Skip to content
Linkdrop
Links

The Sweet Side of the Bitter Lesson — Thariq Shihipar

At CODE Summit 2026, Thariq Shihipar reflects on the Bitter Lesson and argues that AI can support both software industrialization and a creative renaissance. A live demo explores readable specifications and custom scripting languages as ways to preserve creative control when working with coding agents.

videoAI Engineer · 19 min · Nov 19, 2026 · youtube.com · Saved Nov 19, 2026

bitter lesson, coding agents, creative experiments, agent prompting

Description

Vibe coding a fighting game left Thariq Shihipar unsure how his own characters worked. At CODE Summit 2026, he revisits the Bitter Lesson, Bash, React and MCP, contrasting AI’s jagged frontier with its smooth interior. His live demo uses readable specifications, custom scripting languages and IDE workflows to show how coding agents can preserve creative control and help bring about a software renaissance.

Speaker info:
- https://x.com/trq212
- https://www.thariq.io/

Timestamps:
0:00 The sweet side of the Bitter Lesson
1:27 Rethinking Bash, React and AI art
4:04 Keeping AI adoption human centered
4:40 Industrial revolution and software renaissance
5:53 The jagged frontier and smooth interior
9:19 Visualizing model knowledge
10:38 When vibe coding loses the developer
12:54 Separating engine code from readable specs
13:47 Live demo: a scripting language for a game
16:40 Designing a software renaissance

Transcript

The sweet side of the Bitter Lesson

0:12 Hey guys, I'm Tharak. Nice to meet you. Yeah, I wanted to talk to you about a couple things I've been thinking about since I left Anthropic. It's been a wild year and there a bunch of reflections I've had. And my conclusion is about, you know, what I'm calling the sweet side of the bitter lesson and the ways I've been wrong. So I I think I got a lot of things right this year.

0:41 I think I was right about models being able to edit and generate videos. I was talking about this like sometime in like June last year. I was right about models being able to use HTML to understand their work. And I was right about models being able to help you understand and figure out your unknown to help like plan and learn about things. So these were all things I was like right about and I was happy and I I think that like a lot of you might be incorporating this sort of stuff in your work.

1:14 But I've been wrong a lot too and I I think this is one of those things in AI where you sort of just like memory hole the things that you're wrong about and you're like, oh like I was just always expecting, you know, AI to be great. But here are some of the things I've

Rethinking Bash, React and AI art

1:27 been wrong about. So first like Bash is all you need. I I think there was this like moment where the agents were really good at using Bash and Claude Code was built around the Bash tool. And so, you know, people are like, well, we'll just turn everything into a CLI. Like, you know, like there is just going to be Bash and CLIs and that's going to like take over the world.

1:48 That's actually kind of not how it turned out. Like I actually prefer MCP to CLIs now, you know, but now there's also code mode and skills and progressive disclosure. Like there are so many other like ways that agents can, you know, act on the world, right? And the fact that they were using Bash was just like this weird remnant of, you know, like them being trained in a certain way at the start.

2:11 Then there's like React is all you need. I think for a moment it seemed like the models were really really good at writing React. And so, you know, if you wanted to create an AI enabled app you should figure out how to use React. Like you should use React Native, you should use, you know, like you should just make everything on the web.

2:30 And that also has not been true. I I think like now, you know, models can write it in HTML, they can write it in React, they can write it in Swift, they can write it in the Figma proprietary JSON format. And they'll all look roughly the same. Right? And so I was afraid that like you know the models have been trained on React and so we'd just be writing React forever and React would be like the last language almost, right? And that's almost certainly not true, right? I think a lot of people are writing their own languages now.

3:01 And I think one of the other things I was really concerned about was like diffusion models will kind of like eat art. Like it just be like, you know, we wouldn't have any control over these models and or the kind of art that these models are creating, but it would just like create the most beautiful stuff. And that hasn't been true at all. Like I think now LLMs can make art through code, through controlling your computer and using, you know, Blender or Photoshop or yeah or Figma or whatever you'd like, right? And so these were all sort of like fears I had about like I think the bitter lesson, you know, our alien intelligence, the Shoggoth, right? Like like these are all words we use I think to describe this kind of like fear that the models will just have this like way of doing things and it will be weird and

3:49 alien to us and we just need to accept it. You know, like you just have to like submit to this like foreign intelligence. But that's just like I don't think that's what happened, you know? And so I think that like the question come back to a lot is like okay, like

Keeping AI adoption human centered

4:05 how can we be AGI piled and human centered and what can we learn about you know, this progression what we like you know, what can I learn about what I was wrong about, right? So I think this is really really important because software engineers are like the leading edge of like AI adoption, right? Like any thing that is happening to software will happen to knowledge work and other parts of the economy and so I think how we deal and engage with uh with AI is um yeah, how other people will as well and you know, I really enjoyed the different talks here talking about things like yeah, fighting slop. I think you know, we're the first defense uh for that,

Industrial revolution and software renaissance

4:40 right? Or like being more ambitious. Um, what I like to think about it as AI is both at the same time an industrial revolution and a Renaissance, right? And I think this year was really about the industrial revolution, right? Like we use the word software factories very often um and uh like just sort of like turning software from this like craft that was very unreliable and hard to execute on into something that like is not the limiting factor, right? Um, and this sort of like industrialization is absolutely happening. This is definitely something that is happening with with software, but at the same time there's also kind of a Renaissance happening, right? Where people want and uh higher and higher quality software and every marginal bit of like creativity, utility, like spirit, like soul that you put into it is like is useful and like economically valuable,

5:41 right? And so simultaneously we've got both of these things that are very hard to juggle and and sort of like hold in your head at the same time, uh but I think it is very true and uh

The jagged frontier and smooth interior

5:53 yeah, we're doing both of them. So often times when we talk about AI capabilities, we talk about like the jagged frontier, right? And so uh you know, you maybe like you go back two years and you're like AI asking AI to like this function for you, um or, you know, like like a year and half ago, a year ago, it's like build this feature, then you're like, okay, build this like to do app, build like a kind of simple to do app, and then clone Slack for me now, right? And And And now it's like, okay, build Mario Kart from scratch, you know, and then then like port it to like the Wii or something.

6:23 There's like the capability increases are the things that you see, right? Where, you know, whenever you look at the evals and benchmarks and stuff, you see this like just um this jagged frontier of like capabilities. But I also like to talk about like the smooth interior where the the way Whoops. the way the models Is there an animation playing here? Um the smooth interior is like the idea is that the models are um as they get better, the the animation here is like this jagged frontier and this circle in between. And so uh the circle sort of represents all of the things that the models are too are smart enough to do. And so like examples being like uh writing React code, right? Or like using Bash. And what you see is that like it's not jagged below the frontier, right? Like if it can write

7:22 React, it can write Figma, it can write JSON, it can do sort of like all of these capabilities all together. And there's like a breadth of tasks that the models are really really good at. And this is something that's also I think really like important and um for me even more impactful because I don't really care that they can build Mario Kart from scratch anymore, right? What I really care about is like, hey, is the software that we're building, is it high quality?

7:48 Is it worthy of like a Renaissance? You know what I mean? So um yeah, like I think that like, you know, you can do things the way you want. So like if you're building mockup, you could be like, hey, can I get an HTML? Can I get a React? Can I get a Figma file? Can I computer use my computer and use After Effects and make a video?

8:07 There are all these different ways of getting the like final result. And I've been really happy with this. I really was afraid that there would just be like one thing that the models love doing and we'd only be doing that. Um and you know like taken to the limit kind of like, you know, we talk about this a lot but like, you know, everything we've talked about so far about like, you know, how you build like these harnesses and software factories and um these like the scaffolding for the models.

8:35 Even things like, you know, for a while we were also like, hey um the model you'll like your job is to QA the model, you know, or review the code. And that's clearly not the case, right? I think as the models get smarter and smarter um it will become more of this like smooth interior, right? Where they're just doing things well, right? And like the perfect model you can imagine won't need to like really verify its work or scan for security vulnerabilities. We'll just like write the code and be like, hey I'm I'm done now. Um and you will check and you'll like, oh you're right. Like, you know, according to the spec and the goals that you that we did you wrote you did it correctly. Um and uh I think that's actually very exciting.

9:18 Um one of one of the ways I like to sort

Visualizing model knowledge

9:20 of think about visualizing how how much the model knows of this eval. I think Aparthy um posted it. But basically you ask any I think these are these are Claude models. You ask a model uh you give it a longitude latitude and you say like, hey is this point uh earth or water, right? Uh ground sorry land or water. And it from that you plot a map of the earth, right? And yeah this is Opus uh 4.6 uh is somewhere over there, right? With quite a good software engineering model, kind of like a blurry view. And now you look at Opus 5.5 and it's like almost perfect, right? Like it knows every piece of, you know, these like these coordinates. It knows like the entire world basically. And so this I I just gives you an idea besides like, you know, the benchmarks about like the the

10:13 the jagged frontier. I feel like that's less and less important to us, and more important is just like the general knowledge that the models have. And what's exciting, I think about this, is that we can use intelligence to understand and shape how we use intelligence, right? And I think this that will be like the theme of the next year, right? Where this year I think was really about the industrialization of software and and creating this like

When vibe coding loses the developer

10:38 reliable way to make software. Um, but what does that look like? So I have been building a game. Um, it's this like thing I've been posting on Twitter about. Um, it's a 2D like isometric game with like it's kind of like a fighting game inspired by things like Avatar and Smash Brothers. Uh, and I have kind of like it's fun, but it's also quite frustrating to do. Like I'm really trying to like, you know, like be like a great vibe coder here. Um, I don't really know what each character does sometimes because like I will say like, hey, I want this character to throw a rock. And then it'll be like, oh yeah, great, but like there are all these like different things that can happen when you're throwing a rock. Like this character can also summon a wall.

11:21 So does the rock go through the wall or not? And like how what speed or like what what about size and charge and all these questions? Um, and I I find that like the model just makes a bunch of choices for me and then I try it out and I'm like, wow, I have no idea what this character actually does.

11:38 I get distracted pretty easily. I don't feel in the zone very much. You know, I I sort of send off a prompt and I start doing something else and I come back. Um, and I honestly make less progress because of these things, right? And so uh, I think this is frustrating and I think for a while I was like, man, I guess this is just like how you work with AI. This is just like what it's going to be like, you know? But I think reflecting I'm like, man, I I don't think that's right. I think that like, you know, we have this like the AI can work in the way that we want, right? And so yeah, I don't think it's acceptable. I don't think that this is what leads to the Renaissance, right? I don't think this is what leads to like incredible

12:14 software. And part of it, you know, going back to what Dex said is like like great work is just about like this this messy process, right? Of like don't know like exactly what you want. It's more about the choices you made and what wasn't involved or wasn't what wasn't included versus like what was included. And so yeah, I think there's like you you like kind of just have to be an expert. You have to like just really know your stuff in order to make something great, right? In order to like make something worthy of a Renaissance, right? And I just like when I look at something I'm vibe coding, I'm like, dude, I have no idea. Like I just don't really get it, you know?

Separating engine code from readable specs

12:55 And so what what am I doing? How am I fixing this? One of the ways I'm doing it, and I'm still quite early on. I'm not saying that I've like figured it all out, but I think one of the ways I'm doing it is that I have two parts of my codebase. I've got like this like engine part of the codebase, and then I have a folder like I think I call it spec, but I think I'm just calling it human or something.

13:16 And that folder is the code that I read, you know? And it's like this code that's linted and like there's all these rules for Claude about like what that code looks like. It's supposed to feel like English. It's kind of like a spec, and I'll I'll show you what it looks like. But it's at the density where if I just read this code, I know what's happening and I can edit it, I can tweak it.

13:42 But more importantly, I just like, you know, can have a bit of a ground truth

Live demo: a scripting language for a game

13:47 of like what the what my game looks like. And so let's do let's do like a quick demo here. By demo I guess I mean I'm just going to whoops. This is my eTalk. Okay. Here we go. Okay. So this is my game. I've got in here this folder game, and then I've got here this code. This is the like agent code, right? Like this is the code where I'm like, dude, I I don't know how you know cameras work or whatever. I you you like there's a bunch of code here that I don't But then in this code, I see have all my characters and I have their abilities.

14:20 And so I can go in here and I can say like, okay, this is the rock move, right? Um and inside of this move, there's a bunch of ways to tune it. Um it's well commented. It uses things like meters, you know what I mean? Instead of like X like 0.5 or whatever. Um it uh tries to like create a bunch of different abstractions along the way.

14:45 And so I can see that like one of the things that, you know, is useful is that like there are three ways to throw for this character to throw a rock, right? Like it can uh grab it from the ground, it can grab it from the wall or something. Um and I can see sort of like the um the true logic of the character, right?

15:05 I I don't really know how the engine works or how, you know, like this is implemented underneath. But I've essentially created this like almost a custom scripting language for my game. Um and so that means like, you know, when I'm playing this game, um I can tweak little things. You know, I'm still I'm still like working on it.

15:23 So like I can say I want the rock to be a lot bigger. So I can say like um you know, 1.5, but I'm using an actual IDE. Uh and so if I go here to my game, uh you know, I can see that like I have a rock, I can charge it, gets really big and I throw it, right? Um but I'm like kind of it it it's really interesting cuz I'm also starting to use cursor more and I'm using the tab model more, you know what I mean? Because now now I can be like, oh hey, what if I just want to say like, you know, uh if you summon a wall in the middle, like queue it up as an action, and then cursor will sort of like, you know, do the the tab complete here. Um and this is kind of like it's kind of

16:06 like spec driven development, you know, but the spec is this like this this scripting language basically. And um yeah, it's like is this the final way I'm going to write game? I'm not sure. It's like, you know, in some ways, uh, when I need to do something complicated, I still need to ask Claude to do it, right? And I need to ask to Claude to like make the primitives, but it's actually really easy for me to understand what's happening and I feel like much more in control and while like Claude is working, I can also sort of review it and work with it.

16:37 Um, and that's been really exciting. And so

Designing a software renaissance

16:40 I think the takeaway for me has been that like um the takeaway for me has been that like I now have to think about the way the best way for me to deliver high quality software, right? And I'm not as constrained and limited by like the model's capabilities and like what it likes to write code in or something. I'm just not thinking about like I need to think about like how can I deliver great software?

17:08 And so I think if we like really internalize that, I think then this the next year will be about a renaissance of software. And you know, I think many people have talked today and you know, this conference about how to be more ambitious and I don't think anything any of this is quite new, but I just think there's like many things to do like new languages, right? Like obviously like you know, uh, we want things like Rust that are very efficient, but we also want things to be uh potentially modular and like hot reloadable. Um, we want it to be potentially readable, right? Uh, new primitives like there are probably new types of databases, new types of servers like um new operating system interactions. I've been sort of like vibe coding a bunch of personal software, um, that sort of messes with my operating system.

17:55 Uh, there are new interaction paradigms possible, I think. Um, a lot of like old HCI research is now like very interesting to revisit. Um, new aesthetics, I think the way that we like use and design software uh can change quite a lot. And um yeah, I I think that like there's so much to do here that I think I I don't think that any of this is everyone every speaker I think has talked about this and um I I don't have too much more to add, but I think that like what I really want us to say is that what we need to do is to decide what the future of work looks like, right? And decide what our ideal experience is because the models are not, you know, this alien intelligence that's like trying to force us into a specific way of working. Uh they can

18:42 kind of do everything and and they will. Um but our job is to like bring about the Renaissance. So thank you guys. That's my talk.