The Sweet Side of the Bitter Lesson — Thariq Shihipar, Independent — Transcript
Full transcript
- 0:12Hey guys,
- 0:13I'm Tharak. Nice to meet you.
- 0:16Yeah,
- 0:17I wanted to talk to you about a couple
- 0:20things I've been thinking about since I
- 0:21left Anthropic.
- 0:23It's been a wild year and
- 0:26there a bunch of reflections I've had.
- 0:30And my conclusion is
- 0:32about, you know, what I'm calling the
- 0:33sweet side of the bitter lesson
- 0:36and the ways I've been wrong. So I I
- 0:37think I got a lot of things right this
- 0:39year.
- 0:41I think
- 0:43I was right about models being able to
- 0:45edit and generate videos. I was talking
- 0:47about this like
- 0:49sometime in like June last year.
- 0:51I was right about models being able to
- 0:53use HTML to understand their work.
- 0:56And I was right about models being able
- 0:59to help you understand and figure out
- 1:02your unknown to help like plan and learn
- 1:05about things. So these were all things I
- 1:06was like right about and I was happy and
- 1:09I I think that like
- 1:11a lot of you might be incorporating this
- 1:12sort of stuff in your work.
- 1:14But I've been wrong a lot too and I I
- 1:16think this is one of those things in AI
- 1:18where you sort of just like memory hole
- 1:20the things that you're wrong about and
- 1:21you're like, oh like I was just always
- 1:23expecting, you know, AI to be great.
- 1:26But here are some of the things I've
- 1:27been wrong about. So first like Bash is
- 1:30all you need. I I think there was this
- 1:31like moment where the agents were really
- 1:33good at using Bash and Claude Code was
- 1:36built around the Bash tool. And so, you
- 1:39know, people are like, well, we'll just
- 1:40turn everything into a CLI. Like, you
- 1:42know, like there is just going to be
- 1:44Bash and CLIs and that's going to like
- 1:46take over the world.
- 1:48That's actually kind of not how it
- 1:49turned out. Like I actually prefer MCP
- 1:51to CLIs now, you know, but now there's
- 1:53also code mode and skills and
- 1:55progressive disclosure. Like there are
- 1:56so many other like ways that agents can,
- 2:00you know, act on the world, right? And
- 2:03the fact that they were using Bash was
- 2:05just like this weird remnant of, you
- 2:07know, like them being trained in a
- 2:09certain way at the start.
- 2:11Then there's like React is all you need.
- 2:14I think for a moment it seemed like the
- 2:16models were really really good at
- 2:17writing React. And so, you know, if you
- 2:20wanted to create an AI enabled app you
- 2:23should figure out how to use React. Like
- 2:24you should use React Native, you should
- 2:26use, you know, like you should just make
- 2:28everything on the web.
- 2:30And
- 2:32that also has not been true. I I think
- 2:33like now, you know, models can write it
- 2:35in HTML, they can write it in React,
- 2:37they can write it in Swift, they can
- 2:39write it in the Figma proprietary JSON
- 2:41format.
- 2:43And they'll all look roughly the same.
- 2:45Right? And so
- 2:47I was afraid that like you know the
- 2:48models have been trained on React and so
- 2:49we'd just be writing React forever and
- 2:51React would be like the last language
- 2:53almost, right? And that's almost
- 2:54certainly not true, right? I think a lot
- 2:56of people are writing their own
- 2:57languages now.
- 3:01And I think one of the other things I
- 3:02was really concerned about was like
- 3:04diffusion models will kind of like eat
- 3:05art. Like it just be like, you know, we
- 3:07wouldn't have any control over these
- 3:09models and or the kind of art that these
- 3:11models are creating, but it would just
- 3:13like create the most beautiful stuff.
- 3:14And that hasn't been true at all. Like I
- 3:17think now LLMs can make art through
- 3:19code, through controlling your computer
- 3:21and using, you know, Blender or
- 3:24Photoshop or yeah or Figma or whatever
- 3:27you'd like, right? And so
- 3:30these were all sort of like fears I had
- 3:33about like I think
- 3:34the bitter lesson, you know, our alien
- 3:37intelligence, the Shoggoth, right? Like
- 3:39like these are all words we use I think
- 3:41to describe this kind of like fear that
- 3:45the models will just have this like way
- 3:47of doing things and it will be weird and
- 3:49alien to us and we just need to accept
- 3:51it. You know, like you just have to like
- 3:53submit to this like foreign
- 3:55intelligence.
- 3:56But that's just like I don't think
- 3:58that's what happened, you know?
- 4:01And so I think that like the question
- 4:03come back to a lot is like okay, like
- 4:05how can we be AGI piled and human
- 4:06centered and what can we learn about you
- 4:09know, this progression what we like you
- 4:11know, what can I learn about what I was
- 4:13wrong about, right?
- 4:14So I think this is really really
- 4:15important because software engineers are
- 4:17like the leading edge of like AI
- 4:19adoption, right? Like any thing that is
- 4:21happening to software will happen to
- 4:23knowledge work and other parts of the
- 4:24economy and so I think how we deal and
- 4:28engage with uh with AI is um
- 4:31yeah, how other people will as well and
- 4:33you know, I really enjoyed the different
- 4:34talks here talking about things like
- 4:36yeah, fighting slop. I think you know,
- 4:38we're the first defense uh for that,
- 4:40right? Or like being more ambitious.
- 4:42Um, what I like to think about it as AI
- 4:45is both at the same time an industrial
- 4:48revolution and a Renaissance, right?
- 4:51And I think this year was really about
- 4:54the industrial revolution, right? Like
- 4:56we use the word software factories very
- 4:58often
- 4:59um
- 5:00and uh like just sort of like turning
- 5:03software from this like
- 5:05craft that was very unreliable and hard
- 5:08to execute on into something that like
- 5:12is not the limiting factor, right? Um,
- 5:15and this sort of like industrialization
- 5:17is absolutely happening. This is
- 5:19definitely something that is happening
- 5:21with with software, but at the same time
- 5:23there's also kind of a Renaissance
- 5:25happening, right? Where people want and
- 5:28uh higher and higher quality software
- 5:30and every marginal bit of like
- 5:34creativity, utility, like spirit, like
- 5:37soul that you put into it is like is
- 5:39useful and like economically valuable,
- 5:41right? And so simultaneously we've got
- 5:44both of these things that are very hard
- 5:45to juggle and and sort of like
- 5:47hold in your head at the same time, uh
- 5:49but I think it is very true and uh
- 5:53yeah, we're doing both of them. So
- 5:55often times when we talk about AI
- 5:56capabilities, we talk about like the
- 5:58jagged frontier, right? And so uh
- 6:01you know, you maybe like you go back two
- 6:03years and you're like AI asking AI to
- 6:04like this function for you,
- 6:06um or, you know, like like a year and
- 6:09half ago, a year ago, it's like build
- 6:10this feature, then you're like, okay,
- 6:11build this like to do app, build like a
- 6:13kind of simple to do app, and then clone
- 6:15Slack for me now, right? And And And now
- 6:18it's like, okay, build Mario Kart from
- 6:19scratch, you know, and then then like
- 6:21port it to like the Wii or something.
- 6:23There's like the capability increases
- 6:26are the things that you see, right?
- 6:28Where, you know, whenever you look at
- 6:30the evals and benchmarks and stuff, you
- 6:33see this like just
- 6:34um
- 6:35this jagged frontier of like
- 6:36capabilities. But I also like to talk
- 6:39about like the smooth interior
- 6:43where the the way Whoops.
- 6:46the way the models
- 6:48Is there an animation playing here? Um
- 6:51the smooth interior is like the idea is
- 6:54that the models are um
- 6:57as they get better, the the animation
- 6:59here is like this jagged frontier and
- 7:01this circle in between. And so uh the
- 7:04circle sort of represents all of the
- 7:06things that the models are too are smart
- 7:09enough to do. And so like examples being
- 7:11like uh writing React code, right? Or
- 7:14like using Bash. And what you see is
- 7:17that like it's not jagged below the
- 7:20frontier, right? Like if it can write
- 7:22React, it can write Figma, it can write
- 7:24JSON, it can do sort of like all of
- 7:27these capabilities all together. And
- 7:29there's like a breadth of tasks that the
- 7:30models are really really good at. And
- 7:32this is something that's also I think
- 7:34really like important and um for me even
- 7:38more impactful because I don't really
- 7:39care that they can build Mario Kart from
- 7:42scratch anymore, right? What I really
- 7:44care about is like, hey, is the software
- 7:46that we're building, is it high quality?
- 7:48Is it worthy of like a Renaissance? You
- 7:51know what I mean? So
- 7:53um
- 7:55yeah, like I think that like, you know,
- 7:56you can do things the way you want. So
- 7:58like if you're building mockup, you
- 7:59could be like, hey, can I get an HTML?
- 8:01Can I get a React? Can I get a Figma
- 8:03file? Can I computer use my computer and
- 8:05use After Effects and make a video?
- 8:07There are all these different ways of
- 8:09getting the like final result. And I've
- 8:12been really happy with this. I really
- 8:14was afraid that there would just be like
- 8:16one thing that the models love doing and
- 8:17we'd only be doing that.
- 8:19Um
- 8:20and you know like taken to the limit
- 8:23kind of like, you know, we talk about
- 8:24this a lot but like, you know,
- 8:26everything we've talked about so far
- 8:28about like, you know, how you build like
- 8:29these harnesses and software factories
- 8:31and um these like the scaffolding for
- 8:33the models.
- 8:35Even things like, you know, for a while
- 8:37we were also like, hey um the model
- 8:40you'll like
- 8:42your job is to QA the model, you know,
- 8:44or review the code. And that's clearly
- 8:46not the case, right? I think as the
- 8:47models get smarter and smarter um it
- 8:50will become more of this like smooth
- 8:52interior, right? Where they're just
- 8:54doing things well, right? And like the
- 8:57perfect model you can imagine won't need
- 8:59to like really verify its work or scan
- 9:01for security vulnerabilities. We'll just
- 9:03like write the code and be like, hey I'm
- 9:05I'm done now. Um and you will check and
- 9:08you'll like, oh you're right. Like, you
- 9:09know, according to the spec and the
- 9:11goals that you that we did you wrote you
- 9:13did it correctly. Um
- 9:16and uh I think that's actually very
- 9:17exciting.
- 9:18Um one of one of the ways I like to sort
- 9:20of think about visualizing how how much
- 9:23the model knows of this eval. I think
- 9:25Aparthy um posted it.
- 9:28But basically you ask any I think these
- 9:30are
- 9:31these are Claude models. You ask a model
- 9:34uh you give it a longitude latitude and
- 9:35you say like, hey is this point
- 9:38uh earth or water, right? Uh ground
- 9:41sorry land or water. And it from that
- 9:43you plot a map of the earth,
- 9:45right? And yeah this is Opus uh 4.6
- 9:50uh is somewhere over there, right? With
- 9:52quite a good software engineering model,
- 9:54kind of like a blurry view. And now you
- 9:56look at Opus 5.5 and it's like almost
- 9:58perfect, right? Like it knows
- 10:00every piece of, you know, these like
- 10:04these coordinates. It knows like the
- 10:06entire world basically. And so this I I
- 10:08just gives you an idea besides like, you
- 10:10know, the benchmarks about like the the
- 10:13the jagged frontier. I feel like
- 10:15that's less and less important to us,
- 10:17and more important is just like the
- 10:18general knowledge that the models have.
- 10:21And what's exciting, I think about this,
- 10:23is that we can use intelligence to
- 10:27understand and shape how we use
- 10:28intelligence, right? And I think this
- 10:30that will be like the theme of the next
- 10:32year, right? Where this year I think was
- 10:34really about the industrialization of
- 10:36software and and creating this like
- 10:38reliable way to make software. Um,
- 10:41but what does that look like? So I have
- 10:43been building a game.
- 10:45Um, it's this like thing I've been
- 10:47posting on Twitter about. Um, it's a 2D
- 10:50like isometric game with like it's kind
- 10:52of like a fighting game inspired by
- 10:54things like Avatar and Smash Brothers.
- 10:57Uh, and
- 10:59I have kind of like it's fun, but it's
- 11:01also quite frustrating to do. Like I'm
- 11:03really trying to like, you know, like be
- 11:05like a great vibe coder here. Um, I
- 11:07don't really know what each character
- 11:08does sometimes because like I will say
- 11:11like, hey, I want this character to
- 11:12throw a rock. And then it'll be like, oh
- 11:14yeah, great, but like there are all
- 11:16these like different things that can
- 11:17happen when you're throwing a rock. Like
- 11:19this character can also summon a wall.
- 11:21So does the rock go through the wall or
- 11:24not? And like how what speed or like
- 11:26what what about size and charge and all
- 11:28these questions? Um,
- 11:31and I I find that like the model just
- 11:33makes a bunch of choices for me and then
- 11:34I try it out and I'm like, wow, I have
- 11:36no idea what this character actually
- 11:37does.
- 11:38I get distracted pretty easily. I don't
- 11:40feel in the zone very much. You know, I
- 11:42I sort of send off a prompt and I start
- 11:44doing something else and I come back.
- 11:46Um,
- 11:47and I honestly make less progress
- 11:49because of these things, right? And so
- 11:51uh, I think this is frustrating and I
- 11:53think for a while I was like, man, I
- 11:55guess this is just like how you work
- 11:56with AI. This is just like what it's
- 11:57going to be like, you know? But I think
- 11:59reflecting I'm like, man, I I don't
- 12:01think that's right. I think that like,
- 12:02you know,
- 12:03we have this like the AI can work in the
- 12:06way that we want, right? And so yeah, I
- 12:09don't think it's acceptable. I don't
- 12:10think that this is what leads to the
- 12:12Renaissance, right? I don't think this
- 12:13is what leads to like incredible
- 12:14software.
- 12:16And part of it, you know, going back to
- 12:18what Dex said is like like great work is
- 12:21just about like this this messy process,
- 12:23right? Of like don't know like exactly
- 12:25what you want. It's more about the
- 12:27choices you made and what wasn't
- 12:29involved or wasn't what wasn't included
- 12:32versus like what was included.
- 12:34And so
- 12:37yeah, I think there's like
- 12:40you you like kind of just have to be an
- 12:41expert. You have to like just really
- 12:42know your stuff in order to make
- 12:44something great, right? In order to like
- 12:46make something worthy of a Renaissance,
- 12:48right? And I just like when I look at
- 12:49something I'm vibe coding, I'm like,
- 12:50dude, I have no idea. Like I just don't
- 12:53really get it, you know?
- 12:55And so
- 12:58what what am I doing? How am I fixing
- 12:59this?
- 13:01One of the ways I'm doing it, and I'm
- 13:02still quite early on. I'm not saying
- 13:04that I've like figured it all out, but I
- 13:06think one of the ways I'm doing it is
- 13:07that I have two parts of my codebase.
- 13:09I've got like this like engine part of
- 13:10the codebase, and then I have a folder
- 13:12like I think I call it spec, but I think
- 13:14I'm just calling it human or something.
- 13:16And that folder is the code that I read,
- 13:19you know? And it's like this code that's
- 13:22linted and like there's all these rules
- 13:25for Claude about like what that code
- 13:27looks like. It's supposed to feel like
- 13:28English.
- 13:30It's kind of like a spec, and I'll I'll
- 13:32show you what it looks like.
- 13:34But it's at the density where if I just
- 13:36read this code, I know what's happening
- 13:39and I can edit it, I can tweak it.
- 13:42But more importantly, I just like, you
- 13:44know, can have a bit of a ground truth
- 13:47of like what the what my game looks
- 13:48like. And
- 13:51so let's do let's do like a quick demo
- 13:53here.
- 13:55By demo I guess I mean I'm just going to
- 13:57whoops. This is my eTalk. Okay.
- 14:00Here we go. Okay. So
- 14:02this is my game. I've got in here this
- 14:04folder game, and then I've got here this
- 14:06code. This is the like agent code,
- 14:08right? Like this is the code where I'm
- 14:09like, dude, I I don't know how you know
- 14:11cameras work or whatever. I you you like
- 14:13there's a bunch of code here that I
- 14:14don't
- 14:15But then in this code, I see have all my
- 14:18characters and I have their abilities.
- 14:20And so I can go in here
- 14:22and I can say like, okay, this is the
- 14:25rock move, right? Um and inside of this
- 14:28move, there's a bunch of ways to tune
- 14:30it. Um it's well commented. It uses
- 14:33things like meters, you know what I
- 14:34mean? Instead of like
- 14:36X like 0.5 or whatever. Um
- 14:40it uh tries to like create a bunch of
- 14:43different abstractions along the way.
- 14:45And so I can see that like one of the
- 14:46things that, you know, is useful is that
- 14:49like there are three ways to throw for
- 14:51this character to throw a rock, right?
- 14:52Like it can uh grab it from the ground,
- 14:54it can grab it from the wall or
- 14:56something. Um
- 14:58and I can see sort of like the
- 15:01um
- 15:02the true logic of the character, right?
- 15:05I I don't really know how the engine
- 15:07works or how, you know, like this is
- 15:10implemented underneath. But I've
- 15:12essentially created this like almost a
- 15:13custom scripting language for my game.
- 15:16Um and so that means like, you know,
- 15:18when I'm playing this game,
- 15:19um I can tweak little things. You know,
- 15:21I'm still I'm still like working on it.
- 15:23So like I can say I want the rock to be
- 15:25a lot bigger. So I can say like um
- 15:28you know, 1.5, but I'm using an actual
- 15:30IDE. Uh and so if I go here to my game,
- 15:33uh you know, I can see that like I have
- 15:36a rock, I can charge it, gets really big
- 15:38and I throw it, right? Um but I'm like
- 15:41kind of it it it's really interesting
- 15:43cuz I'm also starting to use cursor more
- 15:46and I'm using the tab model more, you
- 15:47know what I mean? Because now now I can
- 15:49be like, oh hey, what if I just want to
- 15:51say like,
- 15:52you know, uh if you summon a wall in the
- 15:56middle, like queue it up as an action,
- 15:59and then cursor will sort of like, you
- 16:00know, do the the tab complete here. Um
- 16:05and this is kind of like it's kind of
- 16:06like spec driven development, you know,
- 16:08but the spec is this like this this
- 16:11scripting language basically.
- 16:14And um
- 16:15yeah, it's like
- 16:16is this the final way I'm going to write
- 16:18game? I'm not sure. It's like, you know,
- 16:20in some ways,
- 16:21uh, when I need to do something
- 16:22complicated, I still need to ask Claude
- 16:24to do it, right? And I need to ask to
- 16:26Claude to like make the primitives, but
- 16:28it's actually really easy for me to
- 16:29understand what's happening and I feel
- 16:31like much more in control and while like
- 16:34Claude is working, I can also sort of
- 16:35review it and work with it.
- 16:37Um,
- 16:38and that's been really exciting. And so
- 16:40I think the
- 16:42takeaway for me has been that like
- 16:45um
- 16:47the takeaway for me has been that like I
- 16:49now have to think about the way
- 16:52the best way for me to deliver high
- 16:54quality software, right? And I'm not as
- 16:58constrained and limited by like the
- 17:00model's capabilities and like what it
- 17:01likes to write code in or something. I'm
- 17:03just not thinking about like I need to
- 17:05think about like how can I deliver great
- 17:07software?
- 17:08And so I think if we like really
- 17:10internalize that, I think then this the
- 17:13next year will be about a renaissance of
- 17:15software. And you know, I think many
- 17:17people have talked
- 17:19today and you know, this conference
- 17:21about how to be more ambitious and I
- 17:22don't think anything any of this is
- 17:23quite new, but I just think there's like
- 17:25many things to do like new languages,
- 17:27right? Like obviously like you know, uh,
- 17:30we want things like Rust that are very
- 17:32efficient, but we also want things to be
- 17:35uh potentially modular and like hot
- 17:37reloadable. Um, we want it to be
- 17:40potentially readable, right? Uh, new
- 17:42primitives like there are probably new
- 17:44types of databases, new types of servers
- 17:46like um new operating system
- 17:49interactions. I've been sort of like
- 17:50vibe coding a bunch of personal
- 17:51software, um, that sort of messes with
- 17:53my operating system.
- 17:55Uh, there are new interaction paradigms
- 17:57possible, I think. Um, a lot of like old
- 18:00HCI research is now like very
- 18:01interesting to revisit. Um, new
- 18:04aesthetics, I think the way that we like
- 18:06use and design software uh
- 18:09can change quite a lot.
- 18:11And um yeah, I I think that like there's
- 18:13so much to do here that I think I I
- 18:15don't think that any of this is
- 18:17everyone every speaker I think has
- 18:19talked about this and
- 18:21um I I don't have too much more to add,
- 18:24but I think that like what I really want
- 18:26us to say is that what we need to do is
- 18:29to decide what the future of work looks
- 18:31like, right? And decide what our ideal
- 18:33experience is because the models are
- 18:35not, you know, this alien intelligence
- 18:38that's like trying to force us into a
- 18:39specific way of working. Uh they can
- 18:42kind of do everything and and they will.
- 18:44Um but our job is to like bring about
- 18:46the Renaissance. So
- 18:48thank you guys. That's my talk.
About this transcript
This page contains the full transcript of The Sweet Side of the Bitter Lesson — Thariq Shihipar, Independent by AI Engineer, generated from the public captions YouTube serves with the video. The transcript has 3,643 words across 534 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.