How I Use AI to Learn Things — Transcript
Full transcript
- 0:00How could we use AI to learn better? I
- 0:02mean, there's got to be some way to
- 0:04optimize learning with AI. I think it's
- 0:07the perfect tool for it. Uh it is not
- 0:10entirely clear yet how. So, yeah. Uh
- 0:13this video will be me sharing my own
- 0:15current approach to this. First, uh I'll
- 0:18go through the reasoning behind the
- 0:21approach and the design, and then we'll
- 0:24do a live demo. All right. So, this is
- 0:26the standard way of learning that we're
- 0:28used to. Each outlet uh so, a teacher, a
- 0:32book, a course, whatever, uh
- 0:35they all are designed to teach many
- 0:38students. And also, each students learns
- 0:42from many outlets. So, there's a
- 0:45many-to-many relation between the
- 0:47learners and the outlets, and both
- 0:50directions of it produce their own kind
- 0:53of inefficiency. First, one outlet
- 0:56teaches many. An outlet that's designed
- 0:59for many people cannot be optimal for
- 1:02any one of them. That's because optimal,
- 1:05quote-unquote optimal, teaching
- 1:07uh that would clearly depend entirely on
- 1:09the learner's current understanding. And
- 1:12this includes both the whole teaching
- 1:15arc from their current understanding to
- 1:18their goal understanding, as well as
- 1:20each explanation along the way. An
- 1:22optimal teaching path, then, uh would be
- 1:25one that minimizes teaching stuff that
- 1:28the learner already holds, and also
- 1:30stuff that they can't yet understand.
- 1:33So, it would work exactly at the edge of
- 1:36their understanding. So, ideally, a
- 1:39teacher would have exactly one student.
- 1:42All right. So, second direction, uh one
- 1:44student learns from many outlets. This
- 1:47means many teaching styles to get used
- 1:51to, many notations, many levels of
- 1:54reliability, many interfaces, and the
- 1:57switching between them uh costs mental
- 2:01effort that's not gone into learning the
- 2:04material. But, I think there's a much
- 2:06deeper cost in this, which is trust.
- 2:10With an unfamiliar outlet, the brain I I
- 2:14think the brain kind of hedges. So, it
- 2:16won't fully commit to accepting a fact
- 2:19until the outlet has proven itself and
- 2:22is familiar enough. I think a good way
- 2:24to illustrate this with is with the
- 2:27following example. Let's say we want to
- 2:29understand the hairy ball theorem, and
- 2:33we have two identical explanations. But,
- 2:35the other one is in a 3Blue1Brown video,
- 2:38and the other is on X posted by some
- 2:42random guy. Now, even though the
- 2:44explanations are identical, I think it
- 2:46is clear that we're going to learn a
- 2:48better from the familiar source. The
- 2:51brain is going to have a much easier
- 2:53time internalizing the information when
- 2:56it trusts the outlet, even though the
- 3:00explanation is identical.
- 3:03All right? So, ideally, a student would
- 3:06have exactly one teacher. Now, we've
- 3:08established that the ideal scenario is
- 3:11one-to-one in both directions, but
- 3:15there's an obvious objection to this,
- 3:18which is that having only one teacher
- 3:22means getting only one perspective. But,
- 3:25I think that the objection conflates a
- 3:27source with an interface. So, a teacher
- 3:31doesn't reduce the number of sources,
- 3:35perspectives.
- 3:36Uh instead, it aggregates all of them
- 3:39and delivers them through one interface.
- 3:42So,
- 3:43we don't lose many perspectives. Now,
- 3:47with AI being the teacher, trust isn't
- 3:49really built over time. Rather, it's
- 3:52engineered into the system. The reason
- 3:55we have to have reliable verification
- 3:58and fact-checking is one, correct
- 4:01information. Obviously, we absolutely
- 4:04don't want the AI to hallucinate false
- 4:08information. But also, two, cuz it makes
- 4:11learning easier when we know that the
- 4:14system is reliable. So, one interface
- 4:18fitted to one mind over all sources.
- 4:22Now, the two inefficiencies that we went
- 4:24over, they give us the two principles
- 4:27that the system runs on. First,
- 4:30optimized teaching, the answer to the
- 4:33first inefficiency. And two, optimized
- 4:37allocation of mental resources, answer
- 4:40to the second one. And this does not
- 4:43mean removing difficulty. Instead, it's
- 4:46about concentrating all cognitive work
- 4:49into the material itself. We want to
- 4:52maximize struggle. We want to learn
- 4:56difficult things, so struggling is very
- 4:58important, but it has to be in the
- 5:01material itself and not in logistics,
- 5:05planning, finding resources, verifying
- 5:07facts, figuring out what to learn and in
- 5:09what order. All of that is for the
- 5:11system to absorb. Next, the actual
- 5:14process,
- 5:16how this is implemented. First,
- 5:18optimized teaching, to teach optimally,
- 5:22whatever this might be very different
- 5:24from person to person, but in order to
- 5:28do that, the system obviously needs to
- 5:31know the person's exact current
- 5:34understanding. So, it has to measure,
- 5:36and it does this with a quiz tool, which
- 5:40lets it ask graded multiple-choice
- 5:42questions. It starts off with very broad
- 5:45of
- 5:46and then basically
- 5:48binary search is the edge on every
- 5:52possible strand that the lesson will
- 5:54depend on. So, it's going to get a very
- 5:57detailed map of the learner's
- 6:00understanding. So, that's phase one,
- 6:03probe. Second, phase two is plan.
- 6:07Basically just reasons everything out.
- 6:09How do I teach this mind this specific
- 6:12thing? And here's also where it fires
- 6:15off its first verification and
- 6:18fact-checking sub-agents. And then
- 6:21presents the plan as a mermaid graph and
- 6:26diagram. And okay, basically two reasons
- 6:29for why why it has to show a graph. One,
- 6:33it gives the learner a better sense of
- 6:36what's to come. And also two, the actual
- 6:39reason I implemented it is to actually
- 6:42force the AI to reason everything out.
- 6:46That's the real reason.
- 6:49So,
- 6:50it cannot just cheat and wing it. And
- 6:53then finally, the teaching itself. And
- 6:56this phase is is going to be very
- 6:58different cuz it's going to depend on it
- 7:00entirely how you want to learn. You want
- 7:02to install the system with your own
- 7:05learning philosophy and how you learn
- 7:07best. But one detail that I think is
- 7:11going to be important regardless of the
- 7:12exact way of learning is feedback. So,
- 7:17the AI is going to quiz you periodically
- 7:20to check whether you actually understood
- 7:22the thing. And this is important for
- 7:25three reasons. One, it's very easy to
- 7:28sort of gaslight yourself into thinking
- 7:31that you understood something,
- 7:33especially when learning with AI. So,
- 7:36actually testing your understanding is
- 7:39important feedback for yourself. Two,
- 7:41the system needs continuous feedback to
- 7:44stay calibrated. And three,
- 7:48applying the material, doing practice
- 7:50problems, using the stuff that you
- 7:53learned, also just helps you learn
- 7:55better. It's It's a part of how the new
- 7:57information and the new understanding
- 8:00locks in. So, yeah.
- 8:02Uh that's the That's the general idea.
- 8:04Now, let's Let's see it in action. All
- 8:08right, we're in the Pi Agent harness now
- 8:11in my learning directory. And yeah, this
- 8:14is where I've got everything set up. So,
- 8:16in the dot Pi folder, I've got the teach
- 8:18skill, some visualization stuff,
- 8:21uh the quiz extension, the MD log
- 8:24extension, which I'll show you in a bit,
- 8:27and two subagents to make visuals.
- 8:30Hopefully, we'll see those also. So,
- 8:33yeah. I think the only way to demo a
- 8:36system like this is actually to try to
- 8:38learn something. And to really go full
- 8:41circle, uh in my previous
- 8:43learning-related video, I mentioned the
- 8:46Maxwell's equations and how they can be
- 8:49expressed in just two equations using
- 8:53differential forms. And as you can see,
- 8:56uh I don't really know anything about
- 8:59differential forms. So, I think this is
- 9:01what we're going to learn today. Now,
- 9:03obviously, I don't We're not going to
- 9:05get to this level, but uh I think this
- 9:07is a very good example to showcase the
- 9:09system with a an actual learning
- 9:12process. So, yeah. Uh let's begin. Let
- 9:15me pull up Obsidian here cuz this is
- 9:17what I use.
- 9:19It's It's sort of like the UI for
- 9:22everything. Learning differential forms.
- 9:25This is what the MD log extension is
- 9:28for.
- 9:28It lets me link a Markdown file to the
- 9:31session, and then everything is going to
- 9:33get printed right here. Just a nicer way
- 9:37to see things and also get the LaTeX
- 9:39rendering. So, yeah, that's my solution.
- 9:42And yeah, it's also nice to have like
- 9:44persistence artifacts from each learning
- 9:47session. So, yeah,
- 9:49let's begin teach. I want to get like a
- 9:52solid introduction to differential
- 9:54forms.
- 9:55And we're using Kimik 3 on Macs cuz I
- 10:00really found that the intelligence of
- 10:02the model really matters
- 10:04in teaching. The teaching instructions
- 10:07are quite
- 10:09They're quite specific.
- 10:10Anyway, [snorts]
- 10:12let's see how this goes. We're going to
- 10:13get questions now.
- 10:15A force field F acts on a particle as it
- 10:18moves along a curve C. What does the
- 10:21line integral
- 10:23compute?
- 10:24Let me also I've got this note here
- 10:26where I can just yap into. Basically,
- 10:29just to give it more context and sort of
- 10:32talk through my reasoning.
- 10:34It's not going to be number one.
- 10:36I mean, number one does hold for
- 10:38conservative fields, but for any force
- 10:41field F, it's going to be number three.
- 10:44The network done by the field on the
- 10:47particle.
- 10:50All right, next one. The divergence of a
- 10:53vector field at a point measures the net
- 10:56outward flux per unit volume at that
- 10:59point.
- 11:02All right, now the Stokes theorem. Yes.
- 11:07And now we get Faraday's law of
- 11:09induction. I mean, you can always just
- 11:10give it more context at the beginning
- 11:13about what stuff you already understand
- 11:17very thoroughly.
- 11:18In this case, I gave you very little
- 11:19context, so we're going to get a long
- 11:22probing phase.
- 11:26Oh we're going to have to go into
- 11:27relativity.
- 11:29In special relativity, when you change
- 11:31to a moving reference frame, what
- 11:33happens to the electric and magnetic
- 11:35fields? Okay, let's let's read through
- 11:38these. Both are invariants. All
- 11:40observers measure the same E and B at
- 11:43each event. They mix. A purely electric
- 11:46field in one frame has both electric and
- 11:48magnetic parts in another. Hold on. Only
- 11:51E changes. B stays fixed.
- 11:54No, obviously not. Three, each
- 11:57transforms independently.
- 12:00And I'll I'll go ahead and answer I
- 12:01don't know. And it was two.
- 12:04They mix.
- 12:06All right, it's stopped with questions.
- 12:09Um so yeah, uh finally done with the
- 12:12probing phase.
- 12:14Uh it did ask a lot of questions, but
- 12:17now it's got a very detailed image of my
- 12:19understanding. And it's also just a nice
- 12:22warm-up as well. And I don't really got
- 12:24to worry about anything. Just answer the
- 12:26questions and it handles all the
- 12:29logistics. Now we've got a researcher
- 12:31still going on fact-checking stuff.
- 12:34Although usually math isn't really
- 12:36something it would need to fact-check.
- 12:39But well, it's it's good practice
- 12:41anyway. Then after this, it's going to
- 12:43fully plan everything out. All right,
- 12:45finally, we're getting the plan.
- 12:50I did start rendering
- 12:52the mermaid in the 2A, but
- 12:55don't know what happened. It was
- 12:56probably
- 12:57in the most recent release, I believe,
- 12:59added
- 13:01mermaids. But anyway, we get it in the
- 13:03obsidian. This is the plan. Let's go. A
- 13:07lot of things I'm going to be reading
- 13:09reading the plan. All right, so that's
- 13:11cool. We're going to get to generalized
- 13:13Stokes well before the actual goal.
- 13:16Okay, perfect. It's it's using the
- 13:18visualization skill. So we're going to
- 13:20get some visuals soon. All right, it's
- 13:23making an SVG. And the reason that these
- 13:26are done in sub agents is to obviously
- 13:29preserve context but also
- 13:32the sub agents will look at the image to
- 13:35verify whether it actually looks
- 13:38correct. It wrote an SVG and now
- 13:41it's going to view it. It's going to
- 13:43edit it a bit, look at it again.
- 13:47Yes. So, we're going to get that in the
- 13:49next message. But now the first
- 13:52node. So, it's going to It's going to
- 13:55slowly
- 13:56walking down the path one reasoning step
- 14:00at a time. Because what usually happens
- 14:03if you're like talking to chat GPT or
- 14:05something trying to get it to explain or
- 14:07teach you anything it usually just it's
- 14:10way too excited and it rushes through
- 14:13the whole thing. And so, I like to keep
- 14:14this very slow, one reasoning step at a
- 14:18time. So, now we're introducing
- 14:20co-vectors. Let me
- 14:22read through this.
- 14:24All right. So, now it's introduced
- 14:26co-vectors
- 14:27and also given a new perspective on the
- 14:31X.
- 14:32This is really cool. Yeah, it's going to
- 14:34quiz me on the this reasoning step to
- 14:38confirm that everything is actually
- 14:39understood, to confirm my understanding.
- 14:41Now, yeah, it's going to incorporate the
- 14:44visual
- 14:45which will be all rendered into Obsidian
- 14:48and automatically here as an embed
- 14:52embedded file. Let's alpha be 3D X minus
- 14:572D Y and V this. What is alpha of V?
- 15:01Which is going to be -4.
- 15:04Okay, this makes perfect sense.
- 15:06The picture helped. Oh, yeah, the
- 15:09Pi also got like built-in LaTeX
- 15:12rendering. Okay, which is kind of cool.
- 15:13Yeah, it does the job but
- 15:16it's definitely not the same as this.
- 15:20But yeah,
- 15:21cool features though. All right, so now
- 15:23we're extending
- 15:25the covectors to a covector field. All
- 15:28right, so now we get a new perspective
- 15:30on what the line integral actually is.
- 15:33Yeah.
- 15:34One thing I don't like about this is
- 15:37it's I haven't really configured its
- 15:41style of speaking. So we get a lot of it
- 15:44sounds very AI, which I don't really
- 15:47mind cuz I don't want to cram it with
- 15:49too many instructions to worry about. So
- 15:53I'm fine with these LLM-isms.
- 15:56It's not X, it's Y.
- 15:59Wow, insane. The quality of the teaching
- 16:02is what matters. But yeah, it's
- 16:04basically walking down the dependency
- 16:07tree. Seems like we're still here.
- 16:10So when we're going to get I think this
- 16:11is the wedge product. But yeah, I like
- 16:14that it moves one reasoning step at a
- 16:17time. So if at any point I have
- 16:18questions, I can always ask.
- 16:21And it doesn't rush forwards. The each
- 16:25step is very easy to digest. And it's
- 16:28going to give me everything that I can
- 16:30accept at face value like this. All
- 16:32integration over a one-dimensional thing
- 16:35is the integration of a one-form. Now, I
- 16:38assume that this continues, yeah.
- 16:41A K-form will be the kind of thing a
- 16:44K-dimensional surface can eat.
- 16:47Sure.
- 16:49All right, we're getting a new new SVG.
- 16:54Seems like our
- 16:56uh SVG maker
- 16:58died. Agent error. Overloaded. Bro, not
- 17:03now. Trying to learn differential forms.
- 17:05Oh no, we're getting it. Now we're
- 17:06getting the wedge products. That's cool.
- 17:10Machines that are bilinear and
- 17:11antisymmetric. Yeah, where do we get
- 17:13them? I'm sure this is not
- 17:15antisymmetric. The minimal fix is the
- 17:18oldest trick in the book. Aha. Yeah,
- 17:21that gives something anti-symmetric.
- 17:24So, is that just what it is? Sure, okay,
- 17:27fine. But yeah, anyway, this is we're
- 17:29obviously not going to go
- 17:31through the whole dependency graph here.
- 17:35Um this is going to be this one, I
- 17:37believe. Yeah.
- 17:40But yeah, anyway, where did we get to in
- 17:42the DAG? Here.
- 17:44Yeah, I think we got here. Yeah, we're
- 17:46approaching generalized Stokes now.
- 17:49Yeah, that's cool. Um but yeah, I think
- 17:53this is enough for a demo. So, this has
- 17:55been sort of a second part or
- 17:57continuation of where we left off last
- 17:59time in my previous learning related
- 18:01video. So, what this video has been
- 18:04really about is how to take a learning
- 18:07philosophy or way of teaching and
- 18:10implement that as an AI system. I think
- 18:13the main reason for why this has worked
- 18:15so well for me is none of the stuff I
- 18:19talked here about, rather just the style
- 18:21of teaching I covered in the previous
- 18:23video, which affects exactly two things:
- 18:26the learning arc, the path that we're
- 18:30going to take from our current
- 18:31understanding to our goal understanding,
- 18:33and two, the individual steps and
- 18:36explanations along the way.
- 18:39Yes. Anyway, just some ideas for you for
- 18:43you to think about. I'd like to hear
- 18:45your your thoughts on this. And how how
- 18:46how could the system be improved? How do
- 18:49you use AI in learning? I'd like to
- 18:51know. I'm very invested into refining
- 18:55this further. So, yeah, uh that's it.
About this transcript
This page contains the full transcript of How I Use AI to Learn Things by Eero Alvar, generated from the public captions YouTube serves with the video. The transcript has 2,662 words across 425 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.