Learning Library

← Back to Library

AI-Native Writing: Next Compute Leap

Key Points

  • Code has evolved dramatically in just a few decades because it was built to work hand‑in‑hand with ever‑more powerful computers, whereas natural language was only later “bolted on” to technology.
  • Modern software engineering practices—DevOps, CI/CD pipelines, testing and staging environments, GitHub, etc.—are recent innovations that exploit code’s computational design to dramatically improve development speed and reliability.
  • While machines can now comprehend and generate natural language far better than before, they still haven’t mastered creating great literature, highlighting the gap between natural‑language understanding and true creative writing.
  • Voice interfaces are essentially a revival of pre‑writing oral communication rather than a brand‑new breakthrough, but the next true compute‑native leap will be AI‑driven writing tools that are built specifically for the machine era.

Full Transcript

# AI-Native Writing: Next Compute Leap **Source:** [https://www.youtube.com/watch?v=l8Po2CyWZag](https://www.youtube.com/watch?v=l8Po2CyWZag) **Duration:** 00:10:29 ## Summary - Code has evolved dramatically in just a few decades because it was built to work hand‑in‑hand with ever‑more powerful computers, whereas natural language was only later “bolted on” to technology. - Modern software engineering practices—DevOps, CI/CD pipelines, testing and staging environments, GitHub, etc.—are recent innovations that exploit code’s computational design to dramatically improve development speed and reliability. - While machines can now comprehend and generate natural language far better than before, they still haven’t mastered creating great literature, highlighting the gap between natural‑language understanding and true creative writing. - Voice interfaces are essentially a revival of pre‑writing oral communication rather than a brand‑new breakthrough, but the next true compute‑native leap will be AI‑driven writing tools that are built specifically for the machine era. ## Sections - [00:00:00](https://www.youtube.com/watch?v=l8Po2CyWZag&t=0s) **Code Evolution Outpaces Language** - The speaker contrasts programming's rapid, compute‑driven evolution and tooling advancements with the slow historical change of natural language, highlighting how machines now finally grasp human language. - [00:04:10](https://www.youtube.com/watch?v=l8Po2CyWZag&t=250s) **AI‑Enhanced Document Production Workflow** - The speaker argues that because knowledge work is linguistically complex, future automation should move beyond simple chatbots toward AI‑driven systems that treat documents like code—offering native variant generation, staged drafting, verification, and continual evolution. - [00:07:26](https://www.youtube.com/watch?v=l8Po2CyWZag&t=446s) **AI Model Development Pipeline** - The speaker describes a multi‑stage workflow that moves a document through different AI models—drafting with 03, deep problem solving with 03 Pro, structuring with Opus 4, validation with Perplexity, and polishing with Sonnet 4—mirroring a software development pipeline and advocating this systematic approach over constantly searching for the “best” model. ## Full Transcript
0:00Code has gone through more evolution in 0:03the last 50 or 60 years than natural 0:06language has gone through since it was 0:09invented 200,000 years ago and since 0:12writing was invented a few tens of 0:14thousands of years ago maybe less. 0:17The point is code is evolving fast 0:21because code was designed to evolve with 0:24developing uh and more powerful computer 0:27systems and natural language has only 0:30been bolted onto computers. It's not 0:32really a computative technology. I want 0:35you to think about it. Code started out 0:38as you compose it, you stick it in, you 0:40hope it runs. We are so far past that 0:42and we have seen step change 0:43improvements as we have leveraged better 0:46code practices with more complex 0:49compute. So now we have DevOps as a 0:51discipline. That wasn't a thing when I 0:53was coming up. It was only a thing after 0:55the 2010s. We have like the idea of 0:58having a testing environment, the idea 1:00of having a staging environment, the 1:02idea of having CI/CD pipelines, the idea 1:04of GitHub. These are all innovations 1:07that combine the power of compute with 1:10code. And they're possible because code 1:13was designed first and foremost to be a 1:16language that worked well with 1:18computers. That's why we have it. It's a 1:21simplified language. It worked well with 1:23traditional computers. 1:25Fast forward now it's 2022. Machines 1:28understand natural language. They 1:30understand the language we have been 1:32speaking for hundreds of thousands of 1:34years. They understand the language we 1:36have been writing. We for the first time 1:40have machines that can master the 1:43semantic technical complexity of 1:46language. Language is much more complex. 1:49Natural language much more complex than 1:51computer code. It can express a much 1:53wider range of meaning. The complexity 1:56it can handle is much greater. Great 1:57literature in particular is extremely 1:59dense, highly complex, etc. Machines can 2:02speak and understand that. Machines have 2:03been trained on that. And yes, I'm the 2:05first person to say machines are not yet 2:07writing great literature. Uh I'm not 2:09trying to make that claim here. The 2:11point is this. 2:14Our tools for writing on computers have 2:18been bolted on for decades. 2:22They've been bolted on and we've just 2:24been tapping and writing out really the 2:27same fundamental technology that we've 2:30had since the beginning. And I know that 2:33one of the hot new things in 2025 is 2:35voice. But guess what? It can't be that 2:38hot. It's actually going back to before 2:40we invented writing to when we were in 2:42oral culture. It's just voice again, 2:45only now we have computers in the mix. 2:47It's not actually a new innovation. 2:50It's just going back to the way our 2:51brains were originally wired. And it's 2:54often easier to engage that part of the 2:56brain because writing is a learned 2:57technology and human brains haven't 2:59evolved to make it truly native yet. So 3:01far so good. But you know what is going 3:05to happen that is truly compute align 3:08that is a computen native innovation on 3:12writing. We are going to have 3:15AI native tooling for writing. I don't 3:19mean you have a chat bar and then 3:21suddenly everything appears. I'm going 3:23to give you a little hint of it and I 3:24think we can infer a lot about how white 3:26collar work is going to go based on this 3:29by the way because if you think about it 3:31how much of our work right now is 3:34document creation if you're not coding 3:36which as we've seen is compute native 3:39it's seen the fastest development before 3:41AI frankly given cursor given windsurf 3:43and others I would argue it's seen the 3:44fastest development since AI the rest of 3:46us we're making documents and you know 3:49what I know there are enterprises that 3:52have document pipelines that use AI. I'm 3:55aware of them. I've advised on a few of 3:58them, but those are all almost without 4:02exception tightly complexity constrained 4:06because they have to be given the scale. 4:09Whereas traditional knowledge work is 4:10actually very complexity expansive. It's 4:12it's closer to the range of natural 4:14language. You have to do a lot of 4:17different things with documents. It's 4:18part of what makes white collar work 4:20actually quite difficult to understand 4:23and automate. It's really complicated. 4:26And what I'm saying here is not like a 4:28direct path to automation A to B. What 4:31I'm saying is that as machines 4:33understand our language, we can finally 4:36develop software that leverages compute 4:40to give us more options. And this is 4:42where I come back around and say again, 4:44it's probably not the chatbot, even 4:46though the chatbot is what we're using 4:48right now. It is probably going to be 4:51something that puts uh optionality and 4:54leverage first and foremost. So, for 4:57example, look at the way you can get 5:01multiple variants easily with AI. 5:03There's no reason that has to be a point 5:05and a click or a type away. It can be 5:07native and obvious. There are a few 5:09tools that already are playing with this 5:11idea where you just have multiple 5:13variants of everything you write. 5:16Another one, why don't we have the 5:20concept of production code for writing 5:23documents? I know we have draft and I 5:25know we have final but we don't really 5:28think of our software for documents as 5:31being something that we could evolve 5:32with AI so that we have like the right 5:35model for drafting we have the right 5:37verification step and staging for 5:39checking our facts and claims and we 5:42treat it like code in the sense that we 5:44check it for clarity we check it for 5:45coherence then we finally deploy it to 5:47production 5:49is that too technical a way of thinking 5:50about it I don't think so because I 5:52think you can take that same princip 5:54principle and then at the end if you 5:55want to dress it up and make it a fancy 5:58report that just becomes a separate step 6:00at the presentation layer. The core of 6:02the context, the core of the text is 6:04still there. Imagine how much easier it 6:07is if you can deploy text across 6:09multiple channels at once. That way, 6:12it's like being able to deploy code to 6:14multiple boxes. You would be able to 6:17say, "Okay, so we're going to tweak this 6:18core message. We're going to send it in 6:21um a multivariant stream, right? We're 6:24going to have cohort one be the 6:25executive team, cohort two be the 6:27marketers, cohort 3 be customer success, 6:30and you're sending the same update, but 6:33you're tuning it to what they need to 6:34hear and focus on. 6:37That's all stuff that was not possible 6:39before because AI did not give us the 6:42chance to understand natural language 6:46until the popularization of large 6:48language models. We had AI before then, 6:51but true large language models were the 6:54breakthrough we needed to grasp the 6:55depth and complexity of text. And that 6:58enables us for the very first time to 7:01have compute platforms that actually 7:04evolve the way we think and write and 7:06take us beyond what we've been doing for 7:09tens of thousands of years. I'm very 7:11excited about it. I can't tell you how 7:14exactly it's going to look, but I will 7:16say I am already seeing professional AI 7:21workers do this manually. 7:24For example, 7:26I am going to be moving my document flow 7:30from chat GPT to drafting with 03 7:35because I think 03 is a good conceptual 7:37thinker. Maybe sometimes if it's a hard 7:39problem, I'll go to 03 Pro and then move 7:42from there into Opus 4 to understand and 7:48structure the problem a little bit. You 7:50see how I'm thinking about it? moving 7:51from the dev environment now I'm 7:53starting to move it into almost a pull 7:56request merge scenario where I have to 7:57think about the structure of what I'm 7:59writing and whether it is congruent with 8:02other things that I've written and other 8:03things I'm focused on then moving it to 8:06what I would call testing where you're 8:08going to perplexity with that document 8:10and you're testing whether the claims 8:12are true and then moving for polishing 8:14to sonnet 4 also a claude model because 8:18sonnet 4 is an exceptional writer and 8:20it's able to polish that text a little 8:22bit more. I would call that, you know, 8:26staging, getting ready for production, 8:28whatever you want to say. The point is, 8:31I am essentially mimicking that dev 8:33pipeline. And I'm not the only one. Lots 8:36of people are doing this, but I think 8:38we've been doing it individually on our 8:40own. And we talk about it as if it's 8:42finding the best model for this and this 8:45and it gives us a headache because it 8:47feels like we have to constantly be 8:49picking better models. I think it's more 8:51stable and helpful to think about it as 8:54we are not equipped yet with tools that 8:56make writing native for AI. We could be 8:59eventually in the meantime. This is our 9:02best way to manually simulate it. And if 9:05we understand the jobs as mapping sort 9:08of loosely to the leverage that code has 9:11been able to get through more compute I 9:14think we're going to get farther because 9:15to be honest it's actually not that much 9:18of a stretch to say most knowledge work 9:20goes through development it goes through 9:22testing it goes through a merge process 9:25we call that peer review and it 9:28eventually gets to production. It's kind 9:30of how it works. If we can figure out 9:33how to do that with computer programs 9:35that are small, with smaller scale 9:37compute, we're going to be able to 9:39figure out how to make that easier for 9:41ourselves, with LLMs, with larger scale 9:45compute, and with all the power of 9:47natural language that makes knowledge 9:49work so interesting. 9:51So there you go. That is the thing that 9:54has been keeping me awake at night. That 9:56is the thing I cannot stop thinking 9:58about. 10:00I 10:02am so excited 10:04because I feel like we are standing on 10:06the edge of a different way of writing 10:09for the first time in a very long time. 10:12If you're out there building on that, if 10:14you're out there working on that, I know 10:15a few founders who are. I'm excited for 10:17you guys and cheering for you guys. In 10:19the meantime, I'm going to keep 10:20documenting how I build, how I learn, 10:23how I think, and uh yeah, good luck out 10:26there, guys. Cheers.