YT YouTube2Text

YouTube2Text → Transcripts → Andrej Karpathy: Software Is Changing (Again)

Andrej Karpathy: Software Is Changing (Again) — Transcript

by Y Combinator · 8,255 words · 1,139 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:01Please welcome former director of AI
0:04Tesla Andre Carpathy.
0:07[Music]
0:11Hello.
0:14[Music]
0:19Wow, a lot of people here. Hello.
0:22Um, okay. Yeah. So I'm excited to be
0:24here today to talk to you about software
0:27in the era of AI. And I'm told that many
0:30of you are students like bachelors,
0:32masters, PhD and so on. And you're about
0:34to enter the industry. And I think it's
0:36actually like an extremely unique and
0:37very interesting time to enter the
0:38industry right now. And I think
0:41fundamentally the reason for that is
0:43that um software is changing uh again.
0:47And I say again because I actually gave
0:49this talk already. Um but the problem is
0:52that software keeps changing. So I
0:54actually have a lot of material to
0:55create new talks and I think it's
0:56changing quite fundamentally. I think
0:58roughly speaking software has not
1:00changed much on such a fundamental level
1:02for 70 years. And then it's changed I
1:04think about twice quite rapidly in the
1:06last few years. And so there's just a
1:08huge amount of work to do a huge amount
1:09of software to write and rewrite. So
1:12let's take a look at maybe the realm of
1:14software. So if we kind of think of this
1:16as like the map of software this is a
1:17really cool tool called map of GitHub.
1:20Um this is kind of like all the software
1:21that's written. Uh these are
1:23instructions to the computer for
1:24carrying out tasks in the digital space.
1:26So if you zoom in here, these are all
1:28different kinds of repositories and this
1:30is all the code that has been written.
1:31And a few years ago I kind of observed
1:33that um software was kind of changing
1:35and there was kind of like a new type of
1:37software around and I called this
1:39software 2.0 at the time and the idea
1:42here was that software 1.0 is the code
1:44you write for the computer. Software 2.0
1:46know are basically neural networks and
1:48in particular the weights of a neural
1:50network and you're not writing this code
1:53directly you are most you are more kind
1:55of like tuning the data sets and then
1:56you're running an optimizer to create to
1:58create the parameters of this neural net
2:00and I think like at the time neural nets
2:02were kind of seen as like just a
2:03different kind of classifier like a
2:04decision tree or something like that and
2:06so I think it was kind of like um I
2:09think this framing was a lot more
2:10appropriate and now actually what we
2:12have is kind of like an equivalent of
2:13GitHub in the realm of software 2.0 And
2:15I think the hugging face is basically
2:18equivalent of GitHub in software 2.0.
2:20And there's also model atlas and you can
2:22visualize all the code written there. In
2:24case you're curious, by the way, the
2:25giant circle, the point in the middle,
2:28uh these are the parameters of flux, the
2:30image generator. And so anytime someone
2:32tunes a on top of a flux model, you
2:34basically create a git commit uh in this
2:37space and uh you create a different kind
2:39of a image generator. So basically what
2:41we have is software 1.0 is the computer
2:43code that programs a computer. Software
2:452.0 are the weights which program neural
2:48networks. Uh and here's an example of
2:50Alexet image recognizer neural network.
2:53Now so far all of the neural networks
2:55that we've been familiar with until
2:56recently where kind of like fixed
2:58function computers image to categories
3:01or something like that. And I think
3:03what's changed and I think is a quite
3:05fundamental change is that neural
3:06networks became programmable with large
3:09language models. And so I I see this as
3:12quite new, unique. It's a new kind of a
3:14computer and uh so in my mind it's uh
3:18worth giving it a new designation of
3:19software 3.0. And basically your prompts
3:22are now programs that program the LLM.
3:25And uh remarkably uh these uh prompts
3:28are written in English. So it's kind of
3:30a very interesting programming language.
3:33Um so maybe uh to summarize the
3:36difference if you're doing sentiment
3:37classification for example you can
3:39imagine writing some uh amount of Python
3:42to to basically do sentiment
3:44classification or you can train a neural
3:46net or you can prompt a large language
3:47model. Uh so here this is a few short
3:50prompt and you can imagine changing it
3:51and programming the computer in a
3:52slightly different way. So basically we
3:54have software 1.0 software 2.0 and I
3:57think we're seeing maybe you've seen a
3:59lot of GitHub code is not just like code
4:01anymore. there's a bunch of like English
4:03interspersed with code and so I think
4:05kind of there's a growing category of
4:07new kind of code. So not only is it a
4:09new programming paradigm, it's also
4:10remarkable to me that it's in our native
4:12language of English. And so when this
4:14blew my mind a few uh I guess years ago
4:17now I tweeted this and um I think it
4:20captured the attention of a lot of
4:21people and this is my currently pinned
4:23tweet uh is that remarkably we're now
4:25programming computers in English. Now,
4:28when I was at uh Tesla, um we were
4:31working on the uh autopilot and uh we
4:34were trying to get the car to drive and
4:37I sort of showed this slide at the time
4:39where you can imagine that the inputs to
4:41the car are on the bottom and they're
4:43going through a software stack to
4:44produce the steering and acceleration
4:47and I made the observation at the time
4:48that there was a ton of C++ code around
4:51in the autopilot which was the software
4:521.0 code and then there was some neural
4:54nets in there doing image recognition
4:56and uh I kind of observed that over time
4:58as we made the autopilot better
5:00basically the neural network grew in
5:02capability and size and in addition to
5:05that all the C++ code was being deleted
5:08and kind of like was um and a lot of the
5:12kind of capabilities and functionality
5:14that was originally written in 1.0 was
5:16migrated to 2.0. So as an example, a lot
5:19of the stitching up of information
5:20across images from the different cameras
5:22and across time was done by a neural
5:24network and we were able to delete a lot
5:26of code and so the software 2.0 stack
5:29quite literally ate through the software
5:32stack of the autopilot. So I thought
5:34this was really remarkable at the time
5:35and I think we're seeing the same thing
5:37again where uh basically we have a new
5:39kind of software and it's eating through
5:40the stack. We have three completely
5:42different programming paradigms and I
5:44think if you're entering the industry
5:45it's a very good idea to be fluent in
5:47all of them because they all have slight
5:49pros and cons and you may want to
5:50program some functionality in 1.0 or 2.0
5:53or 3.0. Are you going to train
5:54neurallet? Are you going to just prompt
5:55an LLM? Should this be a piece of code
5:57that's explicit etc. So we all have to
5:59make these decisions and actually
6:00potentially uh fluidly trans transition
6:03between these paradigms. So what I
6:06wanted to get into now is first I want
6:09to in the first part talk about LLMs and
6:11how to kind of like think of this new
6:13paradigm and the ecosystem and what that
6:15looks like. Uh like what are what is
6:17this new computer? What does it look
6:18like and what does the ecosystem look
6:20like? Um I was struck by this quote from
6:23Anduring actually uh many years ago now
6:25I think and I think Andrew is going to
6:27be speaking right after me. Uh but he
6:29said at the time AI is the new
6:30electricity and I do think that it um
6:33kind of captures something very
6:34interesting in that LLMs certainly feel
6:36like they have properties of utilities
6:38right now. So
6:41um LLM labs like OpenAI, Gemini,
6:44Enthropic etc. They spend capex to train
6:47the LLMs and this is kind of equivalent
6:48to building out a grid and then there's
6:51opex to serve that intelligence over
6:53APIs to all of us and this is done
6:56through metered access where we pay per
6:58million tokens or something like that
7:00and we have a lot of demands that are
7:01very utility- like demands out of this
7:03API we demand low latency high uptime
7:06consistent quality etc. In electricity,
7:08you would have a transfer switch. So you
7:10can transfer your electricity source
7:12from like grid and solar or battery or
7:14generator. In LLM, we have maybe open
7:16router and easily switch between the
7:18different types of LLMs that exist.
7:20Because the LLM are software, they don't
7:23compete for physical space. So it's okay
7:25to have basically like six electricity
7:26providers and you can switch between
7:28them, right? Because they don't compete
7:29in such a direct way. And I think what's
7:31also a little fascinating and we saw
7:33this in the last few days actually a lot
7:36of the LLMs went down and people were
7:38kind of like stuck and unable to work.
7:41And uh I think it's kind of fascinating
7:42to me that when the state-of-the-art
7:43LLMs go down, it's actually kind of like
7:45an intelligence brownout in the world.
7:47It's kind of like when the voltage is
7:49unreliable in the grid and uh the planet
7:52just gets dumber the more reliance we
7:55have on these models, which already is
7:56like really dramatic and I think will
7:58continue to grow. But LLM's don't only
8:00have properties of utilities. I think
8:02it's also fair to say that they have
8:03some properties of fabs. And the reason
8:06for this is that the capex required for
8:09building LLM is actually quite large. Uh
8:12it's not just like building some uh
8:14power station or something like that,
8:15right? You're investing a huge amount of
8:17money and I think the tech tree and uh
8:20for the technology is growing quite
8:22rapidly. So we're in a world where we
8:24have sort of deep tech trees, research
8:26and development secrets that are
8:28centralizing inside the LLM labs. Um and
8:32but I think the analogy muddies a little
8:34bit also because as I mentioned this is
8:36software and software is a bit less
8:38defensible because it is so malleable.
8:40And so um I think it's just an
8:43interesting kind of thing to think about
8:44potentially. There's many analogy
8:46analogies you can make like a 4
8:48nanometer process node maybe is
8:49something like a cluster with certain
8:51max flops. You can think about when
8:53you're use when you're using Nvidia GPUs
8:54and you're only doing the software and
8:56you're not doing the hardware. That's
8:57kind of like the fabless model. But if
8:59you're actually also building your own
9:00hardware and you're training on TPUs if
9:02you're Google, that's kind of like the
9:03Intel model where you own your fab. So I
9:05think there's some analogies here that
9:06make sense. But actually I think the
9:08analogy that makes the most sense
9:09perhaps is that in my mind LLM have very
9:12strong kind of analogies to operating
9:15systems. Uh in that this is not just
9:17electricity or water. It's not something
9:19that comes out of the tap as a
9:20commodity. uh this is these are now
9:22increasingly complex software ecosystems
9:25right so uh they're not just like simple
9:28commodities like electricity and it's
9:30kind of interesting to me that the
9:32ecosystem is shaping in a very similar
9:33kind of way where you have a few closed
9:36source providers like Windows or Mac OS
9:38and then you have an open source
9:39alternative like Linux and I think for u
9:42neural for LLMs as well we have a kind
9:45of a few competing closed source
9:47providers and then maybe the llama
9:49ecosystem is currently like maybe a
9:51close approximation to something that
9:53may grow into something like Linux.
9:55Again, I think it's still very early
9:56because these are just simple LLMs, but
9:58we're starting to see that these are
9:59going to get a lot more complicated.
10:01It's not just about the LLM itself. It's
10:02about all the tool use and the
10:03multiodalities and how all of that
10:05works. And so when I sort of had this
10:07realization a while back, I tried to
10:09sketch it out and it kind of seemed to
10:11me like LLMs are kind of like a new
10:12operating system, right? So the LLM is a
10:15new kind of a computer. It's sitting
10:17it's kind of like the CPU equivalent. uh
10:19the context windows are kind of like the
10:21memory and then the LLM is orchestrating
10:24memory and compute uh for problem
10:26solving um using all of these uh
10:29capabilities here and so definitely if
10:32you look at it looks very much like
10:34operating system from that perspective.
10:36Um, a few more analogies. For example,
10:38if you want to download an app, say I go
10:41to VS Code and I go to download, you can
10:43download VS Code and you can run it on
10:46Windows, Linux or or Mac in the same way
10:50as you can take an LLM app like cursor
10:53and you can run it on GPT or cloud or
10:55Gemini series, right? It's just a drop
10:57down. So, it's kind of like similar in
10:59that way as well.
11:00uh more analogies that I think strike me
11:02is that we're kind of like in this
11:041960sish
11:05era where LLM compute is still very
11:09expensive for this new kind of a
11:10computer and that forces the LLMs to be
11:13centralized in the cloud and we're all
11:15just uh sort of thing clients that
11:18interact with it over the network and
11:20none of us have full utilization of
11:22these computers and therefore it makes
11:24sense to use time sharing where we're
11:26all just you know a dimension of the
11:28batch when they're running the computer
11:30in the cloud. And this is very much what
11:32computers used to look like at during
11:33this time. The operating systems were in
11:35the cloud. Everything was streamed
11:36around and there was batching. And so
11:39the p the personal computing revolution
11:41hasn't happened yet because it's just
11:42not economical. It doesn't make sense.
11:44But I think some people are trying. And
11:46it turns out that Mac minis, for
11:48example, are a very good fit for some of
11:50the LLMs because it's all if you're
11:52doing batch one inference, this is all
11:53super memory bound. So this actually
11:55works.
11:56And uh I think these are some early
11:58indications maybe of personal computing.
12:00Uh but this hasn't really happened yet.
12:02It's not clear what this looks like.
12:03Maybe some of you get to invent what
12:05what this is or how it works or uh what
12:08this should what this should be. Maybe
12:10one more analogy that I'll mention is
12:12whenever I talk to Chach or some LLM
12:14directly in text, I feel like I'm
12:16talking to an operating system through
12:18the terminal. Like it's just it's it's
12:21text. It's direct access to the
12:22operating system. And I think a guey
12:24hasn't yet really been invented in like
12:26a general way like should chatt have a
12:29guey like different than just a tech
12:31bubbles. Uh certainly some of the apps
12:33that we're going to go into in a bit
12:35have guey but there's no like guey
12:38across all the tasks if that makes
12:40sense. Um there are some ways in which
12:43LLMs are different from kind of
12:45operating systems in some fairly unique
12:47way and from early computing. And I
12:49wrote about uh this one particular
12:52property that strikes me as very
12:54different uh this time around. It's that
12:57LLMs like flip they flip the direction
12:59of technology diffusion uh that is
13:02usually uh present in technology. So for
13:05example with electricity, cryptography,
13:07computing, flight, internet, GPS, lots
13:09of new transformative technologies that
13:10have not been around. Typically it is
13:12the government and corporations that are
13:14the first users because it's new and
13:16expensive etc. and it only later
13:18diffuses to consumer. Uh, but I feel
13:20like LLMs are kind of like flipped
13:22around. So maybe with early computers,
13:24it was all about ballistics and military
13:26use, but with LLMs, it's all about how
13:29do you boil an egg or something like
13:30that. This is certainly like a lot of my
13:32use. And so it's really fascinating to
13:33me that we have a new magical computer
13:35and it's like helping me boil an egg.
13:37It's not helping the government do
13:38something really crazy like some
13:40military ballistics or some special
13:42technology. Indeed, corporations are
13:43governments are lagging behind the
13:45adoption of all of us, of all of these
13:47technologies. So, it's just backwards
13:48and I think it informs maybe some of the
13:50uses of how we want to use this
13:52technology or like where are some of the
13:53first apps and so on.
13:56So, in summary so far, LLM labs LLMs. I
14:01think it's accurate language to use, but
14:03LLMs are complicated operating systems.
14:06They're circa 1960s in computing and
14:08we're redoing computing all over again.
14:10and they're currently available via time
14:11sharing and distributed like a utility.
14:13What is new and unprecedented is that
14:16they're not in the hands of a few
14:17governments and corporations. They're in
14:18the hands of all of us because we all
14:20have a computer and it's all just
14:21software and Chaship was beamed down to
14:24our computers like billions of people
14:26like instantly and overnight and this is
14:28insane. Uh and it's kind of insane to me
14:30that this is the case and now it is our
14:33time to enter the industry and program
14:34these computers. This is crazy. So I
14:37think this is quite remarkable. Before
14:39we program LLMs, we have to kind of like
14:42spend some time to think about what
14:43these things are. And I especially like
14:45to kind of talk about their psychology.
14:48So the way I like to think about LLMs is
14:50that they're kind of like people
14:51spirits. Um they are stoastic
14:54simulations of people. Um and the
14:56simulator in this case happens to be an
14:58auto reggressive transformer. So
14:59transformer is a neural net. Uh it's and
15:02it just kind of like is goes on the
15:04level of tokens. It goes chunk chunk
15:06chunk chunk chunk. And there's an almost
15:08equal amount of compute for every single
15:10chunk. Um and um this simulator of
15:14course is is just is basically there's
15:16some weights involved and we fit it to
15:19all of text that we have on the internet
15:20and so on. And you end up with this kind
15:22of a simulator and because it is trained
15:24on humans, it's got this emergent
15:26psychology that is humanlike. So the
15:28first thing you'll notice is of course
15:30uh LLM have encyclopedic knowledge and
15:32memory. uh and they can remember lots of
15:34things, a lot more than any single
15:36individual human can because they read
15:37so many things. It's it actually kind of
15:39reminds me of this movie Rainman, which
15:41I actually really recommend people
15:43watch. It's an amazing movie. I love
15:44this movie. Um and Dustin Hoffman here
15:46is an autistic savant who has almost
15:49perfect memory. So, he can read a he can
15:51read like a phone book and remember all
15:53of the names and phone numbers. And I
15:55kind of feel like LM are kind of like
15:57very similar. They can remember Shaw
15:58hashes and lots of different kinds of
16:00things very very easily. So they
16:02certainly have superpowers in some set
16:04in some respects. But they also have a
16:06bunch of I would say cognitive deficits.
16:08So they hallucinate quite a bit. Um and
16:11they kind of make up stuff and don't
16:13have a very good uh sort of internal
16:15model of self-nowledge, not sufficient
16:17at least. And this has gotten better but
16:19not perfect. They display jagged
16:21intelligence. So they're going to be
16:22superhuman in some problems solving
16:24domains. And then they're going to make
16:26mistakes that basically no human will
16:27make. like you know they will insist
16:29that 9.11 is greater than 9.9 or that
16:32there are two Rs in strawberry these are
16:34some famous examples but basically there
16:36are rough edges that you can trip on so
16:38that's kind of I think also kind of
16:40unique um they also kind of suffer from
16:43entrograde amnesia um so uh and I think
16:46I'm alluding to the fact that if you
16:48have a co-orker who joins your
16:49organization this co-orker will over
16:51time learn your organization and uh they
16:54will understand and gain like a huge
16:55amount of context on the organization
16:57and they go home and they sleep and they
16:59consolidate knowledge and they develop
17:01expertise over time. LLMs don't natively
17:03do this and this is not something that
17:04has really been solved in the R&D of
17:06LLM. I think um and so context windows
17:09are really kind of like working memory
17:10and you have to sort of program the
17:12working memory quite directly because
17:13they don't just kind of like get smarter
17:15by uh by default and I think a lot of
17:17people get tripped up by the analogies
17:19uh in this way. Uh in popular culture I
17:22recommend people watch these two movies
17:23uh Momento and 51st dates. In both of
17:26these movies, the protagonists, their
17:27weights are fixed and their context
17:29windows gets wiped every single morning
17:32and it's really problematic to go to
17:34work or have relationships when this
17:35happens and this happens to all the
17:37time. I guess one more thing I would
17:39point to is security kind of related
17:42limitations of the use of LLM. So for
17:44example, LLMs are quite gullible. Uh
17:46they are susceptible to prompt injection
17:48risks. They might leak your data etc.
17:50And so um and there's many other
17:52considerations uh security related. So,
17:55so basically long story short, you have
17:57to load your you have to load your you
18:00have to simultaneously think through
18:01this superhuman thing that has a bunch
18:03of cognitive deficits and issues. How do
18:05we and yet they are extremely like
18:07useful and so how do we program them and
18:10how do we work around their deficits and
18:12enjoy their superhuman powers.
18:15So what I want to switch to now is talk
18:17about the opportunities of how do we use
18:18these models and what are some of the
18:20biggest opportunities. This is not a
18:22comprehensive list just some of the
18:23things that I thought were interesting
18:24for this talk. The first thing I'm kind
18:26of excited about is what I would call
18:29partial autonomy apps. So for example,
18:32let's work with the example of coding.
18:34You can certainly go to chacht directly
18:36and you can start copy pasting code
18:38around and copyping bug reports and
18:40stuff around and getting code and copy
18:42pasting everything around. Why would you
18:44why would you do that? Why would you go
18:45directly to the operating system? It
18:47makes a lot more sense to have an app
18:48dedicated for this. And so I think many
18:50of you uh use uh cursor. I do as well.
18:53And uh cursor is kind of like the thing
18:56you want instead. You don't want to just
18:57directly go to the chash apt. And I
18:59think cursor is a very good example of
19:01an early LLM app that has a bunch of
19:03properties that I think are um useful
19:06across all the LLM apps. So in
19:08particular, you will notice that we have
19:09a traditional interface that allows a
19:12human to go in and do all the work
19:13manually just as before. But in addition
19:16to that, we now have this LLM
19:17integration that allows us to go in
19:19bigger chunks. And so some of the
19:21properties of LLM apps that I think are
19:23shared and useful to point out. Number
19:25one, the LLMs basically do a ton of the
19:28context management. Um, number two, they
19:31orchestrate multiple calls to LLMs,
19:33right? So in the case of cursor, there's
19:34under the hood embedding models for all
19:36your files, the actual chat models,
19:39models that apply diffs to the code, and
19:41this is all orchestrated for you. A
19:43really big one that uh I think also
19:46maybe not fully appreciated always is
19:48application specific uh GUI and the
19:50importance of it. Um because you don't
19:53just want to talk to the operating
19:54system directly in text. Text is very
19:56hard to read, interpret, understand and
19:59also like you don't want to take some of
20:00these actions natively in text. So it's
20:03much better to just see a diff as like
20:05red and green change and you can see
20:06what's being added is subtracted. It's
20:08much easier to just do command Y to
20:10accept or command N to reject. I
20:11shouldn't have to type it in text,
20:13right? So, a guey allows a human to
20:15audit the work of these fallible systems
20:17and to go faster. I'm going to come back
20:20to this point a little bit uh later as
20:21well. And the last kind of feature I
20:23want to point out is that there's what I
20:25call the autonomy slider. So, for
20:27example, in cursor, you can just do tap
20:29completion. You're mostly in charge. You
20:31can select a chunk of code and command K
20:33to change just that chunk of code. You
20:36can do command L to change the entire
20:37file. Or you can do command I which just
20:40you know let it rip do whatever you want
20:42in the entire repo and that's the sort
20:44of full autonomy agent agentic version
20:46and so you are in charge of the autonomy
20:48slider and depending on the complexity
20:50of the task at hand you can uh tune the
20:53amount of autonomy that you're willing
20:54to give up uh for that task maybe to
20:57show one more example of a fairly
20:58successful LLM app uh perplexity um it
21:03also has very similar features to what
21:04I've just pointed out to in cursor uh it
21:07packages up a lot of the information. It
21:08orchestrates multiple LLMs. It's got a
21:10GUI that allows you to audit some of its
21:13work. So, for example, it will site
21:15sources and you can imagine inspecting
21:17them. And it's got an autonomy slider.
21:18You can either just do a quick search or
21:20you can do research or you can do deep
21:22research and come back 10 minutes later.
21:24So, this is all just varying levels of
21:25autonomy that you give up to the tool.
21:27So, I guess my question is I feel like a
21:30lot of software will become partially
21:32autonomous. I'm trying to think through
21:33like what does that look like? And for
21:35many of you who maintain products and
21:36services, how are you going to make your
21:38products and services partially
21:40autonomous? Can an LLM see everything
21:42that a human can see? Can an LLM act in
21:45all the ways that a human could act? And
21:47can humans supervise and stay in the
21:49loop of this activity? Because again,
21:50these are fallible systems that aren't
21:52yet perfect. And what does a diff look
21:54like in Photoshop or something like
21:56that? You know, and also a lot of the
21:58traditional software right now, it has
22:00all these switches and all this kind of
22:01stuff that's all designed for human. All
22:03of this has to change and become
22:04accessible to LLMs.
22:07So, one thing I want to stress with a
22:09lot of these LLM apps that I'm not sure
22:11gets as much attention as it should is
22:14um we we're now kind of like cooperating
22:16with AIS and usually they are doing the
22:18generation and we as humans are doing
22:20the verification. It is in our interest
22:22to make this loop go as fast as
22:24possible. So, we're getting a lot of
22:25work done. There are two major ways that
22:28I think uh this can be done. Number one,
22:30you can speed up verification a lot. Um,
22:32and I think guies, for example, are
22:34extremely important to this because a
22:36guey utilizes your computer vision GPU
22:39in all of our head. Reading text is
22:41effortful and it's not fun, but looking
22:43at stuff is fun and it's it's just a
22:45kind of like a highway to your brain.
22:47So, I think guies are very useful for
22:49auditing systems and visual
22:51representations in general. And number
22:53two, I would say is we have to keep the
22:56AI on the leash. We I think a lot of
22:58people are getting way over excited with
23:00AI agents and uh it's not useful to me
23:03to get a diff of 10,000 lines of code to
23:05my repo. Like I have to I'm still the
23:07bottleneck, right? Even though that
23:0910,00 lines come out instantly, I have
23:11to make sure that this thing is not
23:12introducing bugs. It's just like and
23:15that it's doing the correct thing,
23:16right? And that there's no security
23:17issues and so on. So um I think that um
23:22yeah basically you we have to sort of
23:25like it's in our interest to make the
23:28the flow of these two go very very fast
23:30and we have to somehow keep the AI on
23:32the leash because it gets way too
23:33overreactive. It's uh it's kind of like
23:35this. This is how I feel when I do AI
23:37assisted coding. If I'm just bite coding
23:39everything is nice and great but if I'm
23:40actually trying to get work done it's
23:42not so great to have an overreactive uh
23:44agent doing all this kind of stuff. So
23:47this slide is not very good. I'm sorry,
23:48but I guess I'm trying to develop like
23:51many of you some ways of utilizing these
23:53agents in my coding workflow and to do
23:55AI assisted coding. And in my own work,
23:58I'm always scared to get way too big
23:59diffs. I always go in small incremental
24:02chunks. I want to make sure that
24:04everything is good. I want to spin this
24:06loop very very fast and um I sort of
24:09work on small chunks of single concrete
24:10thing. Uh and so I think many of you
24:13probably are developing similar ways of
24:14working with the with LLMs.
24:17Um, I also saw a number of blog posts
24:19that try to develop these best practices
24:22for working with LLMs. And here's one
24:24that I read recently and I thought was
24:25quite good. And it kind of discussed
24:26some techniques and some of them have to
24:28do with how you keep the AI on the
24:29leash. And so, as an example, if you are
24:32prompting, if your prompt is vague, then
24:34uh the AI might not do exactly what you
24:36wanted and in that case, verification
24:38will fail. You're going to ask for
24:40something else. If a verification fails,
24:42then you're going to start spinning. So
24:43it makes a lot more sense to spend a bit
24:45more time to be more concrete in your
24:46prompts which increases the probability
24:48of successful verification and you can
24:50move forward. And so I think a lot of us
24:52are going to end up finding um kind of
24:54techniques like this. I think in my own
24:56work as well I'm currently interested in
24:57uh what education looks like in um
25:00together with kind of like now that we
25:01have AI uh and LLMs what does education
25:04look like? And I think a a large amount
25:07of thought for me goes into how we keep
25:09AI on the leash. I don't think it just
25:11works to go to chat and be like, "Hey,
25:13teach me physics." I don't think this
25:14works because the AI is like gets lost
25:16in the woods. And so for me, this is
25:18actually two separate apps. For example,
25:20there's an app for a teacher that
25:22creates courses and then there's an app
25:24that takes courses and serves them to
25:26students. And in both cases, we now have
25:29this intermediate artifact of a course
25:31that is auditable and we can make sure
25:32it's good. We can make sure it's
25:33consistent. and the AI is kept on the
25:35leash with respect to a certain
25:37syllabus, a certain like um progression
25:40of projects and so on. And so this is
25:42one way of keeping the AI on leash and I
25:44think has a much higher likelihood of
25:45working and the AI is not getting lost
25:47in the woods.
25:49One more kind of analogy I wanted to
25:51sort of allude to is I'm not I'm no
25:54stranger to partial autonomy and I kind
25:56of worked on this I think for five years
25:57at Tesla and this is also a partial
26:00autonomy product and shares a lot of the
26:01features like for example right there in
26:03the instrument panel is the GUI of the
26:05autopilot so it's showing me what the
26:07what the neural network sees and so on
26:09and we have the autonomy slider where
26:10over the course of my tenure there we
26:13did more and more autonomous tasks for
26:15the user and maybe the story that I
26:18wanted to tell very briefly is uh
26:21actually the first time I drove a
26:22self-driving vehicle was in 2013 and I
26:25had a friend who worked at Whimo and uh
26:27he offered to give me a drive around
26:29Palo Alto. I took this picture using
26:31Google Glass at the time and many of you
26:33are so young that you might not even
26:35know what that is. Uh but uh yeah, this
26:37was like all the rage at the time. And
26:39we got into this car and we went for
26:40about a 30-minute drive around Palo Alto
26:42highways uh streets and so on. And this
26:45drive was perfect. There was zero
26:46interventions and this was 2013 which is
26:49now 12 years ago. And it kind of struck
26:52me because at the time when I had this
26:54perfect drive, this perfect demo, I felt
26:56like, wow, self-driving is imminent
26:59because this just worked. This is
27:00incredible. Um, but here we are 12 years
27:03later and we are still working on
27:04autonomy. Um, we are still working on
27:07driving agents and even now we haven't
27:09actually like really solved the problem.
27:10like you may see Whimos going around and
27:12they look driverless but you know
27:14there's still a lot of teleoperation and
27:16a lot of human in the loop of a lot of
27:18this driving so we still haven't even
27:20like declared success but I think it's
27:22definitely like going to succeed at this
27:24point but it just took a long time and
27:26so I think like like this is software is
27:29really tricky I think in the same way
27:31that driving is tricky and so when I see
27:34things like oh 2025 is the year of
27:36agents I get very concerned and I kind
27:38of feel like you know this is the decade
27:41of agents and this is going to be quite
27:44some time. We need humans in the loop.
27:45We need to do this carefully. This is
27:47software. Let's be serious here. One
27:51more kind of analogy that I always think
27:52through is the Iron Man suit. Uh I think
27:56this is I always love Iron Man. I think
27:58it's like so um correct in a bunch of
28:01ways with respect to technology and how
28:02it will play out. And what I love about
28:04the Iron Man suit is that it's both an
28:05augmentation and Tony Stark can drive it
28:08and it's also an agent. And in some of
28:10the movies, the Iron Man suit is quite
28:11autonomous and can fly around and find
28:13Tony and all this kind of stuff. And so
28:15this is the autonomy slider is we can be
28:17we can build augmentations or we can
28:19build agents and we kind of want to do a
28:21bit of both. But at this stage I would
28:23say working with fallible LLMs and so
28:25on. I would say you know it's less Iron
28:29Man robots and more Iron Man suits that
28:31you want to build. It's less like
28:33building flashy demos of autonomous
28:35agents and more building partial
28:36autonomy products. And these products
28:39have custom gueies and UIUX. And we're
28:41trying to um and this is done so that
28:43the generation verification loop of the
28:45human is very very fast. But we are not
28:48losing the sight of the fact that it is
28:49in principle possible to automate this
28:51work. And there should be an autonomy
28:52slider in your product. And you should
28:54be thinking about how you can slide that
28:55autonomy slider and make your product uh
28:58sort of um more autonomous over time.
29:01But this is kind of how I think there's
29:02lots of opportunities in these kinds of
29:04products. I want to now switch gears a
29:06little bit and talk about one other
29:08dimension that I think is very unique.
29:09Not only is there a new type of
29:11programming language that allows for
29:12autonomy in software but also as I
29:15mentioned it's programmed in English
29:16which is this natural interface and
29:19suddenly everyone is a programmer
29:20because everyone speaks natural language
29:22like English. So this is extremely
29:24bullish and very interesting to me and
29:26also completely unprecedented. I would
29:28say it it used to be the case that you
29:29need to spend five to 10 years studying
29:31something to be able to do something in
29:32software. this is not the case anymore.
29:35So, I don't know if by any chance anyone
29:37has heard of vibe coding.
29:40Uh, this this is the tweet that kind of
29:42like introduced this, but I'm told that
29:44this is now like a major meme. Um, fun
29:46story about this is that I've been on
29:49Twitter for like 15 years or something
29:51like that at this point and I still have
29:53no clue which tweet will become viral
29:56and which tweet like fizzles and no one
29:58cares. And I thought that this tweet was
30:00going to be the latter. I don't know. It
30:01was just like a shower of thoughts. But
30:03this became like a total meme and I
30:05really just can't tell. But I guess like
30:06it struck a chord and it gave a name to
30:08something that everyone was feeling but
30:10couldn't quite say in words. So now
30:13there's a Wikipedia page and everything.
30:17This is like
30:18[Applause]
30:25yeah this is like a major contribution
30:27now or something like that. So,
30:30um, so Tom Wolf from HuggingFace shared
30:32this beautiful video that I really love.
30:34Um,
30:37these are kids vibe coding.
30:42And I find that this is such a wholesome
30:44video. Like, I love this video. Like,
30:46how can you look at this video and feel
30:48bad about the future? The future is
30:49great.
30:52I think this will end up being like a
30:53gateway drug to software development.
30:56Um, I'm not a doomer about the future of
30:59the generation and I think yeah, I love
31:02this video. So, I tried by coding a
31:04little bit uh as well because it's so
31:07fun. Uh, so bike coding is so great when
31:09you want to build something super duper
31:10custom that doesn't appear to exist and
31:12you just want to wing it because it's a
31:13Saturday or something like that. So, I
31:15built this uh iOS app and I don't I
31:18can't actually program in Swift, but I
31:20was really shocked that I was able to
31:21build like a super basic app and I'm not
31:23going to explain it. It's really uh
31:24dumb, but uh I kind of like this was
31:27just like a day of work and this was
31:28running on my phone like later that day
31:30and I was like, "Wow, this is amazing."
31:32I didn't have to like read through Swift
31:33for like five days or something like
31:35that to like get started. I also
31:38vipcoded this app called Menu Genen. And
31:40this is live. You can try it in
31:41menu.app. And I basically had this
31:44problem where I show up at a restaurant,
31:45I read through the menu, and I have no
31:46idea what any of the things are. And I
31:48need pictures. So this doesn't exist. So
31:51I was like, "Hey, I'm going to bite code
31:52it." So, um, this is what it looks like.
31:55You go to menu.app,
31:58um, and, uh, you take a picture of a of
32:01a menu and then menu generates the
32:03images and everyone gets $5 in credits
32:06for free when you sign up. And
32:08therefore, this is a major cost center
32:10in my life. So, this is a negative
32:13negative uh, revenue app for me right
32:16now.
32:17I've lost a huge amount of money on
32:19menu.
32:21Okay. But the fascinating thing about
32:23menu genen for me is that the code of
32:28the v the vite coding part the code was
32:30actually the easy part of v of v coding
32:32menu and most of it actually was when I
32:35tried to make it real so that you can
32:36actually have authentication and
32:37payments and the domain name and averal
32:39deployment. This was really hard and all
32:41of this was not code. All of this devops
32:44stuff was in me in the browser clicking
32:47stuff and this was extreme slo and took
32:49another week. So it was really
32:51fascinating that I had the menu genen um
32:54basically demo working on my laptop in a
32:57few hours and then it took me a week
32:59because I was trying to make it real and
33:01the reason for this is this was just
33:02really annoying. Um, so for example, if
33:05you try to add Google login to your web
33:07page, I know this is very small, but
33:09just a huge amount of instructions of
33:11this clerk library telling me how to
33:13integrate this. And this is crazy. Like
33:15it's telling me go to this URL, click on
33:17this dropdown, choose this, go to this,
33:19and click on that. And it's like telling
33:21me what to do. Like a computer is
33:22telling me the actions I should be
33:24taking. Like you do it. Why am I doing
33:26this?
33:28What the hell?
33:31I had to follow all these instructions.
33:33This was crazy. So I think the last part
33:36of my talk therefore focuses on can we
33:39just build for agents? I don't want to
33:41do this work. Can agents do this? Thank
33:44you.
33:46Okay. So roughly speaking, I think
33:48there's a new category of consumer and
33:50manipulator of digital information. It
33:53used to be just humans through GUIs or
33:55computers through APIs. And now we have
33:57a completely new thing and agents are
34:00they're computers but they are humanlike
34:02kind of right they're people spirits
34:04there's people spirits on the internet
34:05and they need to interact with our
34:06software infrastructure like can we
34:08build for them it's a new thing so as an
34:10example you can have robots.txt on your
34:12domain and you can instruct uh or like
34:15advise I suppose um uh web crawlers on
34:18how to behave on your website in the
34:19same way you can have maybe lm.txt txt
34:21file which is just a simple markdown
34:23that's telling LLMs what this domain is
34:25about and this is very readable to a to
34:28an LLM. If it had to instead get the
34:30HTML of your web page and try to parse
34:32it, this is very errorprone and
34:33difficult and will screw it up and it's
34:35not going to work. So we can just
34:36directly speak to the LLM. It's worth
34:38it. Um a huge amount of documentation is
34:41currently written for people. So you
34:42will see things like lists and bold and
34:45pictures and this is not directly
34:47accessible by an LLM. So I see some of
34:51the services now are transitioning a lot
34:52of the their docs to be specifically for
34:54LLMs. So Versell and Stripe as an
34:57example are early movers here but there
34:59are a few more that I've seen already
35:01and they offer their documentation in
35:04markdown. Markdown is super easy for LMS
35:06to understand. This is great. Um maybe
35:10one simple example from from uh my
35:12experience as well. Maybe some of you
35:14know three blue one brown. He makes
35:15beautiful animation videos on YouTube.
35:19[Applause]
35:23Yeah, I love this library. So that he
35:25wrote uh Manon and I wanted to make my
35:27own and uh there's extensive
35:30documentations on how to use manon and
35:32so I didn't want to actually read
35:34through it. So I copy pasted the whole
35:35thing to an LLM and I described what I
35:37wanted and it just worked out of the box
35:39like LLM just bcoded me an animation
35:41exactly what I wanted and I was like wow
35:43this is amazing. So if we can make docs
35:45legible to LLMs, it's going to unlock a
35:48huge amount of um kind of use and um I
35:51think this is wonderful and should
35:52should happen more. The other thing I
35:55wanted to point out is that you do
35:56unfortunately have to it's not just
35:57about taking your docs and making them
35:58appear in markdown. That's the easy
36:00part. We actually have to change the
36:01docs because anytime your docs say click
36:04this is bad. An LLM will not be able to
36:06natively take this action right now. So,
36:09Verscell, for example, is replacing
36:11every occurrence of click with an
36:13equivalent curl command that your LM
36:15agent could take on your behalf. Um, and
36:18so I think this is very interesting. And
36:19then, of course, there's a model context
36:21protocol from Enthropic. And this is
36:23also another way, it's a protocol of
36:24speaking directly to agents as this new
36:26consumer and manipulator of digital
36:28information. So, I'm very bullish on
36:29these ideas. The other thing I really
36:31like is a number of little tools here
36:33and there that are helping ingest data
36:36that in like very LLM friendly formats.
36:38So for example, when I go to a GitHub
36:40repo like my nanoGPT repo, I can't feed
36:42this to an LLM and ask questions about
36:44it uh because it's you know this is a
36:46human interface on GitHub. So when you
36:48just change the URL from GitHub to get
36:50ingest then uh this will actually
36:52concatenate all the files into a single
36:54giant text and it will create a
36:55directory structure etc. And this is
36:57ready to be copy pasted into your
36:59favorite LLM and you can do stuff. Maybe
37:01even more dramatic example of this is
37:03deep wiki where it's not just the raw
37:05content of these files. uh this is from
37:08Devon but also like they have Devon
37:10basically do analysis of the GitHub repo
37:12and Devon basically builds up a whole
37:14docs uh pages just for your repo and you
37:18can imagine that this is even more
37:19helpful to copy paste into your LLM. So
37:22I love all the little tools that
37:23basically where you just change the URL
37:24and it makes something accessible to an
37:26LLM. So this is all well and great and u
37:29I think there should be a lot more of
37:30it. One more note I wanted to make is
37:32that it is absolutely possible that in
37:35the future LLMs will be able to this is
37:38not even future this is today they'll be
37:39able to go around and they'll be able to
37:40click stuff and so on but I still think
37:42it's very worth u basically meeting LLM
37:46halfway LLM's halfway and making it
37:48easier for them to access all this
37:49information uh because this is still
37:51fairly expensive I would say to use and
37:54uh a lot more difficult and so I do
37:56think that lots of software there will
37:58be a long tail where it won't like adapt
38:00apps because these are not like live
38:02player sort of repositories or digital
38:04infrastructure and we will need these
38:06tools. Uh but I think for everyone else
38:08I think it's very worth kind of like
38:09meeting in some middle point. So I'm
38:11bullish on both if that makes sense.
38:14So in summary, what an amazing time to
38:17get into the industry. We need to
38:18rewrite a ton of code. A ton of code
38:20will be written by professionals and by
38:23coders. These LLMs are kind of like
38:25utilities, kind of like fabs, but
38:27they're kind of especially like
38:28operating systems. But it's so early.
38:30It's like 1960s of operating systems and
38:34uh and I think a lot of the analogies
38:36cross over. Um and these LMS are kind of
38:38like these fallible uh you know people
38:41spirits that we have to learn to work
38:43with. And in order to do that properly,
38:45we need to adjust our infrastructure
38:47towards it. So when you're building
38:48these LLM apps, I describe some of the
38:50ways of working effectively with these
38:52LLMs and some of the tools that make
38:54that uh kind of possible and how you can
38:57spin this loop very very quickly and
38:59basically create partial tunneling
39:00products and then um yeah, a lot of code
39:03has to also be written for the agents
39:04more directly. But in any case, going
39:07back to the Iron Man suit analogy, I
39:09think what we'll see over the next
39:10decade roughly is we're going to take
39:12the slider from left to right. And I'm
39:15very interesting. It's going to be very
39:17interesting to see what that looks like.
39:19And I can't wait to build it with all of
39:21you. Thank you.

About this transcript

This page contains the full transcript of Andrej Karpathy: Software Is Changing (Again) by Y Combinator, generated from the public captions YouTube serves with the video. The transcript has 8,255 words across 1,139 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 translate the transcript into 100+ languages.

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, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.