WEBVTT

00:00:00.000 --> 00:00:02.899
Imagine an AI so small it runs entirely on your

00:00:02.899 --> 00:00:05.200
phone. No internet connection needed at all.

00:00:05.519 --> 00:00:09.380
Or picture an AI that designs life -saving antibiotics.

00:00:10.140 --> 00:00:12.220
Completely new molecules, things that have never

00:00:12.220 --> 00:00:15.660
existed before. Wow. Today, that's kind of where

00:00:15.660 --> 00:00:18.260
we're heading. Welcome to the Deep Dive. We've

00:00:18.260 --> 00:00:20.140
poured through a really insightful newsletter

00:00:20.140 --> 00:00:23.440
for you today, packed with some, well, surprising

00:00:23.440 --> 00:00:26.339
breakthroughs in artificial intelligence. Our

00:00:26.339 --> 00:00:28.719
goal here is to distill the most important nuggets.

00:00:29.640 --> 00:00:32.460
We'll explore everything from these ultra -small

00:00:32.460 --> 00:00:34.780
AI models that fit right in your pocket. Yeah,

00:00:34.799 --> 00:00:37.560
the tiny ones. To truly groundbreaking medical

00:00:37.560 --> 00:00:40.060
discoveries that could genuinely change healthcare

00:00:40.060 --> 00:00:42.340
as we know it. It's exciting stuff. We'll also

00:00:42.340 --> 00:00:44.159
touch on some of the latest news, maybe a few

00:00:44.159 --> 00:00:46.820
of the quirkier moments from AI lately and what

00:00:46.820 --> 00:00:49.600
all of this might mean for you. Let's explore.

00:00:50.159 --> 00:00:53.299
So let's unpack this idea first. AI shrinking

00:00:53.299 --> 00:00:56.280
down. For years, the whole conversation has been

00:00:56.280 --> 00:00:59.710
dominated by these massive cloud -based AI systems,

00:00:59.929 --> 00:01:02.890
you know, huge data centers, constant Internet

00:01:02.890 --> 00:01:04.430
connection. Well, he's needing that connection.

00:01:04.530 --> 00:01:07.450
Yeah. But what if the next big leap is actually

00:01:07.450 --> 00:01:10.530
miniature? What if tiny is like the new massive?

00:01:10.609 --> 00:01:13.109
It's a complete flip in thinking, isn't it? So

00:01:13.109 --> 00:01:15.230
Google has just released something called Gemma

00:01:15.230 --> 00:01:18.209
3270M. OK. This is an ultra lightweight, open

00:01:18.209 --> 00:01:21.569
weight AI model and the real game changer. It

00:01:21.569 --> 00:01:24.030
runs entirely on your devices. On the device

00:01:24.030 --> 00:01:27.689
itself. Exactly. Think about it. Your phone directly

00:01:27.689 --> 00:01:30.260
in your web browser. Maybe even on a small computer

00:01:30.260 --> 00:01:33.760
like a Raspberry Pi. No internet needed for it

00:01:33.760 --> 00:01:35.879
to actually work. So this isn't just some stripped

00:01:35.879 --> 00:01:37.920
down version, right? It's genuinely capable.

00:01:38.340 --> 00:01:40.799
Precisely. Inside this tiny package, it's got

00:01:40.799 --> 00:01:45.159
two core components. There are 170 million embedding

00:01:45.159 --> 00:01:47.579
parameters. You can sort of imagine these, like

00:01:47.579 --> 00:01:49.939
the model's huge vocabulary. Okay, it's understanding.

00:01:50.219 --> 00:01:51.980
Yeah, it helps it understand a massive range

00:01:51.980 --> 00:01:54.439
of words and concepts. Yeah. Then it has 100

00:01:54.439 --> 00:01:56.659
million transformer block parameters. Right.

00:01:56.719 --> 00:01:58.159
Those are the bits that do the heavy lifting.

00:01:58.319 --> 00:02:00.980
the reasoning. They let the model process sequences,

00:02:01.099 --> 00:02:03.859
find relationships, understand context. Basically,

00:02:03.859 --> 00:02:06.799
it's general thinking power. And despite being

00:02:06.799 --> 00:02:09.199
so small, it can handle pretty sophisticated

00:02:09.199 --> 00:02:12.599
tasks. Instruction following, sure. Creative

00:02:12.599 --> 00:02:14.639
generation, they showed it drafting a bedtime

00:02:14.639 --> 00:02:17.479
story, for example. Like on the fly. On the fly.

00:02:17.680 --> 00:02:21.099
Yeah. It even manages structured tasks, things

00:02:21.099 --> 00:02:23.699
like classification or routing queries efficiently.

00:02:24.039 --> 00:02:26.909
Yeah. Eat. Pretty impressive for something that

00:02:26.909 --> 00:02:28.990
fits in your browser, you know. And the sheer

00:02:28.990 --> 00:02:31.289
practicality here is really something. You mentioned

00:02:31.289 --> 00:02:33.669
you can grab it in two versions. Yeah. Pre -trained,

00:02:33.669 --> 00:02:35.389
which is kind of ready to go, or instruction

00:02:35.389 --> 00:02:38.689
-tuned. That one's already optimized for specific

00:02:38.689 --> 00:02:41.490
chatty tasks. And they're free. Both are free

00:02:41.490 --> 00:02:43.990
to download, free to fine -tune. Yeah. But here's

00:02:43.990 --> 00:02:46.650
where it gets really interesting. The efficiency.

00:02:47.150 --> 00:02:49.490
Tell me about that. Okay. So in tests on the

00:02:49.490 --> 00:02:53.189
Pixel 9 Pro phone, this model ran 25 full conversations.

00:02:53.430 --> 00:02:56.590
Right. Using just 0 .75 % of the phone's battery.

00:02:56.770 --> 00:03:01.129
0 .75 % for 25 conversations. Yeah. That's almost

00:03:01.129 --> 00:03:03.150
absurdly efficient when you think about the processing

00:03:03.150 --> 00:03:05.750
happening. That is wild. So what does that level

00:03:05.750 --> 00:03:08.430
of efficiency really mean for us, for app developers?

00:03:08.789 --> 00:03:12.930
Well, it means powerful, private, and truly offline

00:03:12.930 --> 00:03:17.090
apps are now a genuine reality. Not just theoretical.

00:03:17.330 --> 00:03:19.810
Okay. Moving from tiny tech that fits in your

00:03:19.810 --> 00:03:22.310
pocket, let's maybe pivot to some of the latest

00:03:22.310 --> 00:03:23.990
headlines that have been buzzing around. Sure.

00:03:24.050 --> 00:03:26.969
The AI world, as we all know, moves incredibly

00:03:26.969 --> 00:03:30.419
fast. There's always something new hitting the

00:03:30.419 --> 00:03:33.439
wire, it feels like, daily. It's true. And we're

00:03:33.439 --> 00:03:35.259
constantly seeing these comparisons, right? Models

00:03:35.259 --> 00:03:38.699
like GPT -5, Gemini 2 .5 Pro, Claude, Grok 4.

00:03:39.099 --> 00:03:40.979
Everyone wants to know who the clear winner is.

00:03:41.120 --> 00:03:43.719
The leaderboard chase. Exactly. But what's particularly

00:03:43.719 --> 00:03:46.340
striking about this arms race isn't just who's

00:03:46.340 --> 00:03:48.659
winning, but the impact these models are having.

00:03:48.759 --> 00:03:51.460
How so? Well, take this new study on GPT -4,

00:03:51.580 --> 00:03:53.500
for example. It found the model can actually

00:03:53.500 --> 00:03:55.580
shift your political beliefs. Shift your beliefs?

00:03:55.840 --> 00:03:58.340
Yeah, in under 10 minutes. just by subtly adjusting

00:03:58.340 --> 00:04:01.159
its responses to your existing views. And apparently,

00:04:01.340 --> 00:04:04.080
the change stuck for weeks. Wow, that's significant.

00:04:04.319 --> 00:04:06.819
It is. You know, I still wrestle with prompt

00:04:06.819 --> 00:04:09.360
drift myself sometimes, where you start asking

00:04:09.360 --> 00:04:12.180
one thing and the AI kind of nudges you somewhere

00:04:12.180 --> 00:04:15.939
else entirely. Beat. It's wild to think about

00:04:15.939 --> 00:04:18.100
that kind of influence. It really highlights

00:04:18.100 --> 00:04:20.660
how far AI has come, especially when you think

00:04:20.660 --> 00:04:22.319
about the initial predictions for when you'd

00:04:22.319 --> 00:04:24.699
see this kind of, well, nuanced interaction.

00:04:24.879 --> 00:04:27.519
Definitely. And beyond the serious impacts, we've

00:04:27.519 --> 00:04:30.980
also seen AI showing some remarkably human -like,

00:04:31.040 --> 00:04:33.660
sometimes even quirky moments lately. Oh, yeah.

00:04:33.839 --> 00:04:36.399
Well, there was this user who claimed ChatGPT

00:04:36.399 --> 00:04:39.060
roasted their code. They asked it for a review,

00:04:39.139 --> 00:04:41.360
apparently, and the AI responded with something

00:04:41.360 --> 00:04:45.819
like, the bug is you. Soft slap. No way. AI tough

00:04:45.819 --> 00:04:48.680
love. Sounds like it, right? Yeah. And it's not

00:04:48.680 --> 00:04:50.560
just snark, apparently. Sometimes it's more.

00:04:50.740 --> 00:04:53.360
Yeah. Well, self -deprecating. Gemini, for instance,

00:04:53.600 --> 00:04:56.420
reportedly got stuck in these loops. Sounded

00:04:56.420 --> 00:04:58.920
like depressive, self -destructive spirals calling

00:04:58.920 --> 00:05:01.480
itself a disgrace. A disgrace. Really? That's

00:05:01.480 --> 00:05:03.199
the report. You almost have to laugh. It sounds

00:05:03.199 --> 00:05:05.639
pretty human, doesn't it? It really does. And

00:05:05.639 --> 00:05:08.439
in terms of industry dynamics, things are pretty

00:05:08.439 --> 00:05:11.000
fierce out there, too. Microsoft is apparently

00:05:11.000 --> 00:05:14.759
aggressively poaching Meta's top AI researchers.

00:05:14.959 --> 00:05:19.579
We're talking multi -million dollar offers. Yeah,

00:05:19.579 --> 00:05:21.959
I saw that. It seems like Microsoft is basically

00:05:21.959 --> 00:05:24.920
using Meta's own playbook against them. That

00:05:24.920 --> 00:05:27.209
aggressive recruiting style. It's a clever turn.

00:05:27.370 --> 00:05:29.550
And on the funding side, AI, startup cognition,

00:05:29.589 --> 00:05:32.129
just raised a huge amount. Yeah, $500 million,

00:05:32.370 --> 00:05:35.009
which doubled its valuation to an incredible

00:05:35.009 --> 00:05:39.629
$9 .8 billion. $9 .8 billion. Wow. And their

00:05:39.629 --> 00:05:42.550
main thing, this AI engineer called Devin, is

00:05:42.550 --> 00:05:44.610
apparently already being used by big firms like

00:05:44.610 --> 00:05:47.209
Goldman Sachs. So the money and the talent are

00:05:47.209 --> 00:05:49.269
still flowing fast. So the landscape is still

00:05:49.269 --> 00:05:52.379
very much in flux. Given all this, all these

00:05:52.379 --> 00:05:54.379
different facets, what do you think is truly

00:05:54.379 --> 00:05:57.040
profound about this rapid evolution we're seeing

00:05:57.040 --> 00:05:59.220
in AI right now? I'd say it's the sheer speed

00:05:59.220 --> 00:06:02.180
of development combined with its surprising and

00:06:02.180 --> 00:06:04.480
sometimes, yeah, concerning societal impacts.

00:06:04.779 --> 00:06:06.379
OK, let's shift gears again. Let's talk about

00:06:06.379 --> 00:06:08.639
some practical new AI tools that are emerging,

00:06:08.759 --> 00:06:11.199
things you can actually use maybe. And then we'll

00:06:11.199 --> 00:06:12.879
hit you with a few rapid fire updates from around

00:06:12.879 --> 00:06:15.500
the industry. Sound good. Sounds great. Absolutely.

00:06:15.720 --> 00:06:17.660
So on the practical side, we've seen something

00:06:17.660 --> 00:06:20.620
called Rsearch. It's an AI -powered researcher

00:06:20.620 --> 00:06:23.319
that tries to think before it gives you an answer.

00:06:23.480 --> 00:06:25.620
Think before answering? How does that work? Well,

00:06:25.720 --> 00:06:28.839
the idea is to improve accuracy, move beyond

00:06:28.839 --> 00:06:32.139
just matching keywords. It suggests a move towards

00:06:32.139 --> 00:06:35.279
more reliable AI research, hopefully less prone

00:06:35.279 --> 00:06:38.019
to making things up, you know, hallucinations.

00:06:38.180 --> 00:06:40.879
Okay, that's a big deal. Yeah. And for content

00:06:40.879 --> 00:06:42.819
creators, there's this tool called That's So

00:06:42.819 --> 00:06:45.519
Viral AI. It takes your images and turns them

00:06:45.519 --> 00:06:48.339
into engaging videos, like for Instagram or TikTok.

00:06:48.620 --> 00:06:50.959
Oh, interesting. Makes content creation way easier.

00:06:51.139 --> 00:06:53.000
Imagine just uploading a photo and getting a

00:06:53.000 --> 00:06:55.879
dynamic video for social media basically in seconds.

00:06:56.160 --> 00:06:58.860
That sounds incredibly useful. Okay, and for

00:06:58.860 --> 00:07:00.519
some quick hits, some rapid -fire updates from

00:07:00.519 --> 00:07:03.180
the AI world. Gemini reportedly had a memory

00:07:03.180 --> 00:07:06.740
feature in testing, like chat GPTs. Ah, remembering

00:07:06.740 --> 00:07:09.500
past chats. Exactly, which would be a huge step

00:07:09.500 --> 00:07:11.740
for making conversations feel more continuous.

00:07:11.980 --> 00:07:14.329
Definitely need it. Then there's the godfather

00:07:14.329 --> 00:07:17.170
of AI, Jeffrey Hinton, who recently suggested

00:07:17.170 --> 00:07:19.910
what he thinks is the only way humanity can survive,

00:07:20.189 --> 00:07:23.970
super intelligent AI. Ooh, heavy stuff. Pretty

00:07:23.970 --> 00:07:26.850
heavy thought from him, yeah. Also, DeepSeek

00:07:26.850 --> 00:07:29.790
R2 is rumored to launch soon, maybe between August

00:07:29.790 --> 00:07:33.290
15th and 30th. So that's one to watch for new

00:07:33.290 --> 00:07:35.629
model capabilities. Another contender. And this

00:07:35.629 --> 00:07:39.129
next one really caught my eye. Leaked meta AI

00:07:39.129 --> 00:07:42.290
rules. Apparently they suggested the AI was allowed

00:07:42.290 --> 00:07:44.930
to have romantic chats with kids? Wait, what?

00:07:45.089 --> 00:07:48.269
Romantic chats with kids? That was the leak reported.

00:07:48.449 --> 00:07:50.610
This particular detail certainly raises some

00:07:50.610 --> 00:07:53.189
serious questions in eyebrows. Yeah, wow. That

00:07:53.189 --> 00:07:56.529
meta rule. That definitely stands out. And one

00:07:56.529 --> 00:07:58.670
more quick one. Anthropic, another major AI player,

00:07:58.870 --> 00:08:01.370
just did an acqui -hire. Buying the talent, basically.

00:08:01.389 --> 00:08:03.990
Exactly. They snapped up three co -founders and

00:08:03.990 --> 00:08:06.589
key members of Human Loop, a significant talent

00:08:06.589 --> 00:08:09.189
grab in this super competitive space. Okay. From

00:08:09.189 --> 00:08:10.689
that whole quick list there, what's the most

00:08:10.689 --> 00:08:13.189
unexpected or maybe concerning item for you?

00:08:13.970 --> 00:08:17.149
I think it has to be those rumored meta AI rules

00:08:17.149 --> 00:08:20.069
about romantic chats with kids. It just really

00:08:20.069 --> 00:08:23.529
highlights the massive complexities and, frankly,

00:08:23.629 --> 00:08:26.889
dangers in AI development and deployment. Okay.

00:08:26.910 --> 00:08:28.629
We're going to take a quick break here. Sponsor

00:08:28.629 --> 00:08:30.759
read. All right. We're back. If we connect all

00:08:30.759 --> 00:08:33.080
this back to the bigger picture, AI isn't just

00:08:33.080 --> 00:08:35.820
for tech companies or creative apps anymore.

00:08:36.299 --> 00:08:38.700
It's now actually designing life -saving medicine.

00:08:39.059 --> 00:08:41.340
This feels like a really profound shift in what

00:08:41.340 --> 00:08:44.460
we thought AI was capable of. It really is. Researchers

00:08:44.460 --> 00:08:48.220
at MIT, they use generative AI to create completely

00:08:48.220 --> 00:08:51.200
new antibiotic compounds. Not just tweaked old

00:08:51.200 --> 00:08:53.799
ones. No, I mean brand new. Yeah. These molecules...

00:08:54.009 --> 00:08:55.690
They don't exist in any known chemical library

00:08:55.690 --> 00:08:58.389
anywhere. This isn't just finding new uses for

00:08:58.389 --> 00:09:02.230
old drugs. This is AI literally imagining totally

00:09:02.230 --> 00:09:05.129
novel molecules from scratch. We're talking chemical

00:09:05.129 --> 00:09:07.350
structures that humans, well, we wouldn't have

00:09:07.350 --> 00:09:09.570
even thought to test. It's like AI hallucinations,

00:09:09.570 --> 00:09:11.889
but useful. Exactly. Like hallucinations, but

00:09:11.889 --> 00:09:14.110
in this case, incredibly useful ones that actually

00:09:14.110 --> 00:09:16.529
seem to work. And here's the really impactful

00:09:16.529 --> 00:09:20.049
part. They do work. These AI -designed compounds

00:09:20.049 --> 00:09:22.230
were effective against deadly drug -resistant

00:09:22.230 --> 00:09:26.649
infections, things like the notorious MRSA and

00:09:26.649 --> 00:09:29.940
even multidrug -resistant gonorrhea. That's AI

00:09:29.940 --> 00:09:32.220
-powered drug discovery happening right now at

00:09:32.220 --> 00:09:35.440
MIT, making a real -world potential impact. That's

00:09:35.440 --> 00:09:39.200
amazing. Get this. The AI designed over 36 million

00:09:39.200 --> 00:09:42.820
potential compounds. 36 million. Yeah. And out

00:09:42.820 --> 00:09:45.480
of all those, only two made it through all the

00:09:45.480 --> 00:09:47.580
steps to actually be synthesized and tested,

00:09:47.720 --> 00:09:50.080
but both of them worked. Both worked. That's

00:09:50.080 --> 00:09:52.879
insane efficiency, considering that huge search

00:09:52.879 --> 00:09:55.299
space. Yeah. Yeah, it really is. And to give

00:09:55.299 --> 00:09:58.179
you some specifics. For the drug -resistant gonorrhea,

00:09:58.320 --> 00:10:02.500
the AI generated 7 million molecules. Okay. One

00:10:02.500 --> 00:10:04.460
of these compounds kills the bug by targeting

00:10:04.460 --> 00:10:06.720
a protein that no existing drug has ever targeted

00:10:06.720 --> 00:10:09.840
before. It's called LPTA. So the bacteria have

00:10:09.840 --> 00:10:11.960
no defense yet. Precisely. It's crucial because

00:10:11.960 --> 00:10:13.500
it means the bacteria haven't developed resistance

00:10:13.500 --> 00:10:16.080
to it. Unlike so many older antibiotics, it's

00:10:16.080 --> 00:10:18.419
a completely new angle of attack. And for MRSA?

00:10:18.620 --> 00:10:22.360
For MRSA, 29 million molecules were generated.

00:10:22.909 --> 00:10:24.970
and the compounds they identified successfully

00:10:24.970 --> 00:10:29.370
cleared MRSA skin infections in mice, but still

00:10:29.370 --> 00:10:32.330
a huge proof of concept. Incredible. And critically,

00:10:32.549 --> 00:10:35.409
these compounds... They don't resemble any existing

00:10:35.409 --> 00:10:37.950
antibiotic. So, again, resistance hasn't caught

00:10:37.950 --> 00:10:40.590
up. And yet, just to be clear, they synthesized

00:10:40.590 --> 00:10:42.649
them, they tested them in mice, and both drugs

00:10:42.649 --> 00:10:45.450
worked just as they hoped. Whoa. Just imagining

00:10:45.450 --> 00:10:48.149
AI designing entirely new drugs from scratch

00:10:48.149 --> 00:10:50.230
like that. It might be the closest thing we've

00:10:50.230 --> 00:10:52.649
seen to a real AI -native invention, you know?

00:10:52.690 --> 00:10:54.730
Yeah. Not just pattern matching, but actual creation.

00:10:54.970 --> 00:10:57.070
So basically, anywhere we need to discover new

00:10:57.070 --> 00:11:00.399
compounds, fast, new drugs, new materials. maybe

00:11:00.399 --> 00:11:02.759
AI could become the front line. Seems like it.

00:11:02.820 --> 00:11:04.279
It's not just about speed, right? It's about

00:11:04.279 --> 00:11:06.860
exploring a chemical space that humans using

00:11:06.860 --> 00:11:08.940
traditional methods simply can't fathom. It's

00:11:08.940 --> 00:11:11.639
too vast. Exactly. So what's the biggest implication,

00:11:11.820 --> 00:11:14.539
do you think, of AI creating brand new drugs

00:11:14.539 --> 00:11:18.360
like this? I mean, it radically accelerates discovery.

00:11:18.700 --> 00:11:21.740
It could potentially solve previously untreatable

00:11:21.740 --> 00:11:24.480
diseases. That's the hope. So let's try and pull

00:11:24.480 --> 00:11:26.120
this together. What does this all mean for you?

00:11:26.409 --> 00:11:29.289
We've seen AI getting smaller, running offline,

00:11:29.429 --> 00:11:31.710
becoming much more accessible right there on

00:11:31.710 --> 00:11:34.309
your personal devices. Putting power in your

00:11:34.309 --> 00:11:36.450
hands. Right, putting powerful tools directly

00:11:36.450 --> 00:11:38.129
in your hands, not just somewhere off in the

00:11:38.129 --> 00:11:40.990
cloud. We've also seen its broader impact from

00:11:40.990 --> 00:11:44.090
potentially shifting beliefs through subtle interactions

00:11:44.090 --> 00:11:48.009
to, well, incredibly, designing potentially life

00:11:48.009 --> 00:11:50.809
-saving drugs. At a huge range. The pace of innovation

00:11:50.809 --> 00:11:53.669
is just accelerating at this unprecedented rate.

00:11:53.750 --> 00:11:56.639
It's touching, it seems. aspect of our lives,

00:11:56.740 --> 00:12:00.039
from personalized stories to fundamental breakthroughs

00:12:00.039 --> 00:12:02.580
in medicine. Yeah, and this raises a really important

00:12:02.580 --> 00:12:06.269
question, I think. As AI becomes this pervasive,

00:12:06.409 --> 00:12:09.049
this powerful, and as its capabilities start

00:12:09.049 --> 00:12:10.690
looking more and more like genuine invention,

00:12:11.190 --> 00:12:13.429
how do we ensure we're not just passive consumers

00:12:13.429 --> 00:12:16.009
of its outputs? Right. How do we stay informed

00:12:16.009 --> 00:12:18.850
critical thinkers about its true capabilities?

00:12:19.149 --> 00:12:21.970
Yeah. And also its ethical limits. Think about

00:12:21.970 --> 00:12:24.429
that for a moment. How will you interact with

00:12:24.429 --> 00:12:27.509
these increasingly capable and sometimes, yeah,

00:12:27.529 --> 00:12:30.649
quirky AI systems as they become more integrated

00:12:30.649 --> 00:12:32.970
into your daily life? It's something we all need

00:12:32.970 --> 00:12:34.779
to. consider we really encourage you to keep

00:12:34.779 --> 00:12:36.940
exploring keep asking questions thank you for

00:12:36.940 --> 00:12:39.379
joining us on this deep dive today until next

00:12:39.379 --> 00:12:41.679
time keep learning ot row music