0:04 Hello, I'm Tony Co, the host of Talkit
0:07 Global. In today's session, we're going
0:10 to explore how we can accelerate our own
0:12 growth using AI.
0:14 When you hear the story of Gabriel
0:16 Peterson, who dropped out of high school
0:20 in Sweden, self-studied AI, math, and
0:23 coding solely through generative AI
0:25 chat, and eventually became an AI
0:28 researcher at OpenAI in his mid20s. You
0:30 can sense that there is such a thing as
0:33 AI native learning and it's not just
0:35 about being good at prompting. We've
0:36 always been taught to build everything
0:39 step by step from the basics. But
0:42 Gabriel chose the exact opposite path.
0:44 Start with the problem, dig downward,
0:46 and pull in the knowledge you need. A
0:48 top- down learning approach.
0:51 Fixing bugs, asking questions, diving
0:53 deeper, the fundamental naturally
0:56 follow. In the AI era, perhaps the
0:58 people who touch the problem first are
1:06 The way I think people learn the fastest
1:09 is by what you would call like a like uh
1:12 um top down approach, right? You'll
1:14 probably learn faster if you start with
1:16 a problem and then you can read about
1:18 everything required to to to start
1:19 solving the problem and then you find
1:21 more problems and you read about those
1:23 and then you go down to like the the
1:24 core of the problem, right? is you start
1:26 with actual task and you go down. But
1:28 that's extremely rare way to learn like
1:31 in school everyone has this mindset
1:32 right of like okay we need to start with
1:35 the foundations we need to start like if
1:36 you want to work with machine learning
1:37 like you can forget about doing any
1:39 machine learning for the first like four
1:41 years right it's like math and then and
1:41 then you have like matrix
1:43 classifications you have linear
1:44 algorithm you have all these things that
1:46 build up and then you have the simpler
1:48 ML that's like super outdated you have
1:50 like you know linear regression all
1:52 these things that are still used partly
1:54 but it's like it will take you very long
1:55 time until you get to like production
1:56 grade ML
1:59 Why is this? Well, it's extremely hard
2:02 to scale the top down approach because
2:04 that requires like a teacher always
2:06 being there for you. It requires you
2:08 being able to know exactly what piece of
2:10 thing you need to learn at any point of
2:12 time. Well, if you do bottom up, you
2:14 know, okay, first you always learn this
2:15 and then you always learn this
2:17 >> and it's it's much easier to scale. It's
2:20 extremely inefficient. And now with
2:22 Shachvt, all this changes like this will
2:24 change. People say education will change
2:26 all the time, but I can barely take
2:28 universities seriously that don't teach
2:30 HGBT as a part of their curriculum. It's
2:32 like actually insane that this is not
2:34 like a a course that's taught from like
2:35 2 years old. Like suddenly foundational
2:38 knowledge universities don't have like
2:40 um a monopoly on on foundational
2:42 knowledge anymore. You can just get any
2:43 foundational knowledge from from from
2:45 shipb and people haven't really
2:47 internalized how top down problem
2:49 solving works. They will always tell you
2:51 things you know like oh but you'll never
2:53 actually understand the problem. you'll
2:54 never actually blah blah blah. And this
2:56 is not true. You start with a problem,
2:58 you recursively go down. Like if I want
3:00 to learn machine learning, I ask
3:02 Chashiv, okay, what project should I do?
3:04 Write a project for me. I have bugs. I
3:06 start fixing the bugs and then things
3:08 work. And from there, I start with a
3:09 specific part of the machine learning
3:11 problem. Like, okay, uh what happens
3:12 here? Can you explain to me with
3:14 intuition why this module here makes the
3:16 model learn? And it will explain to you.
3:17 And then oh, it uses matrix
3:19 multiplication and linear algebra, you
3:21 know. Okay. How do they work? What's the
3:23 math intuition behind this? like show me
3:25 like make up a couple graphs to really
3:27 make me get an intuition for this part
3:29 of ML and then suddenly you have all the
3:30 foundational knowledge like it doesn't
3:33 need to go bottom up anymore and this
3:35 shift will will yeah I think this shift
3:36 will like fundamentally change how
3:39 education is done and you just continue
3:41 to ask the model constantly until you
3:42 really understand and when you
3:44 understand you can just tell the model
3:45 okay this is my understanding of this is
3:47 this completely correct and then you
3:48 you'll also start learning about all
3:49 these like small tricks you can do right
3:51 like explain this concept like I'm 12
3:53 years That one is really good. It will
3:55 start like super easily like imagine
3:57 you're in a bookstore and you can
3:58 imagine the embeddings being the
4:01 different books in the store and then
4:03 you can imagine you know all this and it
4:04 will connect everything that has to do
4:06 with AI to like real world concepts
4:08 which makes it really easy to to reason
4:11 about for for someone like like me.
4:13 Gabriel emphasizes finding the gaps in
4:16 your knowledge and digging in until you
4:18 reach that aha moment. It means asking
4:21 AI to explain things more precisely and
4:23 more directly and reconstructing your
4:26 learning by exposing every intermediate
4:28 step in the code. How often you can
4:30 create these click moments
4:32 that becomes a new metric of learners in
4:39 The skill that's required here is
4:42 knowing what gaps you have in your
4:44 knowledge. like see you have an AI model
4:45 or like whatever else you want to learn
4:48 and understanding when you don't really
4:50 understand the part it's actually pretty
4:51 hard to do like it's something you need
4:52 to train up and practice on yourself
4:54 like wait a second do I really
4:57 understand this part and then so that's
4:58 one signal you need the second signal is
5:00 when you start asking questions you need
5:01 to have a really strong signal for when
5:03 it clicks when you're like ah there it clicked
5:04 clicked
5:06 >> okay just understand like fundamentally
5:08 why this thing is as it is how would
5:10 someone else learn how to learn with AI
5:11 >> this is a very good question I mean
5:12 first of all just change like the
5:14 misconception of AI being used to do the
5:18 work for you to instead you know use the
5:23 AI to explicitly help you learn like you
5:25 you don't you don't just use it to get
5:27 work done you actually learn from it I
5:28 mean the moment you just switch that
5:30 mindset which seems still fairly
5:31 uncommon but it's becoming more and more
5:33 common all the time you have most of the
5:34 things to to get there right and and
5:36 then to become really good first of all
5:37 like I said like know when you have gaps
5:40 in your knowledge understand what it
5:42 feels like when you fundamentally grasp
5:44 something and you know you you you
5:45 constantly come up with all these hacks
5:49 like uh you you'll notice
5:52 will respond in a fairly standard way
5:54 and your way of learning is probably not
5:56 exactly what it responds like because it
5:58 wants to you know make sure everyone has
5:59 a good experience
6:00 >> but you probably want it to respond in
6:02 another way. I very often tell it for
6:04 example be extremely direct and
6:06 concrete. Always show me all the
6:08 intermediate states and the shapes of
6:11 the code you produce. make sure to to
6:13 make sure I have like a really intuitive
6:16 understanding of why it happens. And if
6:18 you're unsure, make sure you show me
6:20 options and like what others have tried
6:21 and why this works and why something
6:23 else didn't work. And you start becoming
6:26 good at like asking these questions that
6:28 give you the aha moment. Like as fast as
6:30 possible, you want to get to the aha
6:31 moment. Yeah. Like the first time you
6:33 understood linear algebra or the first
6:34 time you understood what back
6:36 propagation works, you probably had a
6:38 very clear like, oh wow, it finally
6:40 clicked. and to chase these clicks and
6:42 to make them appear like as frequent as
6:43 possible, right? That's like kind of
6:46 your utility function.
6:48 The era of listing specs in an interview
6:50 is over. Gabriel stresses just one
6:52 thing, the proof that you built
6:54 something yourself. A demo that anyone
6:56 can understand at a glance, one that
6:58 makes people think within 3 seconds, ah,
7:00 this person can actually build things.
7:01 That simple proof carried more weight
7:04 than educational words, and it explained
7:06 Gabriel, a high school dropout, far
7:08 better than any resume ever could. The
7:10 number one thing I recommend to people
7:13 is making a demo that is super super
7:15 simple. It's actually really hard to
7:16 make a good demo for a lot of reasons.
7:18 Everyone thinks it's hard because they
7:20 need to make a demo that is hard and
7:22 they don't have the skills. This is very
7:25 not true. You can make very simple like
7:27 you don't need that much code knowledge
7:29 to make a really cool cool demo. The
7:31 hard part of making a demo is making
7:33 sure that people understand why you can
7:35 code within 3 seconds. You know you have
7:37 like 100 like applicants for something.
7:38 If you apply with one link and they
7:40 press the link and you know you have one
7:41 shot, right? Like making sure you build
7:43 a really cool demo where people
7:44 understand what they're looking at,
7:46 which is really hard, and where people
7:47 understand that you're a really good
7:49 engineer, which is really hard, but then
7:51 you're there. I mean, that's all they
7:52 want to see. I mean, companies just want
7:54 to make money. You show them how to make
7:55 money, that you can code, and they'll
7:56 hire you. And then you might say, "Oh,
7:58 but they only hire people with degrees."
8:00 Well, yeah, because literally no one has
8:02 ever showed them that they can do their
8:03 work. They're like, "Oh, I had these
8:05 internships." And the interviewer would
8:07 be like, "Okay, what did you do there?"
8:09 Oh, I streamlined pipelines and made
8:12 things 30% more efficient. And usually
8:15 like, uh, okay, well, that tells me
8:17 literally nothing. Okay, what what else?
8:18 What have you done? Oh, I went to
8:20 Harvard. I have the best grades. Well, I
8:22 still don't know if you can do the job,
8:24 right? Oh, but I have all this
8:26 extracurricular. I was debate champion.
8:28 You start going on about all these
8:30 things that your parents will tell you,
8:32 people around you will tell you. Nothing matters.
8:34 matters.
8:36 Lastly, your likes and subscriptions
8:37 greatly empower us to continue creating
