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