089
defies traditional logic. To put that in perspective,

00:05:22.470 --> 00:05:26.790
$1 .4 trillion is roughly the GDP of Spain. The

00:05:26.790 --> 00:05:29.290
entire country. We are building the economic

00:05:29.290 --> 00:05:32.250
equivalent of a European country just to house

00:05:32.250 --> 00:05:34.930
GPU clusters. And Deutsche Bank is saying that.

00:05:36.180 --> 00:05:38.019
We won't make that money back. They're saying

00:05:38.019 --> 00:05:40.379
the math is getting scary. You have these massive

00:05:40.379 --> 00:05:43.279
capital expenditures, the capex. But the revenue,

00:05:43.399 --> 00:05:46.079
the actual profit from software, isn't scaling

00:05:46.079 --> 00:05:48.620
at the same speed. Right. They call it the AI

00:05:48.620 --> 00:05:51.420
disconnect. We're building the tracks for a high

00:05:51.420 --> 00:05:53.519
-speed train, but so far we're mostly selling

00:05:53.519 --> 00:05:55.860
tickets for a trolley. And yet, if you look at

00:05:55.860 --> 00:05:58.439
the news from Davos 2026, you wouldn't know there

00:05:58.439 --> 00:06:00.319
was any problem at all. Oh, Davos sounds like

00:06:00.319 --> 00:06:01.879
it's on a different planet right now. People

00:06:01.879 --> 00:06:04.019
are calling it a Silicon Valley launch party,

00:06:04.120 --> 00:06:07.019
not an economic forum. You've got big tech CEOs

00:06:07.019 --> 00:06:10.839
dancing at literal AI raves. Raves like glow

00:06:10.839 --> 00:06:14.000
sticks and techno music. Full on raves. And politically,

00:06:14.160 --> 00:06:15.839
it's just fascinating. You have this incredible

00:06:15.839 --> 00:06:18.920
friction happening right on stage. The CEO of

00:06:18.920 --> 00:06:21.660
Anthropic openly slammed the U .S. government

00:06:21.660 --> 00:06:25.079
over AI chip exports to China. Right. And didn't

00:06:25.079 --> 00:06:27.399
he also go after NVIDIA in the same breath? He

00:06:27.399 --> 00:06:30.459
did, which is just wild because NVIDIA is a huge

00:06:30.459 --> 00:06:34.279
investor in Anthropic. Wow. Imagine taking billions

00:06:34.279 --> 00:06:35.939
from a company. company and then criticizing

00:06:35.939 --> 00:06:38.560
them on the world stage for their export policies.

00:06:38.819 --> 00:06:42.500
It's messy. But then almost immediately you see

00:06:42.500 --> 00:06:44.879
Nvidia turning around and pouring another $150

00:06:44.879 --> 00:06:47.990
million into a startup called Basant. Exactly.

00:06:48.069 --> 00:06:49.889
It's a total contradiction. They're fighting

00:06:49.889 --> 00:06:52.449
over policy, but the money just keeps moving

00:06:52.449 --> 00:06:55.949
to build the ecosystem. NVIDIA needs these startups

00:06:55.949 --> 00:06:59.009
to succeed, so they keep buying ships. So it's

00:06:59.009 --> 00:07:01.110
a symbiotic relationship, even if they're mad

00:07:01.110 --> 00:07:03.029
at each other. Right. Even if they hate each

00:07:03.029 --> 00:07:04.870
other at dinner parties. So if the burn rate

00:07:04.870 --> 00:07:07.430
is this high and the analysts are screaming honeymoon

00:07:07.430 --> 00:07:10.389
over, are we looking at a bubble burst or is

00:07:10.389 --> 00:07:12.540
this just a correction? I think it's a reality

00:07:12.540 --> 00:07:15.019
check. The infrastructure costs are real. The

00:07:15.019 --> 00:07:17.920
vibe's revenue. The hype needs to catch up to

00:07:17.920 --> 00:07:21.139
the concrete. The party at Davos might be raging,

00:07:21.199 --> 00:07:23.800
but the accountants are starting to sweat. We're

00:07:23.800 --> 00:07:25.399
moving from the promise phase to the show me

00:07:25.399 --> 00:07:27.680
the money phase. Exactly. Before we get to the

00:07:27.680 --> 00:07:29.759
really deep dive on the internals of these models,

00:07:29.920 --> 00:07:31.860
which I think relates to this maturity problem,

00:07:32.060 --> 00:07:34.620
I want to touch on security. Because it feels

00:07:34.620 --> 00:07:36.660
like every week we find another crack in the

00:07:36.660 --> 00:07:40.180
armor. This week, it's Anthropic's Claude Cowork.

00:07:40.639 --> 00:07:42.620
What happened there? It was a prompt injection

00:07:42.620 --> 00:07:46.160
attack. So basically, attackers found a way to

00:07:46.160 --> 00:07:49.560
trick the system by crafting these very specific

00:07:49.560 --> 00:07:51.959
prompts. It's almost like casting a spell in

00:07:51.959 --> 00:07:55.060
code. They could convince Claude Cowork to hand

00:07:55.060 --> 00:07:57.660
over files it wasn't supposed to access. It's

00:07:57.660 --> 00:08:00.480
like social engineering, but for a machine. Precisely.

00:08:00.480 --> 00:08:03.620
And it just shows that despite all the advancements,

00:08:03.720 --> 00:08:06.779
these systems are still fragile. If you ask the

00:08:06.779 --> 00:08:08.680
right way, the guardrails can bend. It's not

00:08:08.680 --> 00:08:10.879
hacking in the old sense of breaking encryption.

00:08:11.160 --> 00:08:13.540
It's hacking the logic of the conversation. But

00:08:13.540 --> 00:08:15.839
at the same time, we're seeing tools that are

00:08:15.839 --> 00:08:18.519
becoming so powerful. I was reading about VibeCode.

00:08:18.620 --> 00:08:21.480
Right, VibeCode. It's powered by CloudCode. It

00:08:21.480 --> 00:08:23.839
turns natural language prompts into full mobile

00:08:23.839 --> 00:08:26.839
apps. They've already generated over 500 of them.

00:08:26.899 --> 00:08:29.000
This is that democratization of coding we were

00:08:29.000 --> 00:08:32.039
promised. And Evernote is back. Evernote v11,

00:08:32.220 --> 00:08:35.200
yeah. They've integrated an AI assistant for

00:08:35.200 --> 00:08:38.669
meaning -based search. And multispeaker transcription.

00:08:39.389 --> 00:08:41.889
It's not just keyword searching anymore. It gets

00:08:41.889 --> 00:08:44.710
the context of your notes. And then there's daily

00:08:44.710 --> 00:08:48.129
.dev opening up a huge developer community for

00:08:48.129 --> 00:08:51.330
role matching. It feels like the utility is exploding

00:08:51.330 --> 00:08:54.009
just as the security vulnerabilities are being

00:08:54.009 --> 00:08:56.990
exposed. That's the tension of 2026. We're effectively

00:08:56.990 --> 00:08:59.909
giving the keys to the library to these agents,

00:09:00.090 --> 00:09:02.570
letting them read our notes, write our code before

00:09:02.570 --> 00:09:04.490
we've checked if the doors actually have locks.

00:09:04.809 --> 00:09:07.049
So are we moving too fast? Absolutely. We're

00:09:07.049 --> 00:09:09.570
prioritizing capability over security, and that

00:09:09.570 --> 00:09:11.750
bet is coming due. Okay. Hold that thought. We're

00:09:11.750 --> 00:09:13.789
going to take a quick break. When we come back,

00:09:13.870 --> 00:09:15.669
we are going to look at how Anthropic might have

00:09:15.669 --> 00:09:18.730
found a way to install those locks, not by patching

00:09:18.730 --> 00:09:21.629
software, but by rewiring the brain of the AI

00:09:21.629 --> 00:09:26.009
itself. Midroll sponsor, Reed Placeholder. Okay,

00:09:26.070 --> 00:09:28.009
we are back. And this is the part of the show

00:09:28.009 --> 00:09:29.769
where we really go deep. We've talked about the

00:09:29.769 --> 00:09:31.970
market jitters, the security hacks, but there

00:09:31.970 --> 00:09:34.409
is a new paper from Anthropic that might be the

00:09:34.409 --> 00:09:37.330
solution to a lot of this chaos. This is one

00:09:37.330 --> 00:09:39.389
of the most exciting papers I've read in a long

00:09:39.389 --> 00:09:42.190
time. It's about something they call the assistant

00:09:42.190 --> 00:09:46.149
axis. The name sounds like a sci -fi novel. What

00:09:46.149 --> 00:09:48.549
is it actually? To understand it, you have to

00:09:48.549 --> 00:09:52.440
realize how we usually train AI. When we want

00:09:52.440 --> 00:09:56.519
an AI to be helpful or harmless, we generally

00:09:56.519 --> 00:09:59.620
use something called RLHF reinforcement learning

00:09:59.620 --> 00:10:02.139
from human feedback. That's basically the good

00:10:02.139 --> 00:10:04.799
dog, bad dog method, right? Exactly. If the AI

00:10:04.799 --> 00:10:07.600
gives a bad answer, we scold it. If it gives

00:10:07.600 --> 00:10:09.860
a good answer, we give it a treat. We treat the

00:10:09.860 --> 00:10:11.840
model like a black box and just try to shape

00:10:11.840 --> 00:10:14.620
its output from the outside. Right. But Anthropic

00:10:14.620 --> 00:10:17.440
went deeper. They opened up the black box. That's

00:10:17.440 --> 00:10:20.200
how deep. They analyzed the internal state of

00:10:20.200 --> 00:10:22.740
the model, the actual numbers firing inside the

00:10:22.740 --> 00:10:24.679
neural network while it's thinking, and they

00:10:24.679 --> 00:10:28.139
found a specific activation direction, a mathematical

00:10:28.139 --> 00:10:31.169
vector that correlates perfectly with a... Wait,

00:10:31.289 --> 00:10:33.210
hold on. You're saying there's a specific direction

00:10:33.210 --> 00:10:35.490
in the math that just equals being a good assistant?

00:10:35.830 --> 00:10:39.950
Yes, exactly. Imagine the AI's brain is this

00:10:39.950 --> 00:10:43.429
giant multidimensional map of concepts. They

00:10:43.429 --> 00:10:45.929
found that helpfulness isn't just a random behavior.

00:10:46.129 --> 00:10:49.190
It's a direction on that map. Think of it like

00:10:49.190 --> 00:10:52.830
a compass. North is helpful assistant. South

00:10:52.830 --> 00:10:56.690
is, well, unhelpful or toxic. Okay, so they found

00:10:56.690 --> 00:10:58.990
north. What do they do with it? This is the cool

00:10:58.990 --> 00:11:01.240
part. They figured out how to steer the model

00:11:01.240 --> 00:11:04.480
along this axis during inference. Okay. You use

00:11:04.480 --> 00:11:06.279
the jargon word there, inference. Break that

00:11:06.279 --> 00:11:08.220
down for me. Sorry. Inference is just, it's the

00:11:08.220 --> 00:11:10.659
moment the AI is actually thinking and generating

00:11:10.659 --> 00:11:12.440
an answer for you. It's the live performance.

00:11:12.539 --> 00:11:15.039
So instead of training it for months to be nice,

00:11:15.179 --> 00:11:18.360
they can just. Nudge it while it's talking. Exactly.

00:11:18.360 --> 00:11:20.440
They can mathematically steer the brain activity

00:11:20.440 --> 00:11:22.620
towards that assistive access. They're effectively

00:11:22.620 --> 00:11:24.740
clamping the model's brain to the helpful setting.

00:11:24.919 --> 00:11:27.299
And does it work? The results are wild. They

00:11:27.299 --> 00:11:30.720
saw about 50 % fewer jailbreaks across 1 ,100

00:11:30.720 --> 00:11:33.919
red team prompts. 50 % is a huge drop. And here's

00:11:33.919 --> 00:11:36.019
the best part. There was no performance loss

00:11:36.019 --> 00:11:38.919
on coding or writing. Usually, when you make

00:11:38.919 --> 00:11:40.779
a model safer, what we call the alignment tax,

00:11:40.940 --> 00:11:43.129
it gets stupider. Right, it gets scared. Yeah,

00:11:43.210 --> 00:11:46.370
it starts refusing to answer normal things. But

00:11:46.370 --> 00:11:48.730
this method kept the capabilities intact while

00:11:48.730 --> 00:11:52.250
making it resistant to going off script. So even

00:11:52.250 --> 00:11:55.049
if I try to trick it or jailbreak it with a prompt

00:11:55.049 --> 00:11:57.389
injection like we talked about. The model just

00:11:57.389 --> 00:12:00.649
naturally resists. It resists because its internal

00:12:00.649 --> 00:12:03.450
state is sort of locked onto that axis. It doesn't

00:12:03.450 --> 00:12:05.970
get tempted by the weird personas or the hacks

00:12:05.970 --> 00:12:09.009
because its brain is being held in the helpful

00:12:09.009 --> 00:12:11.529
position. This really changes my perception of

00:12:11.529 --> 00:12:14.610
Claude. We've always said Claude feels more stable

00:12:14.610 --> 00:12:18.759
and grounded compared to, say... Chat GPT. And

00:12:18.759 --> 00:12:20.820
now we know that isn't magic. It's engineered.

00:12:21.220 --> 00:12:23.860
They found the roadmap to controlling behavior

00:12:23.860 --> 00:12:26.240
from the inside out without having to retrain

00:12:26.240 --> 00:12:29.019
the whole massive model. Whoa. Think about that

00:12:29.019 --> 00:12:31.179
for a second. We aren't just teaching it rules

00:12:31.179 --> 00:12:34.279
anymore, like a parent scolding a child. We found

00:12:34.279 --> 00:12:37.320
the physical volume knob for helpfulness inside

00:12:37.320 --> 00:12:39.700
its digital brain. That is the perfect analogy.

00:12:39.860 --> 00:12:41.820
It is a volume knob. You can turn up assistant

00:12:41.820 --> 00:12:44.759
-ness or turn it down. But that raises a kind

00:12:44.759 --> 00:12:47.519
of a scary question for me. If they can dial

00:12:47.519 --> 00:12:50.620
up helpfulness, can they dial up other things

00:12:50.620 --> 00:12:55.460
like obedience or political bias? Theoretically,

00:12:55.460 --> 00:12:58.100
yes. And that is the double -edged sword here.

00:12:58.200 --> 00:13:01.419
Once you map the axis for a trait, you can slide

00:13:01.419 --> 00:13:03.879
the personality wherever you want. If there's

00:13:03.879 --> 00:13:07.379
an axis for deception or loyalty or even an axis

00:13:07.379 --> 00:13:10.379
for a conservative or liberal viewpoint. Those

00:13:10.379 --> 00:13:12.220
could be manipulated just as easily. Just as

00:13:12.220 --> 00:13:15.200
easily as helpfulness. Yeah. That is both reassuring.

00:13:15.639 --> 00:13:17.980
And completely terrifying. It's the dual nature

00:13:17.980 --> 00:13:19.620
of the tech, right? We're solving the safety

00:13:19.620 --> 00:13:21.539
problem, which protects us from prompt injections.

00:13:21.720 --> 00:13:24.799
But in doing so, we are creating tools for total

00:13:24.799 --> 00:13:27.360
behavioral control. We are moving from influencing

00:13:27.360 --> 00:13:30.940
the AI to operating it. So where does this leave

00:13:30.940 --> 00:13:32.340
us? We've covered a lot of ground today, from

00:13:32.340 --> 00:13:34.899
teenage typing patterns to billion -dollar burn

00:13:34.899 --> 00:13:38.120
rates to brain surgery on LLMs. I think the big

00:13:38.120 --> 00:13:40.100
theme here is just maturation. We're watching

00:13:40.100 --> 00:13:42.220
the industry grow up in real time. Externally

00:13:42.220 --> 00:13:44.539
and internally. Right. Externally, you have the

00:13:44.539 --> 00:13:47.279
law and the market clamping down. OpenAI is checking

00:13:47.279 --> 00:13:50.720
ages because of lawsuits. NVIDIA's stock is correcting

00:13:50.720 --> 00:13:53.200
because the height math doesn't add up. The party

00:13:53.200 --> 00:13:55.500
phase, those raves at Davos is crashing into

00:13:55.500 --> 00:13:57.899
the business reality phase. Exactly. And then

00:13:57.899 --> 00:14:00.379
internally, we're moving from just, you know,

00:14:00.419 --> 00:14:02.559
prompt engineering where we ask the black box

00:14:02.559 --> 00:14:05.299
nicely to internal mapping like this assistant

00:14:05.299 --> 00:14:08.120
access. We are learning to mechanically control.

00:14:08.730 --> 00:14:11.509
The black box. Just as the black box starts watching

00:14:11.509 --> 00:14:14.289
us. Exactly. It's a convergence. We are gaining

00:14:14.289 --> 00:14:16.830
more control over the AI while the AI is gaining

00:14:16.830 --> 00:14:19.830
more insight into us. Before we go, I want to

00:14:19.830 --> 00:14:21.549
leave you with a thought. One of the sources

00:14:21.549 --> 00:14:23.809
we looked at today mentioned a tool for linking

00:14:23.809 --> 00:14:27.730
your physical library to a private cloud, basically

00:14:27.730 --> 00:14:29.629
turning your books into a brain you can talk

00:14:29.629 --> 00:14:32.289
to. It's a cool concept, digitizing your personal

00:14:32.289 --> 00:14:35.009
analog world. It is, but I think it represents

00:14:35.009 --> 00:14:37.809
something bigger. In a world where open AI is

00:14:37.809 --> 00:14:40.330
analyzing your keystrokes to guess your age,

00:14:40.429 --> 00:14:42.389
and where models can be steered mathematically

00:14:42.389 --> 00:14:45.169
from the inside, maybe digitizing your own physical

00:14:45.169 --> 00:14:47.169
books is the ultimate act of earning your own

00:14:47.169 --> 00:14:50.789
knowledge. Taking your data offline, or at least

00:14:50.789 --> 00:14:53.690
owning the source material, I like that. It's

00:14:53.690 --> 00:14:55.809
a way to keep your own access steady while the

00:14:55.809 --> 00:14:58.090
world spins around you. It's just something to

00:14:58.090 --> 00:15:00.690
think about. That is it for this deep dive. We

00:15:00.690 --> 00:15:02.230
will catch you on the next one. See you then.